From 996a33d88ae83faf01fdcaca9e453a542d76a0be Mon Sep 17 00:00:00 2001 From: rosong11 Date: Tue, 18 Aug 2026 03:09:44 -0700 Subject: [PATCH 1/4] add nvfp4 situ moe cubins Signed-off-by: rosong11 --- .../BatchedGemmInterface.h | 9 +- .../trtllmGen_bmm_export/GemmOptions.h | 4 +- .../trtllmGen_bmm_export/KernelMetaInfo.h | 13086 +++++++++++----- .../trtllmGen_bmm_export/KernelParams.h | 2 +- .../trtllmGen_bmm_export/KernelTraits.h | 6 + .../trtllmGen_bmm_export/config.json | 40 + ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2N_bM_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...Ov_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...Ov_bM_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...asFp32M_bN_rgTma_clmp_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...asFp32M_bN_rgTma_clmp_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst | 4 +- ...x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...dS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...dS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...dS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...dS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...dS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...dS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...W8_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...W8_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...dS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...dS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...dS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...dS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...dS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...dS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ..._schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst | 4 +- ...x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...dS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ..._schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst | 4 +- ...x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...dS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...dS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...dS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ..._schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst | 4 +- ...S_relu2_bN_rgTma_clmp_sm100f.cubin.tar.zst | 4 +- ..._schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst | 4 +- ...S_relu2_bN_rgTma_clmp_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...Ov_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...2M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 3 + ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...stsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...stsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...lmp_geGlu_lsfaW2_dynB_sm100f.cubin.tar.zst | 4 +- ...lmp_geGlu_lsfaW4_dynB_sm100f.cubin.tar.zst | 4 +- ...p_siTuGlu_lsfaW2_dynB_sm100f.cubin.tar.zst | 3 + ...p_siTuGlu_lsfaW4_dynB_sm100f.cubin.tar.zst | 3 + ...mp_swiGlu_lsfaW2_dynB_sm100f.cubin.tar.zst | 4 +- ...mp_swiGlu_lsfaW4_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_lsfaW2_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_lsfaW4_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_lsfaW2_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_lsfaW4_dynB_sm100f.cubin.tar.zst | 4 +- ...rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 3 + ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...stsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...stsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...eGlu_lbW4_lsfbW4_dynB_sm100f.cubin.tar.zst | 4 +- ...uGlu_lbW4_lsfbW4_dynB_sm100f.cubin.tar.zst | 3 + ...iGlu_lbW4_lsfbW4_dynB_sm100f.cubin.tar.zst | 4 +- ...clmp_lbW4_lsfbW4_dynB_sm100f.cubin.tar.zst | 4 +- ...clmp_lbW4_lsfbW4_dynB_sm100f.cubin.tar.zst | 4 +- ...rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 3 + ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...stsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...stsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 3 + ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...stsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...stsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 3 + ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...stsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...stsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 3 + ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...stsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...stsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 3 + ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 3 + ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 3 + ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 3 + ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 3 + ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 3 + ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...eGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst | 4 +- ...uGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst | 3 + ...iGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst | 4 +- ..._clmp_geGlu_lbW4_dynB_sm100f.cubin.tar.zst | 4 +- ...lmp_siTuGlu_lbW4_dynB_sm100f.cubin.tar.zst | 3 + ...clmp_swiGlu_lbW4_dynB_sm100f.cubin.tar.zst | 4 +- ...clmp_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst | 4 +- ..._rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst | 4 +- ...clmp_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst | 4 +- ..._rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst | 4 +- ...eGlu_laW8_lsfaW4_dynB_sm100f.cubin.tar.zst | 4 +- ...uGlu_laW8_lsfaW4_dynB_sm100f.cubin.tar.zst | 3 + ...iGlu_laW8_lsfaW4_dynB_sm100f.cubin.tar.zst | 4 +- ...clmp_laW8_lsfaW4_dynB_sm100f.cubin.tar.zst | 4 +- ...clmp_laW8_lsfaW4_dynB_sm100f.cubin.tar.zst | 4 +- ...rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 3 + ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...stsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...stsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 3 + ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...stsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...stsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 3 + ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...stsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...stsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 3 + ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...stsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...stsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 3 + ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 3 + ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 3 + ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 3 + ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 3 + ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 3 + ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...stsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...stsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 3 + ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...stsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...stsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 3 + ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 3 + ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 3 + ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...stsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...stsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 3 + ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...stsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...stsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 3 + ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...stsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...stsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 3 + ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...stsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...stsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...asFp32M_bN_rgTma_clmp_sm100f.cubin.tar.zst | 4 +- ...rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 3 + ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 3 + ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 3 + ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 3 + ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 3 + ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...asFp32M_bN_rgTma_clmp_sm100f.cubin.tar.zst | 4 +- ...rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 3 + ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 3 + ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst | 4 +- ...clmp_swiGlu_lbW8_dynB_sm100f.cubin.tar.zst | 4 +- ..._rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst | 4 +- ..._rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst | 4 +- ...clmp_swiGlu_lbW8_dynB_sm100f.cubin.tar.zst | 4 +- ..._rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst | 4 +- ..._rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...clmp_swiGlu_lbW8_dynB_sm100f.cubin.tar.zst | 4 +- ..._rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst | 4 +- ..._rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst | 4 +- ...clmp_swiGlu_lbW8_dynB_sm100f.cubin.tar.zst | 4 +- ..._rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst | 4 +- ..._rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst | 4 +- ...clmp_swiGlu_lbW8_dynB_sm100f.cubin.tar.zst | 4 +- ..._rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst | 4 +- ..._rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst | 4 +- ...clmp_swiGlu_lbW8_dynB_sm100f.cubin.tar.zst | 4 +- ..._rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst | 4 +- ..._rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...clmp_swiGlu_lbW8_dynB_sm100f.cubin.tar.zst | 4 +- ..._rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst | 4 +- ..._rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst | 4 +- ...clmp_swiGlu_lbW8_dynB_sm100f.cubin.tar.zst | 4 +- ..._rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst | 4 +- ..._rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ...tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst | 4 +- ..._rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst | 4 +- ..._rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst | 4 +- ..._rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst | 4 +- ..._rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst | 4 +- ..._rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst | 4 +- ..._rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst | 4 +- ..._rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst | 4 +- ..._rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst | 4 +- ..._rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst | 4 +- ..._rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst | 4 +- ..._rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst | 4 +- ..._rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst | 4 +- ..._rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst | 4 +- ..._rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst | 4 +- ..._rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst | 4 +- ..._rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst | 4 +- ..._schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst | 4 +- ..._rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst | 4 +- ..._rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst | 4 +- ..._schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst | 4 +- ..._rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst | 4 +- ..._rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst | 4 +- ..._schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst | 4 +- ...S_relu2_bN_rgTma_clmp_sm100f.cubin.tar.zst | 4 +- ..._schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst | 4 +- ...S_relu2_bN_rgTma_clmp_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...asFp32M_bN_rgTma_clmp_sm100f.cubin.tar.zst | 4 +- ...asFp32M_bN_rgTma_clmp_sm100f.cubin.tar.zst | 4 +- ..._schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst | 4 +- ..._schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst | 4 +- ..._schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst | 4 +- ...S_relu2_bN_rgTma_clmp_sm100f.cubin.tar.zst | 4 +- ..._schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst | 4 +- ...S_relu2_bN_rgTma_clmp_sm100f.cubin.tar.zst | 4 +- ...uGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst | 4 +- ...iGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst | 4 +- ...iGlu_lbW4_lsfbW4_dynB_sm100f.cubin.tar.zst | 4 +- ...uGlu_lbW4_lsfbW4_dynB_sm100f.cubin.tar.zst | 4 +- ...uGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst | 4 +- ...iGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst | 4 +- ...uGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst | 4 +- ...iGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst | 4 +- ...iGlu_lbW4_lsfbW4_dynB_sm100f.cubin.tar.zst | 4 +- ...uGlu_lbW4_lsfbW4_dynB_sm100f.cubin.tar.zst | 4 +- ...uGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst | 4 +- ...iGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...iGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst | 4 +- ...uGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst | 4 +- ...iGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst | 4 +- ...uGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...uGlu_lbW1_lsfbW1_dynB_sm100f.cubin.tar.zst | 4 +- ...iGlu_lbW1_lsfbW1_dynB_sm100f.cubin.tar.zst | 4 +- ...uGlu_lbW1_lsfbW1_dynB_sm100f.cubin.tar.zst | 4 +- ...iGlu_lbW1_lsfbW1_dynB_sm100f.cubin.tar.zst | 4 +- ...uGlu_lbW1_lsfbW1_dynB_sm100f.cubin.tar.zst | 4 +- ...iGlu_lbW1_lsfbW1_dynB_sm100f.cubin.tar.zst | 4 +- ...uGlu_lbW1_lsfbW1_dynB_sm100f.cubin.tar.zst | 4 +- ...iGlu_lbW1_lsfbW1_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...Tma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...gTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst | 4 +- ...iGlu_laW8_lsfaW4_dynB_sm100f.cubin.tar.zst | 4 +- ...iGlu_laW8_lsfaW4_dynB_sm100f.cubin.tar.zst | 4 +- .../modules/fused_moe/fused_moe_trtllm_gen.py | 69 +- .../modules/fused_moe/moe_op_backend.py | 54 +- .../_torch/modules/fused_moe/quantization.py | 45 +- .../_torch/modules/moe/test_nvfp4_situ_moe.py | 313 + tests/unittest/_torch/thop/serial/test_moe.py | 17 + 692 files changed, 11141 insertions(+), 5186 deletions(-) create mode 100644 cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst create mode 100644 cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_siTuGlu_lsfaW2_dynB_sm100f.cubin.tar.zst create mode 100644 cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_siTuGlu_lsfaW4_dynB_sm100f.cubin.tar.zst create mode 100644 cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst create mode 100644 cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_siTuGlu_lbW4_lsfbW4_dynB_sm100f.cubin.tar.zst create mode 100644 cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst create mode 100644 cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst create mode 100644 cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst create mode 100644 cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst create mode 100644 cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst create mode 100644 cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst create mode 100644 cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst create mode 100644 cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst create mode 100644 cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst create mode 100644 cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst create mode 100644 cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst create mode 100644 cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_siTuGlu_lbW4_dynB_sm100f.cubin.tar.zst create mode 100644 cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_siTuGlu_laW8_lsfaW4_dynB_sm100f.cubin.tar.zst create mode 100644 cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst create mode 100644 cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst create mode 100644 cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst create mode 100644 cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst create mode 100644 cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst create mode 100644 cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst create mode 100644 cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst create mode 100644 cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst create mode 100644 cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst create mode 100644 cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst create mode 100644 cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst create mode 100644 cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst create mode 100644 cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst create mode 100644 cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst create mode 100644 cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst create mode 100644 cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst create mode 100644 cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst create mode 100644 cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst create mode 100644 cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst create mode 100644 cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst create mode 100644 cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst create mode 100644 cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst create mode 100644 cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst create mode 100644 cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst create mode 100644 tests/unittest/_torch/modules/moe/test_nvfp4_situ_moe.py diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/BatchedGemmInterface.h b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/BatchedGemmInterface.h index 18411f44c831..38b8135ac8f2 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/BatchedGemmInterface.h +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/BatchedGemmInterface.h @@ -958,14 +958,9 @@ class BatchedGemmInterface auto options = getOptionsFromConfigAndData(config, data); // Check options without modifications. - // - // NOTE(TRT-LLM local fix over producer output a2ad0544-dirty): the generator - // inserted the `worldSize` parameter into checkAndUpdateBatchedGemmOptions - // without updating this call, so the literal `false` bound positionally to - // `worldSize` (=0) and `updateOptions` silently took its default `true`. - // Pass both arguments explicitly. Report upstream before the next re-export. return checkAndUpdateBatchedGemmOptions(options, config.mSm, - /* worldSize */ 1, /* updateOptions */ false); + /* worldSize */ 1, + /* updateOptions */ false); } ////////////////////////////////////////////////////////////////////////////////////////////////// diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/GemmOptions.h b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/GemmOptions.h index 95555c93d86b..cf1569e2fe7f 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/GemmOptions.h +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/GemmOptions.h @@ -2148,8 +2148,8 @@ inline CUresult loadCubinData(CUmodule* module, Config const& config) // Trtllm links the cubin into the executable while Flashinfer loads the cubin from storage. #ifdef TLLM_GEN_EXPORT_FLASHINFER #ifdef TLLM_GEN_GEMM_CUBIN_PATH - static std::string const tllm_gen_gemm_cubin_path = std::string(TLLM_GEN_GEMM_CUBIN_PATH); - std::string const sha256 = config.mHash ? config.mHash : ""; + static const std::string tllm_gen_gemm_cubin_path = std::string(TLLM_GEN_GEMM_CUBIN_PATH); + const std::string sha256 = config.mHash ? config.mHash : ""; std::string fileName = config.mFunctionName; if (!fileName.empty()) { diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/KernelMetaInfo.h b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/KernelMetaInfo.h index 49adcadb7273..3dc3a48ec24d 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/KernelMetaInfo.h +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/KernelMetaInfo.h @@ -16,16 +16,19 @@ */ #pragma once -#include "BatchedGemmOptions.h" #include "tensorrt_llm/common/config.h" +#include "BatchedGemmOptions.h" -namespace batchedGemm { +namespace batchedGemm +{ TRTLLM_NAMESPACE_BEGIN -namespace kernels { + +namespace kernels +{ // clang-format off -#define TLLM_GEN_COMMIT "a2ad0544-dirty" +#define TLLM_GEN_COMMIT "fe08bc1f-dirty" #define TLLM_GEN_EXPORT_VERSION "7.0.4.0.4.0" #ifndef EXCLUDE_SM_100 @@ -263,11 +266,14 @@ extern unsigned char const Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2 extern unsigned char const Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin[]; +extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_geGlu_lsfaW2_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_geGlu_lsfaW4_dynB_sm100f_cubin[]; +extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_siTuGlu_lsfaW2_dynB_sm100f_cubin[]; +extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_siTuGlu_lsfaW4_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_lsfaW2_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_lsfaW4_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_relu2_bM_tma_ldgstsSf_rgTma_clmp_lsfaW2_dynB_sm100f_cubin[]; @@ -275,154 +281,193 @@ extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6 extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_silu_bM_tma_ldgstsSf_rgTma_clmp_lsfaW2_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_silu_bM_tma_ldgstsSf_rgTma_clmp_lsfaW4_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin[]; +extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_geGlu_lbW4_lsfbW4_dynB_sm100f_cubin[]; +extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_siTuGlu_lbW4_lsfbW4_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_lsfbW4_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_lsfbW4_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_lsfbW4_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin[]; +extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin[]; +extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin[]; +extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin[]; +extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin[]; +extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin[]; +extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin[]; +extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin[]; +extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin[]; +extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin[]; +extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_geGlu_lbW8_lsfbW4_dynB_sm100f_cubin[]; +extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_geGlu_lbW4_dynB_sm100f_cubin[]; +extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_siTuGlu_lbW4_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_relu2_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_lbW8_lsfbW4_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_relu2_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_silu_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_lbW8_lsfbW4_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_silu_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_geGlu_laW8_lsfaW4_dynB_sm100f_cubin[]; +extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_siTuGlu_laW8_lsfaW4_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_swiGlu_laW8_lsfaW4_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_relu2_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_laW8_lsfaW4_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_silu_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_laW8_lsfaW4_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin[]; +extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin[]; +extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin[]; +extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin[]; +extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin[]; +extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin[]; +extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin[]; +extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin[]; +extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin[]; +extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin[]; +extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin[]; +extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin[]; +extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin[]; +extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin[]; +extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin[]; +extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin[]; +extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin[]; +extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin[]; +extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin[]; +extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin[]; +extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin[]; +extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin[]; +extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin[]; +extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin[]; +extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin[]; extern unsigned char const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin[]; @@ -675,655 +720,697 @@ extern unsigned char const Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_ #endif // EXCLUDE_SM_103 #ifndef EXCLUDE_SM_100 -extern unsigned int const Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len; -extern unsigned int const Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len; -extern unsigned int const Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len; -extern unsigned int const Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len; -extern unsigned int const Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len; -extern unsigned int const Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len; -extern unsigned int const Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len; -extern unsigned int const Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len; -extern unsigned int const Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len; -extern unsigned int const Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len; -extern unsigned int const Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len; -extern unsigned int const Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len; -extern unsigned int const Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len; -extern unsigned int const Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len; -extern unsigned int const Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len; -extern unsigned int const Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len; -extern unsigned int const Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len; -extern unsigned int const Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len; -extern unsigned int const Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len; -extern unsigned int const Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len; -extern unsigned int const Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len; +extern unsigned int Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len; +extern unsigned int Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len; +extern unsigned int Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len; +extern unsigned int Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len; +extern unsigned int Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len; +extern unsigned int Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len; +extern unsigned int Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len; +extern unsigned int Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len; +extern unsigned int Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len; +extern unsigned int Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len; +extern unsigned int Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len; +extern unsigned int Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len; +extern unsigned int Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len; +extern unsigned int Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len; +extern unsigned int Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len; +extern unsigned int Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len; +extern unsigned int Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len; +extern unsigned int Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len; +extern unsigned int Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len; +extern unsigned int Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len; #endif // EXCLUDE_SM_100 #ifndef EXCLUDE_SM_100F -extern unsigned int const Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x128x256_s6_et128x128_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x128x256_s6_et128x128_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x128x256u2_s6_et128x128_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x256x256_s5x5x5x5x1x1_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x256x256_s5x5x5x5x1x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_fCp_tmOv_bM_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x64x512_s4_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x64x512u2_s4_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPdx3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPdx3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPdx3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128u2_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E4m3E4m3_Fp32_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E4m3E4m3_Fp32_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E4m3E4m3_Fp32_t128x256x128_s6_et128x32_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E4m3E4m3_Fp32_t128x256x128u2_s6_et128x32_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128_s8_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128u2_s8_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128u2_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128u2_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128u2_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128u2_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128u2_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128u2_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128u2_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128u2_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128u2_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x128_s7_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x128u2_s7_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x256_s4_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x256u2_s4_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x256x128_s5x5x5x5x1x1_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x128_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x128u2_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_geGlu_lsfaW2_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_geGlu_lsfaW4_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_lsfaW2_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_lsfaW4_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_relu2_bM_tma_ldgstsSf_rgTma_clmp_lsfaW2_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_relu2_bM_tma_ldgstsSf_rgTma_clmp_lsfaW4_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_silu_bM_tma_ldgstsSf_rgTma_clmp_lsfaW2_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_silu_bM_tma_ldgstsSf_rgTma_clmp_lsfaW4_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_geGlu_lbW4_lsfbW4_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_lsfbW4_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_lsfbW4_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_lsfbW4_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_geGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_geGlu_lbW4_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_relu2_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_lbW8_lsfbW4_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_relu2_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_silu_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_lbW8_lsfbW4_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_silu_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_geGlu_laW8_lsfaW4_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_swiGlu_laW8_lsfaW4_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_relu2_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_laW8_lsfaW4_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_silu_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_laW8_lsfaW4_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128u2_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128u2_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128u2_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128u2_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128u2_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128u2_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256u2_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256u2_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256u2_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128u2_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_Fp16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f_cubin_len; -extern unsigned int const Bmm_Fp16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f_cubin_len; -extern unsigned int const Bmm_Fp16_E4m3E4m3_Fp32_t128x8x128_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f_cubin_len; -extern unsigned int const Bmm_Fp16_E4m3E4m3_Fp32_t128x8x128u2_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f_cubin_len; -extern unsigned int const Bmm_Fp16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f_cubin_len; -extern unsigned int const Bmm_Fp16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f_cubin_len; -extern unsigned int const Bmm_Fp16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f_cubin_len; -extern unsigned int const Bmm_Fp16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7x7x7x7x1x7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_lsfbW4_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7x7x7x7x1x7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_siTuGlu_lbW4_lsfbW4_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128u2_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128u2_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4x4x4x4x1x4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_lsfbW4_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4x4x4x4x1x4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_siTuGlu_lbW4_lsfbW4_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256u2_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256u2_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x128_s6x6x6x6x1x6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x128_s6x6x6x6x1x6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x256_s3x3x3x3x1x3_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x256_s3x3x3x3x1x3_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128u2_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128u2_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256u2_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256u2_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE4m3MxE2m1_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7_et128x64_m256x128x32_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_laW8_lsfaW4_dynB_sm100f_cubin_len; -extern unsigned int const Bmm_MxE4m3_MxE4m3MxE2m1_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4_et128x64_m256x128x32_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_laW8_lsfaW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x128x256_s6_et128x128_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x128x256_s6_et128x128_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x128x256u2_s6_et128x128_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x256x256_s5x5x5x5x1x1_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x256x256_s5x5x5x5x1x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_fCp_tmOv_bM_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x64x512_s4_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x64x512u2_s4_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPdx3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPdx3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPdx3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128u2_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E4m3E4m3_Fp32_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E4m3E4m3_Fp32_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E4m3E4m3_Fp32_t128x256x128_s6_et128x32_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E4m3E4m3_Fp32_t128x256x128u2_s6_et128x32_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128_s8_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128u2_s8_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128u2_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128u2_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128u2_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128u2_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128u2_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128u2_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128u2_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128u2_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128u2_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x128_s7_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x128u2_s7_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x256_s4_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x256u2_s4_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x256x128_s5x5x5x5x1x1_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x128_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x128u2_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_geGlu_lsfaW2_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_geGlu_lsfaW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_siTuGlu_lsfaW2_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_siTuGlu_lsfaW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_lsfaW2_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_lsfaW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_relu2_bM_tma_ldgstsSf_rgTma_clmp_lsfaW2_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_relu2_bM_tma_ldgstsSf_rgTma_clmp_lsfaW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_silu_bM_tma_ldgstsSf_rgTma_clmp_lsfaW2_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_silu_bM_tma_ldgstsSf_rgTma_clmp_lsfaW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_geGlu_lbW4_lsfbW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_siTuGlu_lbW4_lsfbW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_lsfbW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_lsfbW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_lsfbW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_geGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_geGlu_lbW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_siTuGlu_lbW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_relu2_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_lbW8_lsfbW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_relu2_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_silu_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_lbW8_lsfbW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_silu_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_geGlu_laW8_lsfaW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_siTuGlu_laW8_lsfaW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_swiGlu_laW8_lsfaW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_relu2_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_laW8_lsfaW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_silu_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_laW8_lsfaW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128u2_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128u2_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128u2_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128u2_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128u2_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128u2_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256u2_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256u2_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256u2_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128u2_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_Fp16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f_cubin_len; +extern unsigned int Bmm_Fp16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f_cubin_len; +extern unsigned int Bmm_Fp16_E4m3E4m3_Fp32_t128x8x128_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f_cubin_len; +extern unsigned int Bmm_Fp16_E4m3E4m3_Fp32_t128x8x128u2_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f_cubin_len; +extern unsigned int Bmm_Fp16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f_cubin_len; +extern unsigned int Bmm_Fp16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f_cubin_len; +extern unsigned int Bmm_Fp16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f_cubin_len; +extern unsigned int Bmm_Fp16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7x7x7x7x1x7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_lsfbW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7x7x7x7x1x7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_siTuGlu_lbW4_lsfbW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128u2_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128u2_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4x4x4x4x1x4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_lsfbW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4x4x4x4x1x4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_siTuGlu_lbW4_lsfbW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256u2_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256u2_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x128_s6x6x6x6x1x6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x128_s6x6x6x6x1x6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x256_s3x3x3x3x1x3_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x256_s3x3x3x3x1x3_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128u2_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128u2_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256u2_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256u2_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE4m3MxE2m1_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7_et128x64_m256x128x32_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_laW8_lsfaW4_dynB_sm100f_cubin_len; +extern unsigned int Bmm_MxE4m3_MxE4m3MxE2m1_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4_et128x64_m256x128x32_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_laW8_lsfaW4_dynB_sm100f_cubin_len; #endif // EXCLUDE_SM_100F #ifndef EXCLUDE_SM_103 -extern unsigned int const Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len; -extern unsigned int const Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len; -extern unsigned int const Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len; -extern unsigned int const Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len; -extern unsigned int const Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len; -extern unsigned int const Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len; -extern unsigned int const Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len; -extern unsigned int const Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len; -extern unsigned int const Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len; -extern unsigned int const Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len; -extern unsigned int const Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len; -extern unsigned int const Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len; -extern unsigned int const Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len; -extern unsigned int const Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len; -extern unsigned int const Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len; -extern unsigned int const Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len; -extern unsigned int const Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len; -extern unsigned int const Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len; -extern unsigned int const Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len; -extern unsigned int const Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len; -extern unsigned int const Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len; -extern unsigned int const Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len; +extern unsigned int Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len; +extern unsigned int Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len; +extern unsigned int Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len; +extern unsigned int Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len; +extern unsigned int Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len; +extern unsigned int Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len; +extern unsigned int Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len; +extern unsigned int Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len; +extern unsigned int Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len; +extern unsigned int Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len; +extern unsigned int Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len; +extern unsigned int Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len; +extern unsigned int Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len; +extern unsigned int Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len; +extern unsigned int Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len; +extern unsigned int Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len; +extern unsigned int Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len; +extern unsigned int Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len; +extern unsigned int Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len; +extern unsigned int Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len; +extern unsigned int Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len; #endif // EXCLUDE_SM_103 static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { #ifndef EXCLUDE_SM_100 -{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len, 116368, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a", 512, "965a6594cf88ccc3cdf8b436af029beedb2bb95e9e49fb8950cf591e679e9f18", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len, 116368, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a", 512, "cf78cfd959740174d7750fdbc3b548ecab4a988bfde235a76d97907f4c98cc9c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -1449,7 +1536,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100a}, -{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len, 116240, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a", 384, "744b3bdebf9b75154d3423ff5aa1a6ce819096a208badeb9c00641be98fe47cb", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len, 116240, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a", 384, "68417a6586f1595f6acbb65af5d4295602e8c5f37d30dc46fe49a5fbd7e95629", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -1575,7 +1662,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100a}, -{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len, 116368, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a", 512, "1ea37fff7d955412fe2a7fea1144eb9e65e1c47f34c2ea4c344cd059524a265c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len, 116368, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a", 512, "e699eb00d15010f819161d54e49c2a31dbd2bdc7b7e8948b19fb603b465e89bb", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -1701,7 +1788,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100a}, -{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len, 116240, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a", 384, "20a899eef3b55e5b7a159bbda0d59bd75b23d9ca7394eda138bb4a11f269d991", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len, 116240, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a", 384, "9751eb91933667215523c290b1c7aaebbacf55ebbe910ad700a67db4741495a4", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -1827,7 +1914,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100a}, -{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len, 140944, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a", 512, "a0a6a9c48d1f6fa3fe4908f13a1fc0a36f92ca56ed5528f73df4da2059cd3d19", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len, 140944, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a", 512, "1f6f6b959e6d9dded1953ad122b2adbcd5c519df45c7caa22c211fd21b07c36a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -1953,7 +2040,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100a}, -{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len, 140816, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a", 384, "e33c28a26a6224ccccf7df9d87c49e7709e649fbbe75c2cd70d82868fac0a18b", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len, 140816, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a", 384, "72dabd361ac002192f560bed24711b27bfc547865e18060bbc9dbcfdc5c25f95", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -2079,7 +2166,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100a}, -{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len, 140944, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a", 512, "96cf4dc4578b1ac8a9f16fc27d2cbdbe3b2692e832e3fcac3082b352a1d7084c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len, 140944, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a", 512, "6a4dd116d62d97a8f1a11beeb4071e750df9b58e21523c4d97b7a1c6df7ce517", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -2205,7 +2292,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100a}, -{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len, 140816, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a", 384, "8b2c5af9de42887bca74bff8436c53c5b85a3db87287ffce1a2f8572ca44af1f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len, 140816, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a", 384, "d471cbf3611473011bda09806fe4543b06fe3570e46315404411c5eaa1877f8d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -2331,7 +2418,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100a}, -{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len, 157328, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a", 512, "fdab85caecfe0b99431e8dd648fe07f330f838e0514211ca0f2726d0ac0e9025", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len, 157328, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a", 512, "90574b45aa8d9580c21838673cd996c3a32dd51d70f26d41e0fc1c1ee480cdea", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -2457,7 +2544,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100a}, -{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len, 157200, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a", 384, "7c413e8cc59ad3b0356c7e8c55ae5452efa0a17cdeb38ef712eb6b2618d707ac", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len, 157200, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a", 384, "0b661a98caa929f67fa2afbdf36454f850f5c57789b74c31845b37c8feac6fa9", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -2583,7 +2670,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100a}, -{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len, 157328, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a", 512, "71e69d7b6228ce0a0389f45043380604934a0a6cb4be25a197eb2d3f5859e06a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len, 157328, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a", 512, "a81554522b9b6b041faee2bc67d783d4c10d57a7f1edffb021f070e3c2cc2f76", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -2709,7 +2796,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100a}, -{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len, 157200, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a", 384, "2bd9d50ccdade28e02be5827cde712400f5bab8eb377f23c0323184a1311eae1", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len, 157200, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a", 384, "e52781060cfd468da21d6eccab3d6c5ede83151a61c2f23f77262d79dd504d48", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -2835,7 +2922,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100a}, -{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len, 104080, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a", 512, "caa5b56acf9deb07a978e0b0e88bdc4e986f1447b843a9b0b873b51d0fa22c26", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len, 104080, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a", 512, "6a649996b0ae263e234ebcff9c0ed40612a86e453cf2fb40d955f51583161e6c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -2961,7 +3048,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100a}, -{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len, 103952, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a", 384, "32bbacd1b75e8498c14c76f7d4e3a826e8aaf0778d10925f8bc0e0172f85b82e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len, 103952, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a", 384, "5b7012a31450a22eaa97ed8a1b6f9ffd5c4faab96fe9659f591f9de7e8bdbfa4", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -3087,7 +3174,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100a}, -{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len, 104080, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a", 512, "f84c3fcba00a7b233fea74a74d7654b000ddb5def3a6d0a2d99b1f6be5dea766", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len, 104080, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a", 512, "8c2f55f9e22721e4dfb71797f0ca75e6ef4c04b9ba11a4de47936595152104ad", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -3213,7 +3300,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100a}, -{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len, 103952, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a", 384, "ad59d64951352b17982d2c8fc526a934c6be327bba3fa7c83aeed6ecd3da9ae7", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len, 103952, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a", 384, "174e40c2dfe3630e3a17faae10c714ba6bacdb58b539228288544a833c5c324a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -3339,7 +3426,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100a}, -{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len, 123536, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a", 512, "ff03a360a0d7a7dad7a81307ebe809879f3ae0fc7c0d0b3e4f4761594c845d98", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len, 123536, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a", 512, "7bfb796f447e257efefa6e08ac6ad29ed78af0ea07a9886d3963e1610515fa3e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -3465,7 +3552,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100a}, -{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len, 123152, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a", 384, "6b429ecb0da0ca2616e024d5163401298272ee935d9f3f2a279212196b83d79d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len, 123152, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a", 384, "8c30637f40daafcc8b9a6f4a298ba231370d97b1bfac55d341a170aab9ed526d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -3591,7 +3678,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100a}, -{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len, 123536, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a", 512, "116243ee6ae2bc3b7c66f71c36b70bbe87d8f2698ec3366faf00a67f8f5e9256", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len, 123536, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a", 512, "c152555cb5e4745cd15148ac62bada9af379707c9a624f4646c0268249f68962", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -3717,7 +3804,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100a}, -{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len, 123152, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a", 384, "efb3849f33afe9af7a03eecf149becf892cdd0ac3c82137ea72d946c1d99b029", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a_cubin_len, 123152, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a", 384, "73e0820d67eb2701e092d8049184fa117f49908feb58db7dfb4ac8a3a1bb6983", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -3843,7 +3930,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100a}, -{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len, 114320, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a", 512, "cd84bfe0a7d03fa0ccf4db6c60a06cda92813ae8e253672487d0f1da87aed5e6", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len, 114320, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a", 512, "aeffb15fbaf38db2219ab4bb5cb35ee457ba6ba830e3099f0f4725699097e55e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -3969,7 +4056,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100a}, -{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len, 114192, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a", 512, "9c4ed0b86d09d5b8c35793477324651c3023f0b634c23373b1d2e804b1811ac5", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len, 114192, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a", 512, "f5be749f00d5ec37b24ce85a2c88ab440db46c0c63cce0415684a95c21434407", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -4095,7 +4182,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100a}, -{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len, 114320, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a", 512, "3c9a421b78042cf35a6f5fb515e95c20d4947f111e9d21585213a2222372fb9a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len, 114320, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a", 512, "a41752acc632132ac9fac1fa690bc652e260e01986d42596437fc160ed157c6d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -4221,7 +4308,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100a}, -{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len, 114192, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a", 512, "229b2b98791c60c27647748ccc5b81cf0efbe969dff9dbf44f4ad14830004693", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len, 114192, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a", 512, "d8ed1b40834886c8e39d235c59aa2838e1d273f7bf9eef2a61743628528f79fd", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -4347,7 +4434,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100a}, -{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len, 136848, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a", 512, "e35c45da758ccc295101ae3b81e3798f1f241b4e10fd188de1b2cc8b6c343004", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len, 136848, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a", 512, "9f3476221e5eda421f4b348e8cf5677c25a7d0b005fbbfe27eb55a386510c2d4", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -4473,7 +4560,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100a}, -{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len, 136720, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a", 512, "075f09d0e0adb46bbd9a2a330a22f914ae893777181e5c348e198e7d14aebffb", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len, 136720, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a", 512, "d27ed09fd633ae642580bd9498177ffbdd691ecf0e70dd1c7cc348be1ea721a6", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -4599,7 +4686,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100a}, -{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len, 136848, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a", 512, "8da4142dda1c730c25048a5adb85dda7862fa32f5b81c3d75c09a313661bfa1a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len, 136848, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a", 512, "8ca06ba4d1ceda682e955ea1b69554cc5097a89825f90b591bf2092bd4b26084", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -4725,7 +4812,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100a}, -{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len, 136720, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a", 512, "1f10c947c51c8c1ff0685f2a69492e9f3c8f5326bce838c317e5a346aac44775", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len, 136720, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a", 512, "56bfeab07aee1e762cd3a35c51c40f065a8da3e30f6ef2c696307b1ff2f26455", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -4851,7 +4938,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100a}, -{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len, 149136, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a", 512, "e4ad20bc661353d925eed9fbcf2c6cb162a303646f37a93830e22bb12ca2e68d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len, 149136, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a", 512, "2fecfe6ddee0503407cc50aafcc1d7ee742765a959edbc7bd5e89be47c810eff", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -4977,7 +5064,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100a}, -{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len, 149008, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a", 512, "b044e59c128a0379b26a1e2da332d41e523aad3a37b27ff72ed513db18026f5c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len, 149008, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a", 512, "630349a09926a065276b605bb2e9146a445917236e793b4801875ba3bd541565", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -5103,7 +5190,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100a}, -{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len, 149136, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a", 512, "c15033f19958a9505d57d7fe1315dbb64d355830f6103df4a4924ef4600ea50b", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len, 149136, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a", 512, "e526c065557422f4ac1f28d7df56b66ccb623a89810829bd8af3d81bdc0b12d8", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -5229,7 +5316,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100a}, -{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len, 149008, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a", 512, "80183b38838e754cac2a2b2413cae34e1c556512d2a2fe9cf07b37743d3e4516", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len, 149008, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a", 512, "0ec3b65f3a849a99870d3f306c78274737aea529e09a517f5566359e43053eae", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -5355,7 +5442,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100a}, -{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len, 103056, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a", 512, "ee31f269ef5f142e18c8080068882dbf4565c4fabf4223da369b1ec236bbd593", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len, 103056, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a", 512, "253a088e6aecc117cf33127ecefc88031962234f2ac79175322ae05ba33ea211", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -5481,7 +5568,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100a}, -{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len, 102928, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a", 512, "c68a34b135d156f1a0da6833f9a6f9697da87a1feead83f10902faa9a882980d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len, 102928, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a", 512, "6b47dad61379b59e766fe631cfffd7d76ce615dd62c7dc99677528963fc87cb8", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -5607,7 +5694,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100a}, -{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len, 103056, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a", 512, "ef26d6c75598e85bca1f98b0c8dfe1631b175333e3d665adab59198fcfa102b2", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len, 103056, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a", 512, "5708785e796389d0a723e9f370087b4cc4604a2722d9f97a0c9e9682867f20dc", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -5733,7 +5820,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100a}, -{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len, 102928, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a", 512, "c60ebb86795820f14f7915e83a865a9e4ab20b62d7fcdbdeb0a44fbe8a0984dd", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len, 102928, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a", 512, "976e137b61b08eca3eee83c6799b8b06be231d133ce81cd133190b021cb391d6", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -5859,7 +5946,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100a}, -{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len, 122512, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a", 512, "2e70ffe6e5f5013b1101409cd4740822ccf5d71cecd228fc7641810df2624f42", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len, 122512, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a", 512, "0a52c309bfe26be2d71343c5a02285c7115f43eb45f66c6d3f58941512408aae", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -5985,7 +6072,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100a}, -{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len, 122128, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a", 512, "b9dc80efc5787e83e18dd660ee5fe3e0a22270cd5b4e3f60d815b3d534c959bc", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len, 122128, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a", 512, "8bd34d0a3461817bd59ef37eda5de54273250c908429cb1cfd7c8333f7551952", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -6111,7 +6198,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100a}, -{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len, 122512, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a", 512, "4dcc53a09ad5dd1a23579c9041ea50369285530abf786d3bbd99bd3d8f1b417a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len, 122512, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a", 512, "ce9c2c6ac1ffdbf68d00c4f17ba559dee94ec3b7e7f799411381552cee9a3258", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -6237,7 +6324,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100a}, -{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len, 122128, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a", 512, "4851b4d886236f65e5973797a30356c447865358d19dd56c9e8f823414759c2f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a_cubin_len, 122128, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a", 512, "1bd50d132a976020c8b472ffbe06c04c8541994db6908bb72c1c4c53d4699c71", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -6365,7 +6452,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { }, gemm::SmVersion::Sm100a}, #endif // EXCLUDE_SM_100 #ifndef EXCLUDE_SM_100F -{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x128x256_s6_et128x128_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x128x256_s6_et128x128_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 211864, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x128x256_s6_et128x128_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "a937714d994d7eb7f2a67f608c7a5f61b84727bf6f02e0905bfaf94710a598e2", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x128x256_s6_et128x128_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x128x256_s6_et128x128_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 211864, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x128x256_s6_et128x128_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "fafedd494367620024c1c18a6c6fb239013993791247f720f2f66d8ef7a6eaba", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -6491,7 +6578,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x128x256_s6_et128x128_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x128x256_s6_et128x128_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_rgTma_clmp_dynB_sm100f_cubin_len, 209816, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x128x256_s6_et128x128_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_rgTma_clmp_dynB_sm100f", 384, "a1234531cbc2d51b88d3d8b9ed05f9c5e48d443d71110b7035613da56cbc02b0", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x128x256_s6_et128x128_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x128x256_s6_et128x128_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_rgTma_clmp_dynB_sm100f_cubin_len, 209816, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x128x256_s6_et128x128_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_rgTma_clmp_dynB_sm100f", 384, "ad74af23306a2b1d79803bbf4a46022c4458389287af1be603d2b3cc89ce58dc", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(2) , /* mBlockK */ -1 @@ -6617,7 +6704,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x128x256u2_s6_et128x128_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x128x256u2_s6_et128x128_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 211864, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x128x256u2_s6_et128x128_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "3269239b9f8ef17955867ae1ed5da6a342e8389a3119ae78488a091c5219c7cd", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x128x256u2_s6_et128x128_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x128x256u2_s6_et128x128_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 211864, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x128x256u2_s6_et128x128_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "2beb73d9a71e55593e4c4daadb141c5ca35c2975e3443fd82820e910d05c8deb", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -6743,7 +6830,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 196752, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "9e527f6dd9cddf757bc0fb387ad0e369aba6be0b17a961c1cde9f7161ce8185b", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 196752, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "8b0310f3867003df7a5d920412a5a9fb3433a35e7d6bdf2ae49ede6b94413ff2", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -6869,7 +6956,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 196624, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "40dc42950c24cefdc74415ac415e07385f77d029e3f968a0b29b7eb6bb2b1dbe", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 196624, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "539dcea4c2c6167722ae33702588323952cf60d94c2ac65da162cedb0f5d1d65", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -6995,7 +7082,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 196752, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "6aa49c46ae1d785cbacbcebd799286866e3b1f71128bd4e6b5d9c5ffb8b0d3c7", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 196752, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "5aa6958e80b9e91bcb680e2ef471bbf47782d0e089166095940cb559de50e282", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -7121,7 +7208,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 196624, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "97747defd698fe5c67c9a1508e88180ffa51bba2eab166ec1f924034bd7c401e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 196624, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "89aa1127df7d218a0a316642476ed3112391c165c499d320b2d2334daadc4283", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -7247,7 +7334,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 206744, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "c46b942dc80dad27e0e99b610b07162b203adf5ccfb34f0db3177787ea0da8a6", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 206744, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "f2719ebd1171a7c99941c41faa48e155321895adbf469ae48df7fb322d9beba3", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -7373,7 +7460,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 206360, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "fc528fa511d3e00ebdbafa822540a0e6e0f2ba138b5d54deedbbe43b1269bc55", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 206360, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "598ef16f9cd71b4b8a8ec95b51fef3f4663b2798de80eef904c6de7a5ccd10f5", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -7499,7 +7586,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 206744, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "3d93107643439f1533a0bbb6cefbafd9ff85b2d4de12a9f7dca922aa64a056cc", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 206744, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "96c3f6a598aeed1e97ef3efb7f6ba7d4b6a76d94d208fa414e1cc834039ad04d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -7625,7 +7712,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 206360, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "2856549f44988ddad33810ba27c713a26036fef5dab5362f2d00494220fe1832", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 206360, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "b0a222728676c032c3eb52f680fa63c0a8ce80487f60498f25a6948c1e546477", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -7751,7 +7838,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x256x256_s5x5x5x5x1x1_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x256x256_s5x5x5x5x1x1_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_rgTma_clmp_dynB_sm100f_cubin_len, 219800, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x256x256_s5x5x5x5x1x1_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_rgTma_clmp_dynB_sm100f", 384, "b19c23c1da645d6db5146d84d685f60c6660eef6241449ad8f8a55167bf6e41b", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x256x256_s5x5x5x5x1x1_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x256x256_s5x5x5x5x1x1_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_rgTma_clmp_dynB_sm100f_cubin_len, 219800, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x256x256_s5x5x5x5x1x1_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_rgTma_clmp_dynB_sm100f", 384, "edb79ae72e0a4093cdbcb6694a063eda77d23968d6684da116a4021330f7223b", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -7877,7 +7964,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x256x256_s5x5x5x5x1x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_fCp_tmOv_bM_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x256x256_s5x5x5x5x1x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_fCp_tmOv_bM_rgTma_clmp_dynB_sm100f_cubin_len, 215704, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x256x256_s5x5x5x5x1x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_fCp_tmOv_bM_rgTma_clmp_dynB_sm100f", 384, "ae82e8f732b917cb54d6bf7d98ca2630b634ccbf5eee1ca68767f46f4380a371", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x256x256_s5x5x5x5x1x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_fCp_tmOv_bM_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x256x256_s5x5x5x5x1x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_fCp_tmOv_bM_rgTma_clmp_dynB_sm100f_cubin_len, 215704, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x256x256_s5x5x5x5x1x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_fCp_tmOv_bM_rgTma_clmp_dynB_sm100f", 384, "e3fb16720e8bd5a26abd92bf9511dd002a28bb2ee765dda8ea6c3021e903cf04", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(2) , /* mBlockK */ -1 @@ -8003,7 +8090,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 221328, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "970f2a89428aa5c550a746b54cfc62ef22b12a662b4716b26e24b4a3fec6ad97", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 221328, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "1d9c868a1d4544e52a9cf6074321a8644296491eff577ae46456cf0faba47cea", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -8129,7 +8216,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 221200, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "c4d8067a82a110d34f7b4b172ad9f4dc559e072392f8a8e33d7279f48c91400f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 221200, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "39021d0a924ec02b9d866fb612c777517d9eefd3c7b35d9ea57a9f2a9a612216", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -8255,7 +8342,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 221328, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "5ddcd34ac1ca8c6c67178a3bf209fa424cccf12bc0bedef3be2b5d995d92909b", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 221328, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "c4f34533529509f32f36e2d12080318335e9d89f25b594731e9a815271341697", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -8381,7 +8468,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 221200, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "61826677f58e6051bdfaf73ae663da6b9e10dc2b928c1bb8490d1a9b15b03a28", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 221200, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "a675d1e2c6e2e1c16bbf4d2ffe9de7c24990403cf5b91dc98d0ed36a7d973e1c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -8507,7 +8594,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 223128, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "97571478a6b3d803b4b6cdec98a144e6fd333299e7bb30f7c14a1d53cbc24709", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 223128, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "a2bf254e454ae3dee3b69642e9776f7165d6ba958d7e26d44e9f099b863115c3", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -8633,7 +8720,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 222744, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "18e909d98a5681eaf4f7b90a583a7549d6237548aafde795afe059d729f95e68", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 222744, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "8dc39ff2cfafafdd933ae9b9ce80228a9dd8833c14a0c0e684d002f37026d741", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -8759,7 +8846,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 223128, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "9835859f8e625856adca1d653adbff667e5a3f23c060667f570be2c4b09b569a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 223128, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "5d30523179bb4cdf303c147bf020a59d692becb6eae0b8e231d5ad8d12c539cb", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -8885,7 +8972,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 222744, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "67e981ad966bbc438e979010cdb1e39b28d1ea9c0bf9b61edb470a8076abdd15", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 222744, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "31563f4722714f09a1bc3fd6f59db1cace0bf7bb72959b4fcdbfa0161b00aade", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -9011,7 +9098,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x64x512_s4_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x64x512_s4_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 210584, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x64x512_s4_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "35480670ac9f9f05f2a40cce7036467da7e6b09aa7bd9cd3186bbb7ec21fb05f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x64x512_s4_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x64x512_s4_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 210584, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x64x512_s4_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "09d5939925c0a7adca5e4d955f4148234a1f765e1011db8ebee0657edbb878cb", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -9137,7 +9224,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x64x512u2_s4_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x64x512u2_s4_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 210584, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x64x512u2_s4_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "a6c29be6b1772a5d01708c04f36cafa9625b8e98bccbb7a1ce59d70ae3946aa1", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x64x512u2_s4_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x64x512u2_s4_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 210584, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x64x512u2_s4_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "132ff6e9ca820521a64d202d1923296fe3f3cad2d232a0bd883db4f1fcff647f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -9263,7 +9350,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 184464, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "dadcbdfcd871ee40b1d1a8bb02874a54e01af6808d2cdabacb34bd28f589aba0", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 184464, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "71ff17ffc086b6e91cbb7a06de8012f83ae04f992eb28a85403d05c514df3ba3", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -9389,7 +9476,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 184336, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "d19b8baa2ea768aa1a545d84b09288e1f9b0d50ca5f6767686f7a4fca42028eb", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 184336, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "42e5d30e53df8bf42b4c9daf8a1396510484cbcb4fab2b9814172f6fa4de563e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -9515,7 +9602,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 184464, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "74f4a6fbb9cec92c3730bc474b091c0f0ab960bb239d141e571dcec4dc59ed18", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 184464, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "418ddddc59a930096bbe1bb1ecf0d37a0daabd7bf1d9014aba8a6066c93b95ac", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -9641,7 +9728,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 184336, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "3c9604d0c49a2428b29d1928759018f3a43aea5779e36513571ce4526629487e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 184336, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "52af7c3d24e4d319fe3bea379ac6a94b21211dbf784425de3b0171c0105d0eed", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -9767,7 +9854,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f_cubin_len, 163344, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f", 512, "18e6c258af654bb21bcd3ee7dd56a57440bc851cd3962e31c55312c08a740740", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f_cubin_len, 163344, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f", 512, "09f6533a5a9945d75af67407c1cb59b419166db1600c34bbd7452db44d2b5462", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -9893,7 +9980,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 203664, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "5fb85a2f4cf4a0ea644ad5f5644dac8efc7f0d161db696972ad7db8f2bb5f18a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 203664, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "428c2a95bd066e9414b8be37baceb8ba05aa25e0836a6ea3f42e57b9bb54f488", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -10019,7 +10106,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 203280, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "cbeddea34380981d827339a7b74a63c31d8fb111e0b5c8d948d0f23ffb65a045", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 203280, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "4821d8c24a9957f4218b24c00d2cdc311c2cfe2ade47acadbefb3da477149e53", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -10145,7 +10232,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPdx3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPdx3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 209816, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPdx3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "ce1bedeb6391dca56a5f425f5fa987958162e039276d213a10f07b449b150cf7", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPdx3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPdx3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 209816, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPdx3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "06b856e2d49b322d41fb4c8cf5fee18d874f60bbee037773c1dec1aa0eacc25d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -10271,7 +10358,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPdx3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPdx3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 213912, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPdx3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "a36ea163e919cb2d0673ac22d99f37cc1c096201cfc6b630d2398a58d8fbd160", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPdx3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPdx3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 213912, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPdx3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "6beb5fedf38b88f973f02a5a29b3a3eadc8ab62835f758509f04a294198a0a97", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -10397,7 +10484,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPdx3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPdx3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 218008, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPdx3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "020f440af5828a27f01ef43dd8ab42e209da8481fa16a4a4f1b03fd5c18f6e47", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPdx3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPdx3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 218008, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPdx3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "c691f422bbcfcb35b166e399b0c22b293375491187adb3eb6d03217ba752ab2e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -10523,7 +10610,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f_cubin_len, 163344, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f", 512, "0999d64d09b451a2d56b500acd036ad93ca69c2d2b87c740e42bf4743653f13e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f_cubin_len, 163344, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f", 512, "ad1b8b2977e5af6bba8f33876512cafa0ccb78015e9a6c15d78f2c63b3fddd86", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -10649,7 +10736,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 203664, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "eb14d82ae6fbb552bd01994ffccb44deef33083b059cab26535cfcfa5739ddda", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 203664, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "e6fbb05a9c46c1eed362633ed3f4afc1d209c23ee3e41d51cfcd233bd6eaf3f5", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -10775,7 +10862,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 203280, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "29069189319ef6c2569e22ffb49284750acea524254be1320edb51f81890f8c1", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 203280, "bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "950b9c05a1c47da00859419776ecf60dda9decde1f2f62a698d3310adf3904b5", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -10901,7 +10988,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len, 207024, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f", 512, "c7076a07becce38262bb2bf1ec8c5372717bf40fd309a8a5e1adda00c1fb2eb3", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len, 207024, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f", 512, "3a935e35472aa4717eab35daa9d747755e60bf32caa85c399817d0d745805ee4", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -11027,7 +11114,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len, 206896, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f", 384, "15ab98b956990c81ee07240b644ec12ab4ad83c40b79a955a44136e34e579215", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len, 206896, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f", 384, "2cb67ff07f0aa8998666c8792737e386b9767be18afd5ba89496c842db78c698", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -11153,7 +11240,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len, 186008, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f", 256, "7128edf9fd2f4dd3079e611b0c60e6e5b25853b2a5975c1c82baad0118121bb6", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len, 186008, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f", 256, "e3058e2abae14dcd905a5adf7da1fdd979467d9920865bd7f25df73cbcbffedd", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -11279,7 +11366,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len, 207024, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f", 512, "5a277b429f57e417b12e05a2d1c099c85935bb5200e67374d3453208b2ed9e6f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len, 207024, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f", 512, "564c33b935edf785c8e4d1b8c81e75cf2ce43185af8bab895431c10bde132b85", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -11405,7 +11492,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len, 206896, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f", 384, "e8a33eb0b61711aaf19375b2e05af2350dd7b2ac2606e89fb76527b7ae66a563", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len, 206896, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f", 384, "13421072801d7df1211559e33b8c7461d257cb7056f9875cc1b3bfaab92ba115", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -11531,7 +11618,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128u2_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128u2_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len, 186008, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128u2_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f", 256, "5f247addc9aefed3366663918da1784dc280f41c2ccf8283e12ca068c0535065", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128u2_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128u2_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len, 186008, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128u2_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f", 256, "c7f993d94b241429793e3db1a1ffde989fe206ad1c3d85b7ffa502aa3c65a768", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -11657,7 +11744,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len, 121264, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f", 512, "3bdc7309ac3672e853be5350f2dd2ed5fa6cacee2a94bd76ce683c541db5492d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len, 121264, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f", 512, "0a12b32892a56283c2618647c45bd5abefb6288b55eca67abf95ca5a92c90e90", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -11783,7 +11870,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len, 121136, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f", 384, "063f448ad6b7ce2b273dadacd5192e7616e5aba2949283428d01c315bfef6b19", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len, 121136, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f", 384, "9a5802da11df382f2184bfd134a0bbd7b5476cac76ea528102992abd44d52792", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -11909,7 +11996,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len, 121264, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f", 512, "f3e75c31605f056fad394a33127715c119804b85dbe8c7be9a58dc998d94972d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len, 121264, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f", 512, "916976dffe7f845b06fbd9fdc8b1b244fd679e7e00c668ae0292abcc937c8862", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -12035,7 +12122,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len, 121136, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f", 384, "978130f8eca8c21e1a63f29cc87b2af31a3887e8a82d8a0608546f94af4d9aba", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len, 121136, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f", 384, "1e7809ad23736c687537732f7e33aa51ba27ad040b32c1d29404c2ce8b2e0aa7", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -12161,7 +12248,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len, 230032, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f", 256, "b8cb10a32f8da3c06f9b8c0183e22f31a7b348745222558ef08b3ad921abe25e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len, 230032, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f", 256, "58310956f2163841a798f7bd834f3c09e059730fc295afb9ca7e5ad6d9a2ef0c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -12287,7 +12374,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len, 229648, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f", 256, "41fc31c7c4a1f4ee4ae1e9e90ca2317ea8ede1af2b95e9928af29f8b0493e034", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len, 229648, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f", 256, "51cb26f9f5e11ad2d08cec5a368597133d7a9955e2721d8eff6b978a7243f0c4", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -12413,7 +12500,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len, 230032, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f", 256, "df16da2a0aa0c9499d92f1b12bfb1865fb1a4c34d208fa75d82120190273195c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len, 230032, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f", 256, "8b6ee20b547bf0fad776f9f4b9647bea33f7e5597da9c272b90858969a98438a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -12539,7 +12626,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len, 229648, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f", 256, "cc5bb964f490c1e86ec87a3a5828356ba4af6098ce643358d5d550db54a96f29", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len, 229648, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f", 256, "897288ed4cb8da5cb37436c9bf5fbe49832b3987b409a52a2d717ba75c267b0d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -12665,7 +12752,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len, 215704, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f", 256, "96fcefdb8923f824dacb266a360daafa89a026362b515e06862ec82e57ae6161", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len, 215704, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f", 256, "5423a6fff8f4d47cc2f4963ad52a2e55b05f8fdb20e102e3c55dc2132aab531a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -12791,7 +12878,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len, 215704, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f", 256, "823fc84fa1124a449729875de22a9f3a8d780fb08e0694e3a7b3fe7c55b55438", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len, 215704, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f", 256, "d1d595affca9fc42a59c8c0190deb5f5ec76e9d6d9f6720c3675853c97db8793", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -12917,7 +13004,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x256x128_s6_et128x32_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x256x128_s6_et128x32_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_rgTma_clmp_dynB_sm100f_cubin_len, 220824, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x256x128_s6_et128x32_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_rgTma_clmp_dynB_sm100f", 384, "fc7e65a2f48d463d8f1b33bb8a76f836fbeb5cb408e66bc7ac577c0af6c4e022", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x256x128_s6_et128x32_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x256x128_s6_et128x32_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_rgTma_clmp_dynB_sm100f_cubin_len, 220824, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x256x128_s6_et128x32_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_rgTma_clmp_dynB_sm100f", 384, "795e353955d46203ecdd002225d0648815158a7ebad3a2595694eb390be86040", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -13043,7 +13130,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x256x128u2_s6_et128x32_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x256x128u2_s6_et128x32_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_rgTma_clmp_dynB_sm100f_cubin_len, 220824, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x256x128u2_s6_et128x32_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_rgTma_clmp_dynB_sm100f", 384, "cd4e9fd0ec7a5012bde5d32982ec000aafc6ba2dad0f7c688132d705dd81814e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x256x128u2_s6_et128x32_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x256x128u2_s6_et128x32_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_rgTma_clmp_dynB_sm100f_cubin_len, 220824, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x256x128u2_s6_et128x32_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_rgTma_clmp_dynB_sm100f", 384, "fbc63f2cff033d5b7b9a166cb0a04ea219703b1b4641620763949beb124140f6", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -13169,7 +13256,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len, 137648, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f", 512, "5dad3f69616bdb3866e8275af5cedc50966b3609d87c5eb940acc53ee9d22aa2", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len, 137648, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f", 512, "c5c3dc46522a72249cedfeb5063890572a5eae5c8a6e4f7dc182957992432414", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -13295,7 +13382,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len, 137520, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f", 384, "adbc8fa5ffca70c872173d9344eef2e4860339895fcb0ecc4f333075b367276a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len, 137520, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f", 384, "1deeba0da8be084d0cdbc1d50a6db4cc8f34e5d5da10cee8f18987a21929fae1", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -13421,7 +13508,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len, 137648, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f", 512, "9092d9f971d793b5d4334f70a3c6c42a7b52593715b9f71768d1fdcf3554b8ca", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len, 137648, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f", 512, "c3f650e584cd1aab8e59c15963d387bae54480232240e2b1b54b286c39be9eb4", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -13547,7 +13634,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len, 137520, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f", 384, "1f79b16012bc62aa3540a52e86809ab248eb38699d5680ead9cb18222e8c0736", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len, 137520, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f", 384, "f938e14911db9a1f9a87f20f20732a34f4bf85a8464ad2921157d5c69decec41", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -13673,7 +13760,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len, 217744, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f", 256, "123f8e4e64460410f5626f4e1666555687ca3708c9df505a0b2d5c77e8de7707", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len, 217744, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f", 256, "4fad86c71ba51df82521bce87b6c4f9829f825ac39678303a73e1b946e181470", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -13799,7 +13886,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len, 217360, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f", 256, "73b74b3e2ef44c643cff9751a519834ff76f0f0e53b26d883dc4b92df016960d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len, 217360, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f", 256, "ac7b004ec433ad81f6d0712482ca755f3ae06a75242a2daa3b39b87d4629f26e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -13925,7 +14012,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len, 217744, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f", 256, "cbe65be83568be1a3efeca3c64a4c39b60f1f14775a2406185828bb0b882a6fb", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len, 217744, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f", 256, "13460b4a526608042cb826fec734623b061a552599552d31d5a19ff36fa80fbf", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -14051,7 +14138,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len, 217360, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f", 256, "8b523256bd34c4db92dfecc924ba4a86f689cef40e4e977ca38031d32cc765b8", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len, 217360, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f", 256, "f4d86b7c19f2a76504f34ca9f1af64fb52142ebc312c0ba824557a101e614aa2", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -14177,7 +14264,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len, 172208, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f", 512, "ecad6e81b5296824f6b69f5b3c9a500b9d88c9ea5f8487df417454c7b7c9bfae", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len, 172208, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f", 512, "fa2e2c5caa91f1ec13b92dfb5c99e26792578f371d6564d8a43f5c1ac54ad9c9", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -14303,7 +14390,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len, 172080, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f", 384, "412e8d953f355174014bd67340fcdf4d9e272d281264344237a39f379d25f51f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len, 172080, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f", 384, "d9795a2b91c78e0f3a725fa491e81f427f1899af77f8df47d706a25e51f9f777", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -14429,7 +14516,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128_s8_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128_s8_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len, 184984, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128_s8_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f", 256, "a0fbbe73d9fde39161bce440dd2c8ce7ca567254e7cfbce64c96f1242092a9c0", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128_s8_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128_s8_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len, 184984, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128_s8_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f", 256, "930d8bf284deaf1c608284e3eda095938fd6df1e3eec26800d1473691c84a135", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -14555,7 +14642,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len, 172208, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f", 512, "c5bd27a6f77ee9aee22e127672bd368498e256d0830d65b77f0b9da5dd0356e1", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len, 172208, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f", 512, "a82c5354f10fdba41c2b347a836ff44a9e15100cd699efdbeb81aa93cc6c69c3", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -14681,7 +14768,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len, 172080, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f", 384, "1df246bc938e50dd2139d56137eac86d65b1f16342e75843e631c3623bca268e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len, 172080, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f", 384, "80c1da5534a646025e7efc3fcb979e3b3afb84b4e9630717dac8fec8c5f2bfc3", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -14807,7 +14894,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128u2_s8_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128u2_s8_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len, 184984, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128u2_s8_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f", 256, "d093cc5e98a84a4b3e44bc970fa5e0046b8abd59bb968c8ba67cb255d5c753c5", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128u2_s8_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128u2_s8_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len, 184984, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128u2_s8_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f", 256, "2160d53f0629f4c6562aad6cefab6008b9b005eb631170fd1dd47ef829ff88c6", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -14933,7 +15020,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f_cubin_len, 77600, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f", 384, "f13e492a388eae9128c7c8157ad2934e01510e61e0c828717d89dfceb533cf43", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f_cubin_len, 77600, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f", 384, "e7728ea79ff10b528772b95f2cd8251aca49f619da1d39eee6b5020ac223f474", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -15059,7 +15146,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len, 148400, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f", 512, "86d1dc59a81f8e1562d76bf25414224e5263782be23a93a1ead9f9a03b8c3512", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len, 148400, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f", 512, "ff1df77eb3e6cea1d2468eed1f0457d57572f1335ddf732a4893b99947d258ae", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -15185,7 +15272,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len, 148016, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f", 384, "4fbb0ef2471b2ce9a1cb7b7da27e94fee9064893d0df03ee93eab12817904271", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len, 148016, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f", 384, "318824a74309c7b85c1078601a68013b86261e002095e91bc4132b9d22ca2623", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -15311,7 +15398,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128u2_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128u2_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f_cubin_len, 77600, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128u2_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f", 384, "3b4fe97a110b534bf76c21e7bcced520c5869a589cc1f58820fb3ba499fbff24", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128u2_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128u2_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f_cubin_len, 77600, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128u2_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f", 384, "decc8227979a5dd6a6aecd21f2d3924c9b1ddb6a6973b4075c35f3c46c0996a9", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -15437,7 +15524,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len, 148400, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f", 512, "9c69aef2a3b75fe49487be1b9fdbdc3af022ba6ed6d718c3246265b155718f0f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len, 148400, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f", 512, "0a24d6c95ec837f7857a6e5ee46e679e29db28373c0489ccd8a93227d8adbc1b", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -15563,7 +15650,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len, 148016, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f", 384, "402ee0070a0f7f72b24b565de39142673957368e8f285d58b3d252e1fbc346c8", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len, 148016, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f", 384, "0a613fce68874588413271e5ecaf1808889a8ca7674f62f92fad6c810a7a4489", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -15689,7 +15776,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len, 215696, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f", 256, "9f8dd9abadfd78f4c73a894d53ba1ea9c43baaf264297fc3f645d6e62b9d43c3", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len, 215696, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f", 256, "b780212256c07a6e5ce4e50d362f7b004f23a03d62bab035515895a55b52b8ae", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -15815,7 +15902,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len, 215312, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f", 256, "7d74759953228b33013cc809f620a6b9501abdc6b3644b4afa12c278f7e64550", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len, 215312, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f", 256, "834670bb14b7a73b4ed461e32a9556068712a2cf0a718c21297fb65175b5af39", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -15941,7 +16028,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len, 215696, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f", 256, "c78c311d7bff9aad6c6e3f1e5edccbecde9c0a7ec075af298f7a04ba28357eed", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f_cubin_len, 215696, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f", 256, "135d764013026a8d610bf960c1f33871eff117aab6597503589bcb7358e2915b", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -16067,7 +16154,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len, 215312, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f", 256, "0fe5d9694eb04ca489b916af2227c0c9b823ef6e7825ea8259ed7094b0915555", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f_cubin_len, 215312, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f", 256, "3205759d92dd71f002197d13dfa62baa0a962eb19061848b48fb5df027d80b6c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -16193,7 +16280,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f_cubin_len, 215312, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f", 256, "d0f8da72cc009b09b440bc5739a69ceaa82eeef1b1efab8b061a14ec034bfc88", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f_cubin_len, 215312, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f", 256, "a2659203144ba075acda5c604c6b66521f6d1f49c318bd5d5d5d7bdd74f2e0cd", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -16319,7 +16406,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f_cubin_len, 215312, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f", 256, "5f4a91492fbe02049074f669da1031f0163326afa11dd3cf7e32c5eed43220c8", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f_cubin_len, 215312, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f", 256, "c686abfea717d0dc4f0ad8d65fceaf1722a3f0cbbcb8b9ee237be401620c1685", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -16445,7 +16532,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f_cubin_len, 215312, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f", 256, "bb1b7c893b276c17111879b59c88f0cf8b277da1b8053f52d86307f3736b93c0", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f_cubin_len, 215312, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f", 256, "6f73070530851acac5442a8be993a165a163198b1884474d39cf693a1ed9f701", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -16571,7 +16658,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f_cubin_len, 215312, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f", 256, "871632efcd4f975a16251bf51e9fe93f14cf9ec1856e9039106f81edde3ca50d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f_cubin, Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f_cubin_len, 215312, "bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f", 256, "8f64c3358da0e58b15db2483740dd965b367eef4048ccffe2c5f3d6b931683d6", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -16697,7 +16784,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 73360, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "8155544a5404b9347d9222113c0124d74318e7faf9cfa4cd3f9d5e8150f61b07", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 73360, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "cf467872fc1e83b61637c02d32ec23706eae6afdb747100df32470beba98984b", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -16823,7 +16910,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 73232, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "f7b979cca3e86e66d3f7b47d61f7ded7bce8d6192e0604477d333019cb072c88", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 73232, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "faf95fb127631cc713b99de2aec98433961b647638fb5731682edc3560a35470", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -16949,7 +17036,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128u2_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128u2_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 73360, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128u2_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "ae623552d7edd0359eb2164cb7f90850d653e416db2f87efab13d42d1f523126", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128u2_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128u2_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 73360, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128u2_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "a930a4247ed1e43f00f9a0cb3006e452c48393291991f0b96255adbc7b97c4ae", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -17075,7 +17162,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128u2_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128u2_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 73232, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128u2_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "05e19f6bab9609b1f39da63fa1c85362b9c777fd11ce1fdf6a7563892e784a2f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128u2_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128u2_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 73232, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128u2_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "b45d616bbc6b244018b279c51e0a3c79ff4072725cef878ebe50bd5ce0d2e554", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -17201,7 +17288,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 85648, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "9c9575fb25aef17bbb0bc48aa218ac5c8d44cbab0f50a49e40ff7bf408ca2312", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 85648, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "7279e4ede9c7dc8d54ba7c974746b0e8a0b6cd2c6ce21cdd6f0b6ce5ce87331f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -17327,7 +17414,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 85648, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "c53987bd8fbcb51fac898808fbd942378c4a04aa2bd264c5c0b738c27f2902e7", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 85648, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "b30e93ba4f1567fc8471cb2f0587de2ccaa5698936111fbd108cfe944c9dd9ae", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -17453,7 +17540,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 85264, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "2d7e836ad7d69a994058c8a9c2616a34e1f043f053539a210825ab2e90f0c118", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 85264, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "a67fa841ae687dfd9b6a63b652ad8bf2000f4cb0e9ae8381d372a48d4788fa7e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -17579,7 +17666,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 85264, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "55246a68d69c49a705bfd786d35c179cf34909b652e49adf0ce00c39d4d8fbaa", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 85264, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "ac0047fbf00295006293082bb79a13fd9ef06b4610768ee85c01bffb79aa330b", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -17705,7 +17792,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 85648, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "7eda59dc95153b78b261b213517f6324c9f64eb74e980add148cc1b665712b99", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 85648, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "f6a07cd84a4a9891817a1a6980a578ac710832229a3f02fabe254c018ba4d91c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -17831,7 +17918,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 85648, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "73ed5fad6dca98def881df3294ce03fd5a2144aa6196c13b40daff931739db2f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 85648, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "44dcd22d7000caee267888cfa0925047fcd80d24f6c18cc0e7a1d8158a441e90", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -17957,7 +18044,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 85264, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "381324bfe07f0a90efe932008c44e7c964bfb825a5f6a0999524921ebbe74aa8", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 85264, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "a4f1ac205050bcf12fd71f9770c69be0561a41a7c09268c18381222cefcf25b1", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -18083,7 +18170,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 85264, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "ceb014566d6e415152dca433499c7bb8ff23dae899ccd622829084e20d27874b", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 85264, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "c006cd1b8b412aeb2296237b741445dcc954e16ec2843cd9edf9cab27222bdbe", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -18209,7 +18296,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 97936, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "6196af0d95f7581c85052aa556b4fd4cf06ae35cd4cb17f651caad20a05eb0aa", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 97936, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "355a2a7a5682e6818a9104c0531a8510d338c5c56faca7e48395c4231a5cc4b1", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -18335,7 +18422,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 97808, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "bf0249e2b97c4f96888c902b015eebfd15ca78e846fc0ad687fac2a8daf1df05", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 97808, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "3089f3a3afba815d7d902589bceae14802f6cdb39f58ab3456019d492c471719", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -18461,7 +18548,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128u2_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128u2_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 97936, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128u2_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "372b5ed01052af0eb8b8c7780deef1d915b0dd6d413bc8ae23e96d9e5583263f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128u2_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128u2_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 97936, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128u2_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "4b2418c2962bf9f522f8fe5f8f05336ff6121d77e4d69f5482538de5cd04ef77", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -18587,7 +18674,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128u2_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128u2_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 97808, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128u2_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "ac259a3feaee0830a400aa0a61b718069a0f7441f364ef24686c370cdf5f3500", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128u2_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128u2_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 97808, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128u2_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "f460fd508e5250a2180b4a7171e4360cc841bc3b18cbe5d42af88de876baf6c4", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -18713,7 +18800,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 114320, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "1c37123698499cf0914588fe7d287dc3985fad556ce7b0057d304f4d0ddd64fb", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 114320, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "0721e158c88acbca6ddf89e94185ed2f0b4b80953c424434a9ab5805b178c3b4", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -18839,7 +18926,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 114320, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "e4c6c9437626ff54ddd4b61823d585eb14296a6be6b8f873e12bffedfaa49e08", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 114320, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "7a5376005afab0109fde5029429c10a34f9469c1b61ffa30bc7c41c766fe0275", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -18965,7 +19052,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 113936, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "82a473b1494e512b666f3683175dbb7a1b6aa8236d767b1820a8a8f6dd8cc1f2", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 113936, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "1117a919074858cd6ac35639f5878555687019131fcd9cd229545301da9d4ad3", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -19091,7 +19178,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 113936, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "1c3e601b4cebd1cc0e8c6ea1d7450a413108a07f4f01be89161ac54773b12f94", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 113936, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "4169b4a28a6fafc40e33e44f66f12f6323e7c5dd7b23fc3efeab3e5de7df3ef2", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -19217,7 +19304,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 114320, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "ddc17d1506b0d4f4ef2c7891f6c9bd12023a302d555f011deff5c4ccb4339c6b", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 114320, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "c8fb1344570dd5382760ba82367b77718ad966a52a91e0b82ee86fdfd3f1d097", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -19343,7 +19430,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 114320, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "fc5b93748711eb953296b599e1b73cbdf1a9a680bf89fe9e4fe1a758ea554db4", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 114320, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "86adb7e49cbce481e24fbd073a2c666e6149345c743609a311b3b56e3571d058", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -19469,7 +19556,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 113936, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "3fa3039e7dc64b9fffffac9bc4b580364c786f4642044fc5a241c126de21753d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 113936, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "67a02de346d180f14ab91f6ea5eeececbece22e6e5c31016390f18bb9fd45a0b", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -19595,7 +19682,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 113936, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "74de757c9f7e73169519b557ba72de0f17742ff845b451ea2c18ca1d9e2a082d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 113936, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "b1434c7742de9a30574dd425ec93e552081f220bfd251580247264849abe4cff", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -19721,7 +19808,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 148112, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "69b2370bd2aeef420b5f0a197610c47fe5032f42983d302ffe2739472d337188", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 148112, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "6c44f78e9dc528390c362d1e1c75043909a5d5688784e223704bb2a5a6f9d20f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -19847,7 +19934,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 147984, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "0dbc78db5d4d7a5ea65bbc394d6c0c1b19207b6847836e0bb8f6c27d53d9e93a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 147984, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "20f3f5347923dab6f7ca341be7c2e586d7aeacb877bbd87e3fd906f286f15611", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -19973,7 +20060,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128u2_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128u2_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 148112, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128u2_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "a3873db946b227b9d52bcb6a94066c1bc41c46274857c983c4f69f9c31bc7314", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128u2_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128u2_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 148112, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128u2_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "9d18b1df681a1d5a6e790214110509cd224d6aaf230c21e1d6363b1fa253506d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -20099,7 +20186,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128u2_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128u2_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 147984, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128u2_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "8621e794901da057609e5403daaf1773a4824eb9e27225d7a5132e47aac14276", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128u2_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128u2_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 147984, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128u2_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "2db7db48b6126d323896d9aa4ef3c65da70e30e405de17e6f8b1c1c12ed46b93", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -20225,7 +20312,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 172688, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "b03fb8f83a7e7af9d003dfe59b8053190fde8d54c0e55aa50b3968ef1f672aec", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 172688, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "474010c19cbc8cbc397b50ebe2f4288aec0992436a5d966bb277453cc851645b", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -20351,7 +20438,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 172688, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "1c743b6b0917bcbf6e17850ce1c9b77a3f193e26e11f5355afc7be2b60912d9e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 172688, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "5b0ea160cfd46e5d8c22c393aa7d702f29e250f5ad3d85eff803d6d2ba6dfb69", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -20477,7 +20564,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 172304, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "ceec67dd522dbec05896349ad717e20899af8258675d9f45b6ef175203517686", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 172304, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "47a6aaae43c4ae4c703bb7ce23fcc611c54b52fa02ab720391eb284bf5580922", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -20603,7 +20690,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 172304, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "5ab61542f6fd0fc3d4e66bedd205adb3347af5e6b7dbb532e241319565ea199f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 172304, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "3fa6f9439a44d36a9f759d20f69521638231d25b6dc4eba7576aad3cb816214c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -20729,7 +20816,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 172688, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "939a16cfcd63913d99e60fa38a77ea0b558e81cddaa27b5e8783a44174335005", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 172688, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "0e0e3d220e7881359670ddcea945e92305cdabce78222ad88d2b942c810fea08", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -20855,7 +20942,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 172688, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "277e2e34995314a3aa82446e285ca6b9796daf3b257418ceabf8ce9d85df5e2c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 172688, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "4e24c775cbbaf81ca22cfef43faa74f0b0f953ed932643564c91b37830cae28d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -20981,7 +21068,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 172304, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "9fc3083ae35ace85c9b7168cb8eebe7a3fe21dfda58dc7c67be4e691c128e9d5", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 172304, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "4b9b569ff93513f45065895a30595316f4cc781600718440dc1e1720cf83f554", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -21107,7 +21194,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 172304, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "d32f9270035b983dcea1c341f16528f0ab6bbdccc373687a273861c86f6a06fa", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 172304, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "57eb96589ef1f240dcba7b9d6d24013f676fb123a3af6f48b022e2679c0002f6", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -21233,7 +21320,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 61072, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "cbc52c49c31203e700a6a270ef2ab70e61a7b640840e2c357a58f5168513a514", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 61072, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "1c291dcdbefee98171c4b6cface5dc6042401a10201e5f4786b33d03806496df", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -21359,7 +21446,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 60944, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "c2117a13246a6d06f495b53eaf841afce884033088413ae55fe6a89758edbfd4", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 60944, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "260b7e7dd0c8698874143cf9bf94c9eace04843c0f4fa455d8c19f9a9221717a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -21485,7 +21572,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128u2_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128u2_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 61072, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128u2_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "50c64cc351b55f77dbdcd6258b30818028c23e1467940cf2a377d311e86074d5", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128u2_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128u2_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 61072, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128u2_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "16a6b258a9f3feb218a89976c35c6f3906ff8cda308ecf6755e5cf26f0b3498f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -21611,7 +21698,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128u2_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128u2_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 60944, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128u2_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "0f87305e47db77e2deb5ede7a708f7cdaece04cedbeccf8c57ff9c868ecee2c6", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128u2_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128u2_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 60944, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128u2_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "e32f2a3d0de42a4ea46ca8b0bcdcafc0089ec745d629d3f6e33421cf87647d79", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -21737,7 +21824,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 71312, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "4692c832bd1df78af88c6c5e05b8c88cf004a585e355ae2215045627ee7b5818", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 71312, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "6be16c0e3c4c96ac449e132078665b65ac198c396ce7626d4cd09ab742ae8cb0", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -21863,7 +21950,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 71312, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "a846c94052b732e65e766b62dce7e3f5f8b34dbc6e7024956591ccf6e870571e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 71312, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "47e03afdcac31a6318a68cd9922f78d36d90bda101379628a56ea156e371f152", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -21989,7 +22076,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 70928, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "b7c9ca44c8565bfc4878a22b0836caad3c77154e20869f1ba1ce5546ffcc7895", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 70928, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "33fe180816b2a8f1fed712a5649b0aff7f22d745b0a3bc25ccc30c6329eaf688", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -22115,7 +22202,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 70928, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "71aa86d00c4109f2c1898e53df0a23db4374ec983586ecb84151648e998b118d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 70928, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "ce44069287c06a280abebf3c98b2b14a17b11ad08d3ce2c2156602344adaea37", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -22241,7 +22328,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 71312, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "59be3d62b3fee330d59883eb8ed69d0f81ecbf0e2bcb48035187cdc749fa09b2", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 71312, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "d06244633aca2565eb100ebe3cc02033a284770b540e3ef980e32b22720a1309", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -22367,7 +22454,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 71312, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "754d23850a2187f06f7b970084e111044dd18c3a67e2ece83794d0cc8d5ec098", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 71312, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "315610868bb672e8d7d21ade2162f16ae3acb3deec0830fcf9e6f5c4cd34ef8f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -22493,7 +22580,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 70928, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "d014728ad17d0ef3dcc9cc12f1b911a9cbd1ad1f588be9fac9a909ca0ec78c6c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 70928, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "21e1c459fb5cceb3f48c03a326cf72efd64f65bc3cbee52ce8fa9f5916306864", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -22619,7 +22706,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 70928, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "aa1dd791baf0417ac058380060a6d80a19fd2d60764b1bb7d65a9fc9fab17e33", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 70928, "bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "902de8c22dff47a1aa14b30b8edb65d38c4f46f218e3d1027c72a212f173378f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -22745,7 +22832,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 199312, "bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "0fece5d6f509a8d23732bd1b72749d9335219122be976c0c2115129824b87eb2", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 199312, "bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "cd6b9dc2ca41ca2ed7f0d37bcdc75985ef2eb40ab74e2e0b48922a01eb34efef", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -22871,7 +22958,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 199184, "bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "5d9a0cb1d8be81b95293ca8fe6b051c3ff25448d071d88217eb5c8e321f84048", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 199184, "bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "38b5700481af972dbaa9de3f8bc15e16d296a28da9fcb53ca8d299419e4776cb", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -22997,7 +23084,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 199312, "bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "4df694a67874b6f3fe70fa3af962bb3f4aff3b358646f0693af1cdc865b23b3b", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 199312, "bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "cd5055fc6fd4f552795ad2777340f3655032bd183b2f908ce191ebc341d990b2", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -23123,7 +23210,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 199184, "bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "41b1a6b2fbd0b259bc45418b2e830238d1c098cb5aef835460cb0119b390d84e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 199184, "bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "865e004a0be6935ba993aa7bae8b2f09bf8625be847dbdeeb182161486748722", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -23249,7 +23336,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 223888, "bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "e8032168d0443fa360c2d035bba29e25f0fd30f57bcaaef99ea902fe9d0f3774", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 223888, "bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "de49e06678e8df42201c47b65264a0030c737a4b14f16b98631af78d770955c7", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -23375,7 +23462,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 223760, "bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "303214d601ba22b816d76bd3eed255cf17248025dfbb3202700591e86ec642b5", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 223760, "bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "fefeb2dd82033029135c117d5d29e37e5a46daf894e2da8b9261208a78c43079", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -23501,7 +23588,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 223888, "bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "9c526f23697b2f36ee00a72167eeccfd82f76474938514fff86bc499de894502", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 223888, "bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "f900a8c9af97be0e2d19504cd208bc5c09bb9ceb98b237c28fd28f190a340955", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -23627,7 +23714,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 223760, "bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "efd919aac9d0bfad0763c822b4816be0c113bf7bcacb941077c3274ac6ed7ae7", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 223760, "bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "677ee1ed43eb0b9b931048bd109c5c6822eb6dbc2800ae4c27b88e0706fc3b38", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -23753,7 +23840,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 222864, "bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "752ecbfa4180314bf8980137341dbd4df672de072a8770ebab578ce1886595b6", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 222864, "bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "97037c2e7512a530134e5d3487779d6d84b7f56daf7ac1096effd63ebe5d2c95", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -23879,7 +23966,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 222736, "bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "e6b24c65a1f13b932131b92edce42430790840432732462c9e9f5257cec476ec", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 222736, "bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "f7d8fe3519ccf7724f65538c796893f8fb2df99446b1f30382642b44fe5b8517", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -24005,7 +24092,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 222864, "bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "79f312523d5fd0a446821a14e4fdf8e7ce12cb1335220db769440b0e00dbc7df", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 222864, "bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "2ec8b2e49ad3bbc1a87f0f3405f1f47967bcf5490bb75cffd8bbf1a2120f632b", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -24131,7 +24218,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 222736, "bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "d4d78dccca24e96674512f431eb0826605456b69e13e8aa2b483d33345b28e58", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 222736, "bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "fd8675b28c1287fde20beadcfc3c3ba98d1cac0a2550c2d629bef46e9249d4e5", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -24257,7 +24344,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 187024, "bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "7846842f19e8a0f8fc85ad57196c0761030f839e6f4b1b19025e3ae00d11d952", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 187024, "bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "c659628ef522ce4eb643c462786e3a695b5cf39652b0e8bc0a71f0f35acbe85f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -24383,7 +24470,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 186896, "bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "25b6c7f2ac8d2d44a54b24f8817831cfeaba1e5d4b33ea2b4f1090e2879403a6", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 186896, "bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "9dd5b3172991a08cba9fffa6889eb1a53e1732e87879f70a8d6b8edfa727bb95", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -24509,7 +24596,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 187024, "bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "809453ee91c1e234f1549a4f9135f2486bddb093e6bdf840fd2bc35593efacd3", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 187024, "bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "bbcc1e57aa1ebd4ce5cd876af4bb0bf41db40233db56d5b07625a3edd10002d9", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -24635,7 +24722,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 186896, "bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "f8dd75e6b16eec6ff9fcf939dd7e081a416900d3b8d30b027ef0ea01dfe078a6", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 186896, "bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "406f27282e80c6f883151b3e76c6747c7db69125b9d1b7160ab78e95189bac05", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -24761,7 +24848,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 222864, "bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "52e54012399a798c495fe2db4799bbe0e99027757fb17ff906ddfd5fdeea3eee", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 222864, "bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "e195737fd1e14c00cd127667914960496f657f488d0abd475b760d11e2e00fc7", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -24887,7 +24974,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 222480, "bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "b8c4647ce371c232b919af211c7e9382d154f9c1623a67cb679fadf837d31384", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 222480, "bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "78b84a18dc04e71121e560f4010f80d9fa64ebe483cf69cfd3993c0c96b776e5", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -25013,7 +25100,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 222864, "bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "967becb2177e400001a46507a544c8b1154570c5bd50e6fbc721177b819737fc", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 222864, "bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "03c71c98769eab9bfbfd9741e2bf094d6260745803b4758fa6407dfa29463e60", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -25139,7 +25226,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 222480, "bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "4a2bd8ba8e59a0aeb406c3571021e80495b1b788a1320b1ca8d94a5df34c106b", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 222480, "bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "86e6159cbf65ebd309c9d9cb32ec9ff5e1d3a7ce22decf711c26ad4974f68891", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -25265,7 +25352,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x128_s7_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x128_s7_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 203672, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x128_s7_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "070afe2bbd0b267f1cae24aa8ed2f7884c9ba230a24edde8371b6cf557a2ce04", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x128_s7_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x128_s7_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 203672, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x128_s7_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "1a2681bb46ff611a73a60a3e45348d42c015339ba02c9359d3a61cc35d85583d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -25391,7 +25478,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x128u2_s7_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x128u2_s7_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 203672, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x128u2_s7_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "4b069cc38bbe35c38d7daf5e566a8b54475879b34aa1af6e8971dd5ffac212fa", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x128u2_s7_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x128u2_s7_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 203672, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x128u2_s7_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "52426d0f8a845ae875dd2ed131b5c4d1d2e96ca136ae0eedce20a5b696993662", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -25517,7 +25604,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x256_s4_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x256_s4_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 227992, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x256_s4_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "8e75d79fff0a703862b424845fd73efe70a9841592eca697da98a0a4e795da2b", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x256_s4_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x256_s4_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 227992, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x256_s4_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "5546d5eb8a5d7d9ed03996a2ac9e5ef95ce6d13b4bfaf7614f29df57e8b0b6ff", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -25643,7 +25730,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x256u2_s4_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x256u2_s4_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 227992, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x256u2_s4_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "ce75f0426d8fdd04e3e84a5054e7272f9fc1c5e16f3ac14d67031543d0b46306", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x256u2_s4_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x256u2_s4_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 227992, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x256u2_s4_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 384, "6934b12fb2b62863e975d8da6723a0e28f678eaab40da477cd12891fe550ab74", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -25769,7 +25856,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 199568, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "766084e4b4325e0d531c9980418b0c23bd9a3e0c8b9e5833840ec7e01c240498", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 199568, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "ad1698ebd29f1b4f63e667287252702870e9ba4504a972d73f6674784a7cedf3", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -25895,7 +25982,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 199184, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "0784dbcc594d03700e334cd29f3a8f1e683bc262650e887b77818b63379d3340", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 199184, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "a76045455f48c6cf38dea8137c617902d937ac88e6c42db911503c63695a2c3e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -26021,7 +26108,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 225176, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "5305503fd91f80b270f9cb379ad50a3d34c0dc6dd3f23be9fc699b42be441531", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 225176, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "adae26099f4810bdc6b63fab717737d379591bc03fd324dd2b004c87fa8d5ab6", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -26147,7 +26234,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 225048, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "4240c86f6a715de3b5a7449530faed7b247890b7772c625d4914dc6dafb07cde", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 225048, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "5ef5697c63191d7501a7b650cfa1a879a3f297526281c5db295604734e5c934d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -26273,7 +26360,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 199568, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "2741aefca2063a5260988819b60b3120b4c60cc3fae9694752f3308ac4f06843", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 199568, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "91e0a2f712fc325701d2f00f66ee9d2545d0967bf54f04098722e7432dc3392c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -26399,7 +26486,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 199184, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "7a6f797718e4e568d6fae1b968d7145c084409d7d177276993b8d122c4ab69e7", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 199184, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "112b7219aca0794c8671d4659988785c876e9741f525453af4bf48d58c7a30e1", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -26525,7 +26612,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 225176, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "808ab62b00db1333027c7318b4af19019d27102cccd338d72003c444f177f886", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 225176, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "6d42e8989862b8d4a0ecb9ff7366456215af7162ff18dfd2954fc04f033d983d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -26651,7 +26738,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 225048, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "8347d3631183c4953125c91dfb536629e5185faf2a8949e4632a314b20a73009", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 225048, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "efee38684748ca1e047191523f1c0b106d962b352a60c426f34d99a1de7cefa1", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -26777,7 +26864,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x256x128_s5x5x5x5x1x1_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x256x128_s5x5x5x5x1x1_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_rgTma_clmp_dynB_sm100f_cubin_len, 197272, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x256x128_s5x5x5x5x1x1_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_rgTma_clmp_dynB_sm100f", 384, "6a69ef3d32e12702fff7f7cc59f3f351db5c75ad041be6250c88769d505c47fd", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x256x128_s5x5x5x5x1x1_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x256x128_s5x5x5x5x1x1_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_rgTma_clmp_dynB_sm100f_cubin_len, 197272, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x256x128_s5x5x5x5x1x1_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_rgTma_clmp_dynB_sm100f", 384, "d08c3bf7d745b20af7bc474a4eeea580519c4aa7a55a7b17d340fca7cbf8f76d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -26903,7 +26990,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 225168, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "b82e9d73de3da8b0637bf2d9b65352ec7f6461143b6e27e84b5596ab9c71f628", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 225168, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "aad93289747ebc2cac843d3dd75b09b9eef4b648b5bda4b44e63abddf9fcfedc", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -27029,7 +27116,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 224784, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "0073994dbb06b866e0b84394c124df8b81e54958143f966a3b701aeeaa5e65a2", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 224784, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "5e072775dc4d55f620b18bbab69d718cbab2521827ce6845a6f796a94ed01283", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -27155,7 +27242,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 204696, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "2ab3f64cdfafe899103ee5fcba1ec7de6e07987dcd504a1e072aaa9258b07ae9", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 204696, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "afd2392bb554d928a33bfa5885cff35f211e14d314f16e8b9924b28dd43722f6", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -27281,7 +27368,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 204312, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "2247acbfa1966fd0e317aad06a76724f408acd8bab63794a65884eaf3689964c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 204312, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "20619360d54578c126b83b76b28cd79d86daad3dd00e9444cf204431407589ab", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -27407,7 +27494,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 225168, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "42af25cc67a24d4eef5558b3a45dd1eeaaaf07d1e9fd6e19a1c9d2b32bc844c7", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 225168, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "30465e6012d83968f34481c2f9f1395663a52b09977de4a9ac20a7e368b1a17c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -27533,7 +27620,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 224784, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "05b41b78aca20e1e40716ccd9194e796331c0ee2e59c869db4f58ff7596eb96e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 224784, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "e24a7a5502ce50cd4f7b6d15cc0e9356b637f27dd7a42e15497e16ad31990345", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -27659,7 +27746,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 204696, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "037254fa03d539fea3355e2d65f2980e1d8c41d1439c50c999e7cb68f121adc1", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 204696, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "c178882383b437f9bf026a31f055abe628c1416c2e171bcdc3ec38208c9b6c72", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -27785,7 +27872,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 204312, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "6d200fd364a2aa515a18e3fdc6e1a533530cee1e96320a459170f9ca2437de77", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 204312, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "cbc5136c7ea55dd48fc3a161c3ee4f4707607453ff9131c048f2d0992e68864a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -27911,7 +27998,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x128_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x128_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 172184, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x128_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "c82fd85a8857028b6ca5900f23b5bd02e3e78332bbb86f5f0df88c33772087be", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x128_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x128_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 172184, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x128_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "1ed6730ee190c336fb2b286ac94287597ec19cdcbd4aa6fe84924d1ad8077b34", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -28037,7 +28124,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x128u2_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x128u2_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 172184, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x128u2_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "40fba5048ce64389188520de43993ae8154aee079bfa7265212ad3f06f4bf9fa", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x128u2_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x128u2_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 172184, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x128u2_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "bd309d726e39a7ee14e634a27978068a17843a8707f24a7584a5ba43b78dc123", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -28163,7 +28250,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 192152, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "ffc9e87aa41ba330df7fe046d78fe2b6491718aab7e07335cd7da4929c109992", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 192152, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "a70e044edbf53c5fa79eb5eddca8f511de5605ff717c46d2597347edbe31b2ba", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -28289,7 +28376,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 192152, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "584096941de4322c66db15e04b305d540d96eb561421a9a7821c79ff3e531a3c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 192152, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "25a6076496ed919f276efffda3e8846da648e981f6ca7a09cc1ba2b823cac893", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -28415,7 +28502,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 223120, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "089ef7628c51440444d43c32289fb25012e4f148addc1321d75c7aa7b97bcbe1", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 223120, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "328d8a76f0cb46aa68a1d36943bbe3d2a1807a78d584b337ce6090147c4e087b", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -28541,7 +28628,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 222992, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "c77fecd8c9fb42b38714597c4912bc2dfca702067f0e23af5107f4fb8fc3c8cb", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 222992, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "9009f3112e1dcc5f608147abae04e2b6c6254f8ed55e17378dcea83f6a31cdc9", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -28667,7 +28754,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 229272, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "85f0e4e97a768cdb9334e11806946f8c44d401ce1507952f848637fd2e5a8038", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 229272, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "abbb480de12b6ec9a669529fb7c1935a717ae06b2bdf0e0a98ce162bf263d916", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -28793,7 +28880,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 220952, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "b3d6f9ccdd1374455ceb9c9b15dd32783eb0d13dc4a71548a44e3bc04b12f144", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 220952, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "9bca090c86b7d164aca73e7875c444a7374a1e2640425da00f5310cfdb271e72", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -28919,7 +29006,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 223120, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "c29b7e964315c40deeec4d62b1cc2e2686b8717353b0dc47dfa4100a1e7b4014", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 223120, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "9c7c9ba1c719085d98088e454f4beb2b8e7c80ea5126e1f23fe51be1beb3b3ac", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -29045,7 +29132,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 222992, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "38fd2e9cf9797c59ca9c4001e5c2a384a19814e3da0d68bb8b90c660d0b4750e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 222992, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "7ad5e073c719a84fa5e92d32bd479ad3d62877f315a1a9550425d020e73ac6c8", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -29171,7 +29258,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 229272, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "b22a179046c59f62468931ab30e0719d25c027156c326058e532caafba6617de", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 229272, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "01d006ffd740c2eef1f256dc54b7817abb2250ff04d243fbc217f29abbe22a78", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -29297,7 +29384,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 220952, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "b1fc27759fc7df0d3f32e30c871ab5bc4707b2b835184a4a1dcde9b80618396a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 220952, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "23c52ee74f1f4e42fcf4ad7b440bac51ce9221aa2e392e77408faaf42f692ccc", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -29423,7 +29510,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 222864, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "171c7bcff2c3f1157619e80fd29932b365028525b453accb5d7f77cde17798c9", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 222864, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "304684abd14ff28417f927112b54d831c846b3d489f3dfb515e148fcd86897ab", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -29549,7 +29636,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 222736, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "773fc7e3ece7872af040624075f86ea2a5fefa1a5d0820886cbbca62d82400a0", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 222736, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "04299904677b1cbec591bc0d7109e6e39c9850d8dd6bdbb6e8bfe7789c93926c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -29675,7 +29762,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 229016, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "a1cfcb3d01af4ae7e6cb78712a7bdbef6f3f7e44c5beb58b9922befe3dbaabd2", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 229016, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "4137774845447bf5b0c79d8feb62dd4ec029cb177f82baa55ed8a45d10c63522", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -29801,7 +29888,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 220696, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "d95a4232c6682463b883dba0480c7093cf0d51de983517eb0c6195bff509ed99", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 220696, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "4269f15f251bc1323547199b33ffa4f896919add565629d25ea30a420f507cee", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -29927,7 +30014,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 222864, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "38e9c7c6f32c97afe32a742500ba5d42de0bafda5110335aae5b56afdd72a626", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 222864, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "9bfd32eca43da4b55d97a5e407cf914d708d730dc7f675528fae937c6ed05573", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -30053,7 +30140,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 222736, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "ef880a9e4748b5c2d920f5212abba6a1e2b2552a62a6a6ea66e1933d0dc722bc", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 222736, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "c2a5e736becca3311c38fcc6418f196cf1a60346897ec66bb1e6dd4d8555fc98", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -30179,7 +30266,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 229016, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "d60e6c1591fba4d6f36b85b067a3ff2df4e2f3f9d1b3f5a9ab15c3b79b33c613", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 229016, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "9bd0f2ec4337352414c4bf78789903bb70a1bbe0dc8fbabcb4a6084e2bf473fd", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -30305,7 +30392,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 220696, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "b860d3462e4dd862078eb62bb2f86f76459c65db8cf77ccd07d3408a729fc291", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin, Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f_cubin_len, 220696, "bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f", 512, "283a842f2d57fdd8c573f1fc2800e12d5e3b1117994dbbc1e68f56f098251054", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -30431,7 +30518,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 181144, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f", 512, "0800fc4f4c9fb56d1876a4065af3b7311ef88970a9c132737b68af7d20a5eafe", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 181144, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f", 512, "4ea12bf2dacf1929b7f6762148dddf320b3ebff4e9268b2ae4a4b62cfde21074", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -30557,7 +30644,133 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 181144, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "ea04344109bd50d6c9d53011c692f104aa619b2987275bbf49b1cf17e7e072b3", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 181144, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "602d0a66755aa09e0757ec1e3b0de5df287953a70b470adda6c59221b80b5a8a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(1) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 2 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 1 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 16 +, /* mEpilogueLdtmBits */ 256 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 32 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 0 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 256 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 256 +, /* mMmaN */ 128 +, /* mMmaTileK */ 256 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 48 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 6 +, /* mNumStagesB */ 6 +, /* mNumStagesMma */ 2 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 2 +, /* mNumStagesSmemSfA */ 6 +, /* mNumStagesSmemSfB */ 6 +, /* mNumStagesTmemSfA */ 6 +, /* mNumStagesTmemSfB */ 6 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(3) +, /* mSfLayoutB */ trtllm::gen::SfLayout(0) +, /* mSfLayoutC */ trtllm::gen::SfLayout(3) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 256 +, /* mTileM */ 128 +, /* mTileN */ 128 +, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTransposeMmaOutput */ 1 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 0 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 0 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 256 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(2) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 1 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 0 +, /* mNumRegsPerThreadLoadSfA */ 0 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 0 +, /* mNumWarpsLoadSfA */ 0 +, /* mNumWarpsLoadSfB */ 0 +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 181144, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "1a82bcedf4bdbb72662b7ddb00140cf0ae2961bb1d9bd090250533e0a648e911", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -30683,7 +30896,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 181144, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 512, "c39768a2366c66f142773fe526a4c93992313590a4cc3fc7a8dc78f35957c157", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 181144, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 512, "74c0ef0b6c9ba01897a0f84a2ef7bf96412af2921fc2444006561546bbdae603", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -30809,7 +31022,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 181144, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 512, "d5aba136f449f01ceed6b1e32e73f25801972c52091d59b1c34faa3f9cf71460", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 181144, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 512, "2dc6aed315b1cb1d069fd084c6e16270d77c7fda83571fdc276fcfd5292535ca", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -30935,7 +31148,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_geGlu_lsfaW2_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_geGlu_lsfaW2_dynB_sm100f_cubin_len, 181144, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_geGlu_lsfaW2_dynB_sm100f", 640, "80980e42090347b6027b71115b985f08349e6f99ed7ad1de4b59a253b3fa11b5", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_geGlu_lsfaW2_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_geGlu_lsfaW2_dynB_sm100f_cubin_len, 181144, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_geGlu_lsfaW2_dynB_sm100f", 640, "bf9493edbfc14d40140f7c3ffba48083deb63e9c2986d4cced824b5b048233b2", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(2) , /* mBlockK */ -1 @@ -31061,7 +31274,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_geGlu_lsfaW4_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_geGlu_lsfaW4_dynB_sm100f_cubin_len, 181144, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_geGlu_lsfaW4_dynB_sm100f", 768, "7af3bf3259318bf008181d564333c637138879849f48cffc37b372fe7b508696", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_geGlu_lsfaW4_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_geGlu_lsfaW4_dynB_sm100f_cubin_len, 181144, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_geGlu_lsfaW4_dynB_sm100f", 768, "4d13496b08e71062058a146c24cefee8f8a1e8d56338e98824631ac402a02134", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(2) , /* mBlockK */ -1 @@ -31187,7 +31400,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_lsfaW2_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_lsfaW2_dynB_sm100f_cubin_len, 181144, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_lsfaW2_dynB_sm100f", 640, "88adedea3da89a8cb12b299ab09d83e5b2f6fc13732a6f3040f8e7e057b1571e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_siTuGlu_lsfaW2_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_siTuGlu_lsfaW2_dynB_sm100f_cubin_len, 181144, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_siTuGlu_lsfaW2_dynB_sm100f", 640, "2d6f40166b81db481bc92fc060278580586bf63cfd3c3840c2d40e17c103fffb", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(2) , /* mBlockK */ -1 @@ -31287,7 +31500,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mValidN */ 256 , /* mValidK */ 256 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(0) +, /* mActType */ gemmGatedAct::ActType(2) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} @@ -31313,7 +31526,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_lsfaW4_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_lsfaW4_dynB_sm100f_cubin_len, 181144, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_lsfaW4_dynB_sm100f", 768, "f14cf6b79fdd401509237e8ecf6106bc3f588debe831205d096d5ebb95f30759", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_siTuGlu_lsfaW4_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_siTuGlu_lsfaW4_dynB_sm100f_cubin_len, 181144, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_siTuGlu_lsfaW4_dynB_sm100f", 768, "89188e4ff6f510e61e4099a5c07cff9c1956a37b5e180fdf6a708fd8397aabc6", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(2) , /* mBlockK */ -1 @@ -31413,7 +31626,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mValidN */ 256 , /* mValidK */ 256 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(0) +, /* mActType */ gemmGatedAct::ActType(2) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} @@ -31439,7 +31652,259 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_relu2_bM_tma_ldgstsSf_rgTma_clmp_lsfaW2_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_relu2_bM_tma_ldgstsSf_rgTma_clmp_lsfaW2_dynB_sm100f_cubin_len, 181144, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_relu2_bM_tma_ldgstsSf_rgTma_clmp_lsfaW2_dynB_sm100f", 640, "a12e57b1bc8caff82de0ebe8d70864ac6625ec76b6c26b20d1e71f434cc00f51", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_lsfaW2_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_lsfaW2_dynB_sm100f_cubin_len, 181144, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_lsfaW2_dynB_sm100f", 640, "7c19b0e67f591dadda663cd7ad980aa104edbcde1bd3efe264c058416179c5bf", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(2) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 2 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 1 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(1) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 32 +, /* mEpilogueLdtmBits */ 32 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 64 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 0 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 256 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 256 +, /* mMmaN */ 128 +, /* mMmaTileK */ 256 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 48 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 6 +, /* mNumStagesB */ 6 +, /* mNumStagesMma */ 2 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 2 +, /* mNumStagesSmemSfA */ 6 +, /* mNumStagesSmemSfB */ 6 +, /* mNumStagesTmemSfA */ 6 +, /* mNumStagesTmemSfB */ 6 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(0) +, /* mSfLayoutB */ trtllm::gen::SfLayout(3) +, /* mSfLayoutC */ trtllm::gen::SfLayout(3) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 256 +, /* mTileM */ 128 +, /* mTileN */ 128 +, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTransposeMmaOutput */ 0 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 0 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 0 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 256 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(0) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(0) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 1 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 0 +, /* mNumRegsPerThreadLoadSfA */ 0 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 0 +, /* mNumWarpsLoadSfA */ 2 +, /* mNumWarpsLoadSfB */ 0 +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_lsfaW4_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_lsfaW4_dynB_sm100f_cubin_len, 181144, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_lsfaW4_dynB_sm100f", 768, "cf7f20a22bac2333c9f35e0d7ba32a67d081ada403789a703fca269c32d9db6a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(2) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 2 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 1 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(1) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 32 +, /* mEpilogueLdtmBits */ 32 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 64 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 0 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 256 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 256 +, /* mMmaN */ 128 +, /* mMmaTileK */ 256 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 48 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 6 +, /* mNumStagesB */ 6 +, /* mNumStagesMma */ 2 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 2 +, /* mNumStagesSmemSfA */ 6 +, /* mNumStagesSmemSfB */ 6 +, /* mNumStagesTmemSfA */ 6 +, /* mNumStagesTmemSfB */ 6 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(0) +, /* mSfLayoutB */ trtllm::gen::SfLayout(3) +, /* mSfLayoutC */ trtllm::gen::SfLayout(3) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 256 +, /* mTileM */ 128 +, /* mTileN */ 128 +, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTransposeMmaOutput */ 0 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 0 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 0 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 256 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(0) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(0) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 1 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 0 +, /* mNumRegsPerThreadLoadSfA */ 0 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 0 +, /* mNumWarpsLoadSfA */ 4 +, /* mNumWarpsLoadSfB */ 0 +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_relu2_bM_tma_ldgstsSf_rgTma_clmp_lsfaW2_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_relu2_bM_tma_ldgstsSf_rgTma_clmp_lsfaW2_dynB_sm100f_cubin_len, 181144, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_relu2_bM_tma_ldgstsSf_rgTma_clmp_lsfaW2_dynB_sm100f", 640, "695774bb0a3ced140cb3200cb039272d34767c87208970f100514a8868df5521", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(2) , /* mBlockK */ -1 @@ -31565,7 +32030,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_relu2_bM_tma_ldgstsSf_rgTma_clmp_lsfaW4_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_relu2_bM_tma_ldgstsSf_rgTma_clmp_lsfaW4_dynB_sm100f_cubin_len, 181144, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_relu2_bM_tma_ldgstsSf_rgTma_clmp_lsfaW4_dynB_sm100f", 768, "67b98f12a4385da35c54feb24004852afcca33ed648cf8a0c1e660b03f7b951e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_relu2_bM_tma_ldgstsSf_rgTma_clmp_lsfaW4_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_relu2_bM_tma_ldgstsSf_rgTma_clmp_lsfaW4_dynB_sm100f_cubin_len, 181144, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_relu2_bM_tma_ldgstsSf_rgTma_clmp_lsfaW4_dynB_sm100f", 768, "d21f3f7a0314b669788b52d17b16933d290a43df1117d61d781785b5fde070f7", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(2) , /* mBlockK */ -1 @@ -31691,7 +32156,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_silu_bM_tma_ldgstsSf_rgTma_clmp_lsfaW2_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_silu_bM_tma_ldgstsSf_rgTma_clmp_lsfaW2_dynB_sm100f_cubin_len, 181144, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_silu_bM_tma_ldgstsSf_rgTma_clmp_lsfaW2_dynB_sm100f", 640, "f2c8ed7d7382b08f06b23df4bb4727f235c34f7c385486feede6c6b7d35af3e6", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_silu_bM_tma_ldgstsSf_rgTma_clmp_lsfaW2_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_silu_bM_tma_ldgstsSf_rgTma_clmp_lsfaW2_dynB_sm100f_cubin_len, 181144, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_silu_bM_tma_ldgstsSf_rgTma_clmp_lsfaW2_dynB_sm100f", 640, "5b433e597c731acc1a2856331882f8628faaef16874c310ce2a70514a5c77bb4", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(2) , /* mBlockK */ -1 @@ -31817,7 +32282,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_silu_bM_tma_ldgstsSf_rgTma_clmp_lsfaW4_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_silu_bM_tma_ldgstsSf_rgTma_clmp_lsfaW4_dynB_sm100f_cubin_len, 181144, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_silu_bM_tma_ldgstsSf_rgTma_clmp_lsfaW4_dynB_sm100f", 768, "0f84220dbb7f4a72c0bd464c842f1c4b1ef749c996525c7344713fa4dd27bd5d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_silu_bM_tma_ldgstsSf_rgTma_clmp_lsfaW4_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_silu_bM_tma_ldgstsSf_rgTma_clmp_lsfaW4_dynB_sm100f_cubin_len, 181144, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_silu_bM_tma_ldgstsSf_rgTma_clmp_lsfaW4_dynB_sm100f", 768, "794aff105e27bce09c88a7cd6fc24132443013c0bd16136a5ae185d8ffa569fe", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(2) , /* mBlockK */ -1 @@ -31943,7 +32408,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 181144, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f", 512, "5139e257fd12c8f8b14e961a68909326f016f4f67f6df7ab35c082eda101a43e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 181144, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f", 512, "e9dc0c484cb1963a0ebacd98cd2ba8dfc9f6b4b205d545519f203a3890150b04", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -32069,7 +32534,133 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 181144, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "fc5842520859a98bd1674c5941d8a6025633759156b776f95d5622da0df6847c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 181144, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "5ff55463fc97b7d0e5adcf86c498c547d448d1257f9b312aef29e075e38c68e7", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(1) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 2 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 1 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 16 +, /* mEpilogueLdtmBits */ 256 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 32 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 0 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 512 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 256 +, /* mMmaN */ 128 +, /* mMmaTileK */ 256 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 48 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 6 +, /* mNumStagesB */ 6 +, /* mNumStagesMma */ 2 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 2 +, /* mNumStagesSmemSfA */ 6 +, /* mNumStagesSmemSfB */ 6 +, /* mNumStagesTmemSfA */ 6 +, /* mNumStagesTmemSfB */ 6 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(3) +, /* mSfLayoutB */ trtllm::gen::SfLayout(0) +, /* mSfLayoutC */ trtllm::gen::SfLayout(3) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 256 +, /* mTileM */ 128 +, /* mTileN */ 128 +, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTransposeMmaOutput */ 1 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 0 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 1 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 512 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(2) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 1 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 0 +, /* mNumRegsPerThreadLoadSfA */ 0 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 0 +, /* mNumWarpsLoadSfA */ 0 +, /* mNumWarpsLoadSfB */ 0 +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 181144, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "e56ebcfca16c973c62cbc795e1091cdc1f30737cb3c707cb416bd06c186bc6ff", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -32195,7 +32786,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 181144, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 512, "353c7ac930d2f30d773d27c86e747c038572e31a9fd05f996422cb9e3d03f230", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 181144, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 512, "cfbc9d7fbf655cd4fcc42c75411f88354e53a4d13a3e32971e04f9e5940217a3", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -32321,7 +32912,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 181144, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 512, "1cbbcdab1ba2744c69f303e899beb4e2ec7b284619360170d6b957cfccf6b8c6", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 181144, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 512, "460c3067ae22785d3d957ae7e25a8bfee061f1372a1d5472b409c025accd7972", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -32447,7 +33038,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_geGlu_lbW4_lsfbW4_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_geGlu_lbW4_lsfbW4_dynB_sm100f_cubin_len, 180888, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_geGlu_lbW4_lsfbW4_dynB_sm100f", 384, "74f91d4c838c4367e2b7a20cdb5a35bf6d986d51e4665af7975f1859b57f3209", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_geGlu_lbW4_lsfbW4_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_geGlu_lbW4_lsfbW4_dynB_sm100f_cubin_len, 180888, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_geGlu_lbW4_lsfbW4_dynB_sm100f", 384, "da5a5c4214e0b0a7940137cdb8bb4065448736ff73e7fdb7cc96c69e6f47f950", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -32573,7 +33164,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_lsfbW4_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_lsfbW4_dynB_sm100f_cubin_len, 180888, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_lsfbW4_dynB_sm100f", 384, "47585f807ef67c2231b1ea0e29fe88455afd4b388e8dca10cf9a53110f3a7f7f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_siTuGlu_lbW4_lsfbW4_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_siTuGlu_lbW4_lsfbW4_dynB_sm100f_cubin_len, 180888, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_siTuGlu_lbW4_lsfbW4_dynB_sm100f", 384, "0d8b63db35e77c8eb1fa107429437ff5013d48768b4b4ac8ac6c6b9b6002bbe6", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -32673,7 +33264,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mValidN */ 256 , /* mValidK */ 512 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(0) +, /* mActType */ gemmGatedAct::ActType(2) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} @@ -32699,7 +33290,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_lsfbW4_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_lsfbW4_dynB_sm100f_cubin_len, 180888, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_lsfbW4_dynB_sm100f", 384, "a073296e8a3594c07ca2eae044c8f6f57f7b05964a0177890fe55dd906d190b4", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_lsfbW4_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_lsfbW4_dynB_sm100f_cubin_len, 180888, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_lsfbW4_dynB_sm100f", 384, "8f5bca84a5ba992febd3a86fe15c7a7385c278eeab4e4ae5e3b7d5c9d072bd1b", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -32715,7 +33306,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(2) +, /* mEltwiseActType */ gemm::EltwiseActType(0) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 @@ -32799,13 +33390,13 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mValidN */ 256 , /* mValidK */ 512 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(3) +, /* mActType */ gemmGatedAct::ActType(0) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 0 +, /* mFusedAct */ 1 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 @@ -32825,7 +33416,133 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_lsfbW4_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_lsfbW4_dynB_sm100f_cubin_len, 180888, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_lsfbW4_dynB_sm100f", 384, "7bb1b3e4502bb18a56a33816595dc9fbbaa1945fcc020cdbfbd2f4c5f3c56cbd", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_lsfbW4_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_lsfbW4_dynB_sm100f_cubin_len, 180888, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_lsfbW4_dynB_sm100f", 384, "405cf50d8cfaf29b2e6c70c6f48975851945bd9ba612dcde1f9124e03a680fd0", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(1) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 2 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 1 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(2) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 16 +, /* mEpilogueLdtmBits */ 256 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 32 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 1 +, /* mFuseUtccpWithUtcmma */ 1 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 512 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 256 +, /* mMmaN */ 128 +, /* mMmaTileK */ 512 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 3 +, /* mNumStagesB */ 3 +, /* mNumStagesMma */ 2 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 2 +, /* mNumStagesSmemSfA */ 3 +, /* mNumStagesSmemSfB */ 3 +, /* mNumStagesTmemSfA */ 1 +, /* mNumStagesTmemSfB */ 3 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(3) +, /* mSfLayoutB */ trtllm::gen::SfLayout(0) +, /* mSfLayoutC */ trtllm::gen::SfLayout(3) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 512 +, /* mTileM */ 128 +, /* mTileN */ 128 +, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTransposeMmaOutput */ 1 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 0 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 0 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 512 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(3) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 0 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 128 +, /* mNumRegsPerThreadLoadSfA */ 0 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 4 +, /* mNumWarpsLoadSfA */ 0 +, /* mNumWarpsLoadSfB */ 4 +, /* mRouteImpl */ batchedGemm::RouteImpl(1) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_lsfbW4_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_lsfbW4_dynB_sm100f_cubin_len, 180888, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_lsfbW4_dynB_sm100f", 384, "0b5b4cb7536873a70cb2b5df592a07068da85676f9d9b250c396dab7baf852f7", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -32951,7 +33668,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 193680, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f", 640, "eb8141c7f37171079a8357474ef61385e2d7a7dc6e345ab1bc50781ff852f7e4", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 193680, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f", 640, "2ffd86683fece167635ce64877eef2591763e0ee277e3244d1203404f7ea26a0", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -33077,7 +33794,133 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 193680, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "273ab8585f04e734258fd6b7e1b5f9c202b7ea98660caecd5a8e5d0cca858e94", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 193680, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 640, "e789a15a4a6a9f2f6d7806074139d5e105b2f5928fed0909f27ef6145ee9e7c4", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(1) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 1 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 1 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 16 +, /* mEpilogueLdtmBits */ 256 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 16 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 0 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 256 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 128 +, /* mMmaN */ 16 +, /* mMmaTileK */ 256 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 9 +, /* mNumStagesB */ 9 +, /* mNumStagesMma */ 2 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 2 +, /* mNumStagesSmemSfA */ 9 +, /* mNumStagesSmemSfB */ 9 +, /* mNumStagesTmemSfA */ 9 +, /* mNumStagesTmemSfB */ 9 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(3) +, /* mSfLayoutB */ trtllm::gen::SfLayout(0) +, /* mSfLayoutC */ trtllm::gen::SfLayout(1) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 256 +, /* mTileM */ 128 +, /* mTileN */ 16 +, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTransposeMmaOutput */ 1 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 0 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 0 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 256 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(2) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 1 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 0 +, /* mNumRegsPerThreadLoadSfA */ 0 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 0 +, /* mNumWarpsLoadSfA */ 0 +, /* mNumWarpsLoadSfB */ 0 +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 193680, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "6e236ae94bb666f84e1801ca5aae21fe56509068a0770f6db9f8c740c33baa4a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -33203,7 +34046,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 193680, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 640, "c882d703d1b37ee920ad62a8e21d220528217d40eb16f7e588da72f2c1802cec", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 193680, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 640, "d1334c6163b0fc7d03f7c60bf4f6c69ad9067757de14f945dd8e4201d49ea823", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -33329,7 +34172,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 193680, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 640, "9e68c6379e718205316a78127f10f9f81db1a810b88fb599f12068b8a3b87552", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 193680, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 640, "24c5d3b9d5fe3fc344d9c4ea271c2e1aeb3a0400c661d8081232514e98695a71", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -33455,7 +34298,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 193552, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f", 640, "9eae9d82cb1311ff1531ea613a5b20a57722c929a007c2fe4dcf5c94a42f47ed", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 193552, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f", 640, "fa3dcf6769bee018d6c8f41f61b2026335f534500836e56b5bcfc39d18f9f716", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -33581,7 +34424,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 193552, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "f6fb6301cfea4c746627d1f6cf3d53c1ff9b7bf4efadac74d1e285f80434060f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 193552, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 640, "4434d78c9313d713164070198627e13f445c0280bcd77a1e532fd60469f99675", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -33681,7 +34524,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mValidN */ 256 , /* mValidK */ 256 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(0) +, /* mActType */ gemmGatedAct::ActType(2) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} @@ -33707,7 +34550,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 193552, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 640, "5e5594af545c789303c934eeb9a5a0848659cab8c70a4588801c0bb3f4bdd6fe", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 193552, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "aaf9728465693444e9641d0c091cce8ef428b9603dba108f7c8d0bd72dd89043", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -33723,7 +34566,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(2) +, /* mEltwiseActType */ gemm::EltwiseActType(0) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 @@ -33813,7 +34656,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 0 +, /* mFusedAct */ 1 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 @@ -33833,7 +34676,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 193552, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 640, "b7e70fea1bd4a252d05a5012c635700a49abcac098df118569b40243cce8cd28", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 193552, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 640, "e1abf085f52330b6962e4480819f269a4b172cce6e149f977d66659aabef304a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -33849,7 +34692,133 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(3) +, /* mEltwiseActType */ gemm::EltwiseActType(2) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 16 +, /* mEpilogueLdtmBits */ 256 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 16 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 0 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 256 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 128 +, /* mMmaN */ 16 +, /* mMmaTileK */ 256 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 9 +, /* mNumStagesB */ 9 +, /* mNumStagesMma */ 1 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 1 +, /* mNumStagesSmemSfA */ 9 +, /* mNumStagesSmemSfB */ 9 +, /* mNumStagesTmemSfA */ 9 +, /* mNumStagesTmemSfB */ 9 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(3) +, /* mSfLayoutB */ trtllm::gen::SfLayout(0) +, /* mSfLayoutC */ trtllm::gen::SfLayout(1) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 256 +, /* mTileM */ 128 +, /* mTileN */ 16 +, /* mTileScheduler */ gemm::TileScheduler(0) +, /* mTransposeMmaOutput */ 1 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 0 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 0 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 256 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(0) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 0 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 0 +, /* mNumRegsPerThreadLoadSfA */ 0 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 0 +, /* mNumWarpsLoadSfA */ 0 +, /* mNumWarpsLoadSfB */ 0 +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 193552, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 640, "2b66c0d4afee610cba58ec909026b7893601d85aaa72340c8ddab0267d032917", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(1) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 1 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 1 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(3) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 @@ -33959,7 +34928,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 193680, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f", 640, "435269db3fa1e4dea3377d39b04bd32c91ff7bc1a9942a85a055abe816b0b45c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 193680, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f", 640, "15d39883998ab40948481dbc751e8cad8df90bd57db51042bfe32ac53b442d18", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -34085,7 +35054,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 193680, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "5e1a564b47d3260bb3463826966c834a4660d9a09450da8c118b39a7be333e80", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 193680, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 640, "92f72142bba40a7479384f43224214a7b0e03dd119f19fa3c337c90f5413194d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -34185,7 +35154,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mValidN */ 256 , /* mValidK */ 512 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(0) +, /* mActType */ gemmGatedAct::ActType(2) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} @@ -34211,7 +35180,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 193680, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 640, "38fabc266c4bbb35fb91e8af53a323a523ca74748f76cafff69dd6d01a6d9fa8", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 193680, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "e54b8bb3879d443e86b093b2653c791815da8172d80ed5654887a633b2bce430", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -34227,7 +35196,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(2) +, /* mEltwiseActType */ gemm::EltwiseActType(0) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 @@ -34317,7 +35286,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 0 +, /* mFusedAct */ 1 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 @@ -34337,7 +35306,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 193680, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 640, "64eb52449bdc7abf6251e912f2a82aeb67d67186fa46ba283dec0bd25d8382a9", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 193680, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 640, "bd6109e876957baf11a96c05a1434a5289ccc0051b954d72254a7c483046f36e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -34353,7 +35322,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(3) +, /* mEltwiseActType */ gemm::EltwiseActType(2) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 @@ -34437,7 +35406,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mValidN */ 256 , /* mValidK */ 512 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(3) +, /* mActType */ gemmGatedAct::ActType(0) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} @@ -34463,7 +35432,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 193552, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f", 640, "db3561d9cc85b49282b0ee4eac0a5671c178e1fc1140f9d1a7f94046f6307fc6", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 193680, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 640, "f0f7114152123f6328afdc29d05d446b6415a685eeeca14a3eb66d2cc21e8a60", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -34479,7 +35448,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEltwiseActType */ gemm::EltwiseActType(3) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 @@ -34520,9 +35489,9 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumSlicesForSplitK */ 1 , /* mNumStagesA */ 9 , /* mNumStagesB */ 9 -, /* mNumStagesMma */ 1 +, /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 -, /* mNumStagesMmaAcrossWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 2 , /* mNumStagesSmemSfA */ 9 , /* mNumStagesSmemSfB */ 9 , /* mNumStagesTmemSfA */ 9 @@ -34544,7 +35513,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mTileK */ 256 , /* mTileM */ 128 , /* mTileN */ 16 -, /* mTileScheduler */ gemm::TileScheduler(0) +, /* mTileScheduler */ gemm::TileScheduler(1) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 , /* mUseCustomMmaSchedule */ 1 @@ -34563,13 +35532,13 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mValidN */ 256 , /* mValidK */ 512 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(1) +, /* mActType */ gemmGatedAct::ActType(3) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 1 +, /* mFusedAct */ 0 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 @@ -34589,7 +35558,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 193552, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "18e2a53492e8246c8d957d53c95d0fc6c421031f9956fcd48452e549e0e934a2", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 193552, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f", 640, "cc54caf0207e397b96fff0eaed5c66add0b0e948faf1d99172d5fdb559ddc517", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -34689,7 +35658,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mValidN */ 256 , /* mValidK */ 512 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(0) +, /* mActType */ gemmGatedAct::ActType(1) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} @@ -34715,7 +35684,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 193552, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 640, "0a7a832b7f94d26c2454e0aed22100ec2bd1f959295643fc328ffcce445280e2", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 193552, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 640, "d9297aca2ae08bfbf9d0dee3495ec5d9b8e6b71232d8a966063f1105e1c53c83", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -34731,7 +35700,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(2) +, /* mEltwiseActType */ gemm::EltwiseActType(0) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 @@ -34815,13 +35784,13 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mValidN */ 256 , /* mValidK */ 512 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(0) +, /* mActType */ gemmGatedAct::ActType(2) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 0 +, /* mFusedAct */ 1 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 @@ -34841,7 +35810,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 193552, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 640, "021a5f047a1ce19631906b61b68a1ab597148f681da461a28072c470c2efa39e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 193552, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "e5bf70ab61ebea147f7f5675526a909d991cd973291871907a7ced97f7ef409d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -34857,7 +35826,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(3) +, /* mEltwiseActType */ gemm::EltwiseActType(0) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 @@ -34941,13 +35910,13 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mValidN */ 256 , /* mValidK */ 512 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(3) +, /* mActType */ gemmGatedAct::ActType(0) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 0 +, /* mFusedAct */ 1 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 @@ -34967,14 +35936,14 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 195224, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f", 768, "60a5e6e439a2467f048ef4fb7abcf80779bdeeae963bab57c5570adc31f7aff9", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 193552, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 640, "8e02c78024f3027d49d11e6378e8092da608d2d6b39c01305b5d10860fb3e056", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 , /* mClusterDimX */ 1 , /* mClusterDimY */ 1 -, /* mClusterDimZ */ 3 +, /* mClusterDimZ */ 1 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) , /* mDtypeAcc */ trtllm::gen::Dtype(1056776) , /* mDtypeA */ trtllm::gen::Dtype(17826818) @@ -34983,7 +35952,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEltwiseActType */ gemm::EltwiseActType(2) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 @@ -35003,7 +35972,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 1536 +, /* mK */ 512 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) @@ -35012,7 +35981,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mMmaKind */ trtllm::gen::MmaKind(4) , /* mMmaM */ 128 , /* mMmaN */ 16 -, /* mMmaTileK */ 512 +, /* mMmaTileK */ 256 , /* mMockAllReduce */ 0 , /* mN */ 256 , /* mNumEpilogueWarps */ 4 @@ -35021,16 +35990,16 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 -, /* mNumSlicesForSplitK */ 3 -, /* mNumStagesA */ 4 -, /* mNumStagesB */ 4 -, /* mNumStagesMma */ 2 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 9 +, /* mNumStagesB */ 9 +, /* mNumStagesMma */ 1 , /* mNumStagesMmaWithinWorkTile */ 1 -, /* mNumStagesMmaAcrossWorkTile */ 2 -, /* mNumStagesSmemSfA */ 4 -, /* mNumStagesSmemSfB */ 4 -, /* mNumStagesTmemSfA */ 4 -, /* mNumStagesTmemSfB */ 4 +, /* mNumStagesMmaAcrossWorkTile */ 1 +, /* mNumStagesSmemSfA */ 9 +, /* mNumStagesSmemSfB */ 9 +, /* mNumStagesTmemSfA */ 9 +, /* mNumStagesTmemSfB */ 9 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -35044,11 +36013,11 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfLayoutC */ trtllm::gen::SfLayout(1) , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) -, /* mSplitK */ gemm::SplitK(2) -, /* mTileK */ 512 +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 256 , /* mTileM */ 128 , /* mTileN */ 16 -, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTileScheduler */ gemm::TileScheduler(0) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 , /* mUseCustomMmaSchedule */ 1 @@ -35062,18 +36031,18 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseTmaStore */ 1 , /* mUseTwoTmaLoadWarps */ 1 , /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 0 +, /* mUseUnrollLoop2xForMma */ 1 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 1536 +, /* mValidK */ 512 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(1) +, /* mActType */ gemmGatedAct::ActType(0) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 1 +, /* mFusedAct */ 0 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 @@ -35088,19 +36057,19 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 , /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 195224, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 768, "e42455edae8bcd32c39b50b233423200a750c923f75e19462ef0c2c38eda96d8", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 193552, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 640, "aa28552efeb666a8803034c835391fd0d860a42d97232745fb49853b37e855ee", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 , /* mClusterDimX */ 1 , /* mClusterDimY */ 1 -, /* mClusterDimZ */ 3 +, /* mClusterDimZ */ 1 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) , /* mDtypeAcc */ trtllm::gen::Dtype(1056776) , /* mDtypeA */ trtllm::gen::Dtype(17826818) @@ -35109,7 +36078,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(2) +, /* mEltwiseActType */ gemm::EltwiseActType(3) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 @@ -35129,7 +36098,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 1536 +, /* mK */ 512 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) @@ -35138,7 +36107,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mMmaKind */ trtllm::gen::MmaKind(4) , /* mMmaM */ 128 , /* mMmaN */ 16 -, /* mMmaTileK */ 512 +, /* mMmaTileK */ 256 , /* mMockAllReduce */ 0 , /* mN */ 256 , /* mNumEpilogueWarps */ 4 @@ -35147,16 +36116,16 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 -, /* mNumSlicesForSplitK */ 3 -, /* mNumStagesA */ 4 -, /* mNumStagesB */ 4 -, /* mNumStagesMma */ 2 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 9 +, /* mNumStagesB */ 9 +, /* mNumStagesMma */ 1 , /* mNumStagesMmaWithinWorkTile */ 1 -, /* mNumStagesMmaAcrossWorkTile */ 2 -, /* mNumStagesSmemSfA */ 4 -, /* mNumStagesSmemSfB */ 4 -, /* mNumStagesTmemSfA */ 4 -, /* mNumStagesTmemSfB */ 4 +, /* mNumStagesMmaAcrossWorkTile */ 1 +, /* mNumStagesSmemSfA */ 9 +, /* mNumStagesSmemSfB */ 9 +, /* mNumStagesTmemSfA */ 9 +, /* mNumStagesTmemSfB */ 9 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -35170,11 +36139,11 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfLayoutC */ trtllm::gen::SfLayout(1) , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) -, /* mSplitK */ gemm::SplitK(2) -, /* mTileK */ 512 +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 256 , /* mTileM */ 128 , /* mTileN */ 16 -, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTileScheduler */ gemm::TileScheduler(0) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 , /* mUseCustomMmaSchedule */ 1 @@ -35188,10 +36157,10 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseTmaStore */ 1 , /* mUseTwoTmaLoadWarps */ 1 , /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 0 +, /* mUseUnrollLoop2xForMma */ 1 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 1536 +, /* mValidK */ 512 , /* mWorldSize */ 1 , /* mActType */ gemmGatedAct::ActType(3) , /* mClampBeforeAct */ 1 @@ -35214,12 +36183,12 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 , /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 195224, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 768, "95d1039fe1e4862bf55c7069e3ac601b293e3e9e0048c533b530d3e877a6b7c9", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 195224, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f", 768, "45ef0f13c327342b7ef07ceccca6346e9e33b5c51bb5babf182cfead8685b6f1", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -35235,7 +36204,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(3) +, /* mEltwiseActType */ gemm::EltwiseActType(0) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 @@ -35319,13 +36288,13 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mValidN */ 256 , /* mValidK */ 1536 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(3) +, /* mActType */ gemmGatedAct::ActType(1) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 0 +, /* mFusedAct */ 1 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 @@ -35345,14 +36314,14 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 203416, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f", 768, "0daf2823ede553a9c49a56afe46d31d537e424f54fa505d225a90cd92b19b097", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 195224, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f", 768, "d7348b8d88baded7cedd86c69e1dba6cddb46212c3a27947eac5c878f0ca5ccf", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 , /* mClusterDimX */ 1 , /* mClusterDimY */ 1 -, /* mClusterDimZ */ 4 +, /* mClusterDimZ */ 3 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) , /* mDtypeAcc */ trtllm::gen::Dtype(1056776) , /* mDtypeA */ trtllm::gen::Dtype(17826818) @@ -35381,7 +36350,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 2048 +, /* mK */ 1536 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) @@ -35399,7 +36368,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 -, /* mNumSlicesForSplitK */ 4 +, /* mNumSlicesForSplitK */ 3 , /* mNumStagesA */ 4 , /* mNumStagesB */ 4 , /* mNumStagesMma */ 2 @@ -35443,9 +36412,9 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 2048 +, /* mValidK */ 1536 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(1) +, /* mActType */ gemmGatedAct::ActType(2) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} @@ -35471,14 +36440,14 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 203416, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 768, "6d91b78fc01924e6f0c7ba11bc0d9ef0cc80f013757f29984871052e7a34dde7", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 195224, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 768, "2de694f14d77bfa8bdc23c226a0fed2a658fabbf2a6d0d64f5902ea890b21888", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 , /* mClusterDimX */ 1 , /* mClusterDimY */ 1 -, /* mClusterDimZ */ 4 +, /* mClusterDimZ */ 3 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) , /* mDtypeAcc */ trtllm::gen::Dtype(1056776) , /* mDtypeA */ trtllm::gen::Dtype(17826818) @@ -35507,7 +36476,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 2048 +, /* mK */ 1536 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) @@ -35525,7 +36494,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 -, /* mNumSlicesForSplitK */ 4 +, /* mNumSlicesForSplitK */ 3 , /* mNumStagesA */ 4 , /* mNumStagesB */ 4 , /* mNumStagesMma */ 2 @@ -35569,7 +36538,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 2048 +, /* mValidK */ 1536 , /* mWorldSize */ 1 , /* mActType */ gemmGatedAct::ActType(3) , /* mClampBeforeAct */ 1 @@ -35597,14 +36566,14 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 203416, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 768, "7c945cb964673ab6fb447c9f19a503d196cd09d24ffb8aa4900a68224de346af", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 195224, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 768, "7c45a10a661bf84a1d3d12d607afe31c82a71d314af64d4e8a46be19742ea389", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 , /* mClusterDimX */ 1 , /* mClusterDimY */ 1 -, /* mClusterDimZ */ 4 +, /* mClusterDimZ */ 3 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) , /* mDtypeAcc */ trtllm::gen::Dtype(1056776) , /* mDtypeA */ trtllm::gen::Dtype(17826818) @@ -35633,7 +36602,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 2048 +, /* mK */ 1536 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) @@ -35651,7 +36620,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 -, /* mNumSlicesForSplitK */ 4 +, /* mNumSlicesForSplitK */ 3 , /* mNumStagesA */ 4 , /* mNumStagesB */ 4 , /* mNumStagesMma */ 2 @@ -35695,7 +36664,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 2048 +, /* mValidK */ 1536 , /* mWorldSize */ 1 , /* mActType */ gemmGatedAct::ActType(3) , /* mClampBeforeAct */ 1 @@ -35723,14 +36692,14 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 203672, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f", 640, "0023fb5528293e1372784e4f4f5eb17f3e5b876f56b6508f3b2316d85b4f6b4f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 203416, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f", 768, "a719b7313f2928c1cc04ec6f397350dcb6cb1388e172af712ed8141f14189f1c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 -, /* mClusterDimX */ 2 +, /* mClusterDimX */ 1 , /* mClusterDimY */ 1 -, /* mClusterDimZ */ 1 +, /* mClusterDimZ */ 4 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) , /* mDtypeAcc */ trtllm::gen::Dtype(1056776) , /* mDtypeA */ trtllm::gen::Dtype(17826818) @@ -35759,14 +36728,14 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 512 +, /* mK */ 2048 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) , /* mM */ 256 , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) -, /* mMmaM */ 256 +, /* mMmaM */ 128 , /* mMmaN */ 16 , /* mMmaTileK */ 512 , /* mMockAllReduce */ 0 @@ -35777,16 +36746,16 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 -, /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 5 -, /* mNumStagesB */ 5 +, /* mNumSlicesForSplitK */ 4 +, /* mNumStagesA */ 4 +, /* mNumStagesB */ 4 , /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 2 -, /* mNumStagesSmemSfA */ 5 -, /* mNumStagesSmemSfB */ 5 -, /* mNumStagesTmemSfA */ 5 -, /* mNumStagesTmemSfB */ 5 +, /* mNumStagesSmemSfA */ 4 +, /* mNumStagesSmemSfB */ 4 +, /* mNumStagesTmemSfA */ 4 +, /* mNumStagesTmemSfB */ 4 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -35800,7 +36769,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfLayoutC */ trtllm::gen::SfLayout(1) , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) -, /* mSplitK */ gemm::SplitK(0) +, /* mSplitK */ gemm::SplitK(2) , /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 16 @@ -35821,7 +36790,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 512 +, /* mValidK */ 2048 , /* mWorldSize */ 1 , /* mActType */ gemmGatedAct::ActType(1) , /* mClampBeforeAct */ 1 @@ -35849,14 +36818,14 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 203672, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "822bc7094959e6e5f648c445167e377d2b9b9386ee09dcb0ae5b33e9d069f0a7", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 203416, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f", 768, "cca439a1385c42137da2345e7ebe4656311ba421b3d3209b6f9832c3747098dc", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 -, /* mClusterDimX */ 2 +, /* mClusterDimX */ 1 , /* mClusterDimY */ 1 -, /* mClusterDimZ */ 1 +, /* mClusterDimZ */ 4 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) , /* mDtypeAcc */ trtllm::gen::Dtype(1056776) , /* mDtypeA */ trtllm::gen::Dtype(17826818) @@ -35885,14 +36854,14 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 512 +, /* mK */ 2048 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) , /* mM */ 256 , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) -, /* mMmaM */ 256 +, /* mMmaM */ 128 , /* mMmaN */ 16 , /* mMmaTileK */ 512 , /* mMockAllReduce */ 0 @@ -35903,16 +36872,16 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 -, /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 5 -, /* mNumStagesB */ 5 +, /* mNumSlicesForSplitK */ 4 +, /* mNumStagesA */ 4 +, /* mNumStagesB */ 4 , /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 2 -, /* mNumStagesSmemSfA */ 5 -, /* mNumStagesSmemSfB */ 5 -, /* mNumStagesTmemSfA */ 5 -, /* mNumStagesTmemSfB */ 5 +, /* mNumStagesSmemSfA */ 4 +, /* mNumStagesSmemSfB */ 4 +, /* mNumStagesTmemSfA */ 4 +, /* mNumStagesTmemSfB */ 4 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -35926,7 +36895,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfLayoutC */ trtllm::gen::SfLayout(1) , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) -, /* mSplitK */ gemm::SplitK(0) +, /* mSplitK */ gemm::SplitK(2) , /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 16 @@ -35947,9 +36916,9 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 512 +, /* mValidK */ 2048 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(0) +, /* mActType */ gemmGatedAct::ActType(2) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} @@ -35975,14 +36944,14 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 203672, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 640, "5892ded9cf2e5d0795195e5e36a01537472c12b24d89b02d85f27dca51ec0d33", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 203416, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 768, "9901164a314df1b2fa111e9063620fa489639c99db06b2ca16904f863248d9ed", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 -, /* mClusterDimX */ 2 +, /* mClusterDimX */ 1 , /* mClusterDimY */ 1 -, /* mClusterDimZ */ 1 +, /* mClusterDimZ */ 4 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) , /* mDtypeAcc */ trtllm::gen::Dtype(1056776) , /* mDtypeA */ trtllm::gen::Dtype(17826818) @@ -36011,14 +36980,14 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 512 +, /* mK */ 2048 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) , /* mM */ 256 , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) -, /* mMmaM */ 256 +, /* mMmaM */ 128 , /* mMmaN */ 16 , /* mMmaTileK */ 512 , /* mMockAllReduce */ 0 @@ -36029,16 +36998,16 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 -, /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 5 -, /* mNumStagesB */ 5 +, /* mNumSlicesForSplitK */ 4 +, /* mNumStagesA */ 4 +, /* mNumStagesB */ 4 , /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 2 -, /* mNumStagesSmemSfA */ 5 -, /* mNumStagesSmemSfB */ 5 -, /* mNumStagesTmemSfA */ 5 -, /* mNumStagesTmemSfB */ 5 +, /* mNumStagesSmemSfA */ 4 +, /* mNumStagesSmemSfB */ 4 +, /* mNumStagesTmemSfA */ 4 +, /* mNumStagesTmemSfB */ 4 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -36052,7 +37021,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfLayoutC */ trtllm::gen::SfLayout(1) , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) -, /* mSplitK */ gemm::SplitK(0) +, /* mSplitK */ gemm::SplitK(2) , /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 16 @@ -36073,9 +37042,9 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 512 +, /* mValidK */ 2048 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(0) +, /* mActType */ gemmGatedAct::ActType(3) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} @@ -36101,14 +37070,14 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 203672, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 640, "23776927d9444a8f84f8b2d706171fb1792b1565d929661066b143a7a1b38131", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 203416, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 768, "224106fb946a2e615f86b5a5e9c6f01b0b7268caa52891817458c07c69107a7d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 -, /* mClusterDimX */ 2 +, /* mClusterDimX */ 1 , /* mClusterDimY */ 1 -, /* mClusterDimZ */ 1 +, /* mClusterDimZ */ 4 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) , /* mDtypeAcc */ trtllm::gen::Dtype(1056776) , /* mDtypeA */ trtllm::gen::Dtype(17826818) @@ -36137,14 +37106,14 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 512 +, /* mK */ 2048 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) , /* mM */ 256 , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) -, /* mMmaM */ 256 +, /* mMmaM */ 128 , /* mMmaN */ 16 , /* mMmaTileK */ 512 , /* mMockAllReduce */ 0 @@ -36155,16 +37124,16 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 -, /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 5 -, /* mNumStagesB */ 5 +, /* mNumSlicesForSplitK */ 4 +, /* mNumStagesA */ 4 +, /* mNumStagesB */ 4 , /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 2 -, /* mNumStagesSmemSfA */ 5 -, /* mNumStagesSmemSfB */ 5 -, /* mNumStagesTmemSfA */ 5 -, /* mNumStagesTmemSfB */ 5 +, /* mNumStagesSmemSfA */ 4 +, /* mNumStagesSmemSfB */ 4 +, /* mNumStagesTmemSfA */ 4 +, /* mNumStagesTmemSfB */ 4 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -36178,7 +37147,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfLayoutC */ trtllm::gen::SfLayout(1) , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) -, /* mSplitK */ gemm::SplitK(0) +, /* mSplitK */ gemm::SplitK(2) , /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 16 @@ -36199,7 +37168,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 512 +, /* mValidK */ 2048 , /* mWorldSize */ 1 , /* mActType */ gemmGatedAct::ActType(3) , /* mClampBeforeAct */ 1 @@ -36227,7 +37196,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 203288, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f", 640, "82eb7aa2c5e511d6aa8419b6082dcf65f5ed647509bb7e07a5a4ab1c8c6d0a77", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 203672, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f", 640, "2cc5fb5106bc9117cf10e65171651411796e779db5e7f3077282883f70483408", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -36284,9 +37253,9 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumSlicesForSplitK */ 1 , /* mNumStagesA */ 5 , /* mNumStagesB */ 5 -, /* mNumStagesMma */ 1 +, /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 -, /* mNumStagesMmaAcrossWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 2 , /* mNumStagesSmemSfA */ 5 , /* mNumStagesSmemSfB */ 5 , /* mNumStagesTmemSfA */ 5 @@ -36308,7 +37277,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 16 -, /* mTileScheduler */ gemm::TileScheduler(0) +, /* mTileScheduler */ gemm::TileScheduler(1) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 , /* mUseCustomMmaSchedule */ 1 @@ -36353,7 +37322,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 203288, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "f7731647196df15ade7ba953bccd68db614269b2d2d59f4d9efd5f0591fd567d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 203672, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f", 640, "f723d1398d4b6f703beec87cc06c7b18613021b12b06692de5411d0d91d2c9c5", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -36410,9 +37379,9 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumSlicesForSplitK */ 1 , /* mNumStagesA */ 5 , /* mNumStagesB */ 5 -, /* mNumStagesMma */ 1 +, /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 -, /* mNumStagesMmaAcrossWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 2 , /* mNumStagesSmemSfA */ 5 , /* mNumStagesSmemSfB */ 5 , /* mNumStagesTmemSfA */ 5 @@ -36434,7 +37403,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 16 -, /* mTileScheduler */ gemm::TileScheduler(0) +, /* mTileScheduler */ gemm::TileScheduler(1) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 , /* mUseCustomMmaSchedule */ 1 @@ -36453,7 +37422,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mValidN */ 256 , /* mValidK */ 512 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(0) +, /* mActType */ gemmGatedAct::ActType(2) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} @@ -36479,7 +37448,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 203288, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 640, "78a0e608204b168fe65dbffd09b19705cfabf8724458410d2b6d64e58afd7813", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 203672, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "b2f9dbbf7b4bee3381a0b0de739085b3ef528765c48e95778b80c754675e6370", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -36495,7 +37464,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(2) +, /* mEltwiseActType */ gemm::EltwiseActType(0) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 @@ -36536,9 +37505,9 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumSlicesForSplitK */ 1 , /* mNumStagesA */ 5 , /* mNumStagesB */ 5 -, /* mNumStagesMma */ 1 +, /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 -, /* mNumStagesMmaAcrossWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 2 , /* mNumStagesSmemSfA */ 5 , /* mNumStagesSmemSfB */ 5 , /* mNumStagesTmemSfA */ 5 @@ -36560,7 +37529,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 16 -, /* mTileScheduler */ gemm::TileScheduler(0) +, /* mTileScheduler */ gemm::TileScheduler(1) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 , /* mUseCustomMmaSchedule */ 1 @@ -36585,7 +37554,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 0 +, /* mFusedAct */ 1 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 @@ -36605,7 +37574,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 203288, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 640, "bf39a381651feea10ab26c855601d6b6ecb919d05a1f1a627f722a49437d3d2c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 203672, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 640, "6761b9b7b118187aa475636ce1a44c39fea6235eed6843b7eb7e5fbb116c4537", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -36621,7 +37590,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(3) +, /* mEltwiseActType */ gemm::EltwiseActType(2) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 @@ -36662,9 +37631,9 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumSlicesForSplitK */ 1 , /* mNumStagesA */ 5 , /* mNumStagesB */ 5 -, /* mNumStagesMma */ 1 +, /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 -, /* mNumStagesMmaAcrossWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 2 , /* mNumStagesSmemSfA */ 5 , /* mNumStagesSmemSfB */ 5 , /* mNumStagesTmemSfA */ 5 @@ -36686,7 +37655,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 16 -, /* mTileScheduler */ gemm::TileScheduler(0) +, /* mTileScheduler */ gemm::TileScheduler(1) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 , /* mUseCustomMmaSchedule */ 1 @@ -36705,7 +37674,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mValidN */ 256 , /* mValidK */ 512 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(3) +, /* mActType */ gemmGatedAct::ActType(0) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} @@ -36731,7 +37700,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 203672, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f", 640, "43d0a3004a82ae69cefb205731030068b632eb29dd882d3792fd0ff207c887a2", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 203672, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 640, "721cc038720c9ba44d884af77ef694ca9a2d9b323fdeca861ff595e79c03005b", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -36747,7 +37716,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEltwiseActType */ gemm::EltwiseActType(3) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 @@ -36767,7 +37736,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 1024 +, /* mK */ 512 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) @@ -36826,18 +37795,18 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseTmaStore */ 1 , /* mUseTwoTmaLoadWarps */ 1 , /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 1 +, /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 1024 +, /* mValidK */ 512 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(1) +, /* mActType */ gemmGatedAct::ActType(3) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 1 +, /* mFusedAct */ 0 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 @@ -36857,7 +37826,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 203672, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "ea45593a1b79998942c457f2882e3efb784e9b312b0e3d65401dc680a8709f0f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 203288, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f", 640, "46c1945d3823a348c377263d4fba049a267e26b3cd2d0d489dbf1db0263288de", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -36893,7 +37862,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 1024 +, /* mK */ 512 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) @@ -36914,9 +37883,9 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumSlicesForSplitK */ 1 , /* mNumStagesA */ 5 , /* mNumStagesB */ 5 -, /* mNumStagesMma */ 2 +, /* mNumStagesMma */ 1 , /* mNumStagesMmaWithinWorkTile */ 1 -, /* mNumStagesMmaAcrossWorkTile */ 2 +, /* mNumStagesMmaAcrossWorkTile */ 1 , /* mNumStagesSmemSfA */ 5 , /* mNumStagesSmemSfB */ 5 , /* mNumStagesTmemSfA */ 5 @@ -36938,7 +37907,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 16 -, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTileScheduler */ gemm::TileScheduler(0) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 , /* mUseCustomMmaSchedule */ 1 @@ -36952,12 +37921,12 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseTmaStore */ 1 , /* mUseTwoTmaLoadWarps */ 1 , /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 1 +, /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 1024 +, /* mValidK */ 512 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(0) +, /* mActType */ gemmGatedAct::ActType(1) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} @@ -36983,7 +37952,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 203672, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 640, "edc06af8e38e340af8a960b20672e9f2524cc3c8e259bde0d5d964ab4827be3a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 203288, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f", 640, "a037550936fa5363a7bd82590e93f23e26272d5b092b2d674050e1f18cad4196", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -36999,7 +37968,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(2) +, /* mEltwiseActType */ gemm::EltwiseActType(0) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 @@ -37019,7 +37988,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 1024 +, /* mK */ 512 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) @@ -37040,9 +38009,9 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumSlicesForSplitK */ 1 , /* mNumStagesA */ 5 , /* mNumStagesB */ 5 -, /* mNumStagesMma */ 2 +, /* mNumStagesMma */ 1 , /* mNumStagesMmaWithinWorkTile */ 1 -, /* mNumStagesMmaAcrossWorkTile */ 2 +, /* mNumStagesMmaAcrossWorkTile */ 1 , /* mNumStagesSmemSfA */ 5 , /* mNumStagesSmemSfB */ 5 , /* mNumStagesTmemSfA */ 5 @@ -37064,7 +38033,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 16 -, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTileScheduler */ gemm::TileScheduler(0) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 , /* mUseCustomMmaSchedule */ 1 @@ -37078,18 +38047,18 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseTmaStore */ 1 , /* mUseTwoTmaLoadWarps */ 1 , /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 1 +, /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 1024 +, /* mValidK */ 512 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(0) +, /* mActType */ gemmGatedAct::ActType(2) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 0 +, /* mFusedAct */ 1 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 @@ -37109,7 +38078,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 203672, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 640, "839bd79351712d5cf5cb0699beb499a0df9993da454fd870fc0ce17d15381868", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 203288, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "a1ae08e9f940257ba93a9a5137e92a08b200959e26f5c21b6caeae7cbb10b48f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -37125,7 +38094,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(3) +, /* mEltwiseActType */ gemm::EltwiseActType(0) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 @@ -37145,7 +38114,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 1024 +, /* mK */ 512 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) @@ -37166,9 +38135,9 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumSlicesForSplitK */ 1 , /* mNumStagesA */ 5 , /* mNumStagesB */ 5 -, /* mNumStagesMma */ 2 +, /* mNumStagesMma */ 1 , /* mNumStagesMmaWithinWorkTile */ 1 -, /* mNumStagesMmaAcrossWorkTile */ 2 +, /* mNumStagesMmaAcrossWorkTile */ 1 , /* mNumStagesSmemSfA */ 5 , /* mNumStagesSmemSfB */ 5 , /* mNumStagesTmemSfA */ 5 @@ -37190,7 +38159,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 16 -, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTileScheduler */ gemm::TileScheduler(0) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 , /* mUseCustomMmaSchedule */ 1 @@ -37204,18 +38173,18 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseTmaStore */ 1 , /* mUseTwoTmaLoadWarps */ 1 , /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 1 +, /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 1024 +, /* mValidK */ 512 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(3) +, /* mActType */ gemmGatedAct::ActType(0) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 0 +, /* mFusedAct */ 1 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 @@ -37235,7 +38204,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 203288, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f", 640, "c3a003d82263aef19d9588e4989936eb3ac0e7daf7c6944aced00ad1d984433c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 203288, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 640, "e2a11b8e75a187bcaed628ee5a88ab6ff337155d1a5e21a93a3211ca90129af6", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -37251,7 +38220,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEltwiseActType */ gemm::EltwiseActType(2) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 @@ -37271,7 +38240,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 1024 +, /* mK */ 512 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) @@ -37330,18 +38299,18 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseTmaStore */ 1 , /* mUseTwoTmaLoadWarps */ 1 , /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 1 +, /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 1024 +, /* mValidK */ 512 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(1) +, /* mActType */ gemmGatedAct::ActType(0) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 1 +, /* mFusedAct */ 0 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 @@ -37361,7 +38330,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 203288, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "d35e2ab5fee7114cf71e0e1055a1122c1b672d5e49518993fd6fcf7d82b7b821", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 203288, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 640, "027b81226dbace4a516e52c68a8c6818129c6f10d69ff8f14e7acfaa04dbc0e9", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -37377,7 +38346,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEltwiseActType */ gemm::EltwiseActType(3) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 @@ -37397,7 +38366,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 1024 +, /* mK */ 512 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) @@ -37456,18 +38425,18 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseTmaStore */ 1 , /* mUseTwoTmaLoadWarps */ 1 , /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 1 +, /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 1024 +, /* mValidK */ 512 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(0) +, /* mActType */ gemmGatedAct::ActType(3) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 1 +, /* mFusedAct */ 0 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 @@ -37487,7 +38456,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 203288, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 640, "b3576dc1a038c18ae9957bd385f917e9c64b3605ef888f5b9c8a92509c1ac6b8", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 203672, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f", 640, "8a82591e51e6ab502039c97104269ef18f4376c2ffce067eeeec9e474444358a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -37503,7 +38472,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(2) +, /* mEltwiseActType */ gemm::EltwiseActType(0) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 @@ -37544,9 +38513,9 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumSlicesForSplitK */ 1 , /* mNumStagesA */ 5 , /* mNumStagesB */ 5 -, /* mNumStagesMma */ 1 +, /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 -, /* mNumStagesMmaAcrossWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 2 , /* mNumStagesSmemSfA */ 5 , /* mNumStagesSmemSfB */ 5 , /* mNumStagesTmemSfA */ 5 @@ -37568,7 +38537,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 16 -, /* mTileScheduler */ gemm::TileScheduler(0) +, /* mTileScheduler */ gemm::TileScheduler(1) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 , /* mUseCustomMmaSchedule */ 1 @@ -37587,13 +38556,13 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mValidN */ 256 , /* mValidK */ 1024 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(0) +, /* mActType */ gemmGatedAct::ActType(1) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 0 +, /* mFusedAct */ 1 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 @@ -37613,7 +38582,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 203288, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 640, "d834b9b5635d9b77c2273a7a0c0c032ee37d0eceadba6b7b9a70264ce94bbdc9", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 203672, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f", 640, "f368746db248723ffb81a6acec0440cdee0b3a7aaa8aa5472766830047d1eb39", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -37629,7 +38598,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(3) +, /* mEltwiseActType */ gemm::EltwiseActType(0) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 @@ -37670,9 +38639,9 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumSlicesForSplitK */ 1 , /* mNumStagesA */ 5 , /* mNumStagesB */ 5 -, /* mNumStagesMma */ 1 +, /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 -, /* mNumStagesMmaAcrossWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 2 , /* mNumStagesSmemSfA */ 5 , /* mNumStagesSmemSfB */ 5 , /* mNumStagesTmemSfA */ 5 @@ -37694,7 +38663,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 16 -, /* mTileScheduler */ gemm::TileScheduler(0) +, /* mTileScheduler */ gemm::TileScheduler(1) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 , /* mUseCustomMmaSchedule */ 1 @@ -37713,13 +38682,13 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mValidN */ 256 , /* mValidK */ 1024 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(3) +, /* mActType */ gemmGatedAct::ActType(2) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 0 +, /* mFusedAct */ 1 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 @@ -37739,7 +38708,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_geGlu_lbW8_lsfbW4_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_geGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len, 172696, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_geGlu_lbW8_lsfbW4_dynB_sm100f", 768, "ecb5360a5e4c1f001fbf9d6f03ff30a79c127926c2fddaa757a14eb23333d8cf", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 203672, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "cd65b96304f5fb5c80381b80e16cba0fba0d67876404e23836a996a1bb846174", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -37761,13 +38730,13 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mEpilogueLdtmDps */ 16 , /* mEpilogueLdtmBits */ 256 , /* mEpilogueTileM */ 128 -, /* mEpilogueTileN */ 64 +, /* mEpilogueTileN */ 16 , /* mFallbackClusterDimX */ 1 , /* mFallbackClusterDimY */ 1 , /* mFallbackClusterDimZ */ 1 , /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) , /* mFuseLoadSfTask */ 0 -, /* mFuseUtccpWithUtcmma */ 1 +, /* mFuseUtccpWithUtcmma */ 0 , /* mGridTriggerSecondaryA */ 0 , /* mGridTriggerSecondaryB */ 1 , /* mGridWaitForPrimaryEarlyExit */ 1 @@ -37775,7 +38744,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 256 +, /* mK */ 1024 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) @@ -37783,26 +38752,26 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) , /* mMmaM */ 256 -, /* mMmaN */ 256 -, /* mMmaTileK */ 256 +, /* mMmaN */ 16 +, /* mMmaTileK */ 512 , /* mMockAllReduce */ 0 , /* mN */ 256 -, /* mNumEpilogueWarps */ 8 +, /* mNumEpilogueWarps */ 4 , /* mNumRegsCastAWarps */ 0 , /* mNumRegsCopySfLdsSttm */ 0 , /* mNumRegsCopySparsityInfo */ 0 -, /* mNumRegsPerThreadEpilogueWarp */ 144 -, /* mNumRegsPerThreadNonEpilogueWarp */ 88 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 4 -, /* mNumStagesB */ 4 -, /* mNumStagesMma */ 1 +, /* mNumStagesA */ 5 +, /* mNumStagesB */ 5 +, /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 -, /* mNumStagesMmaAcrossWorkTile */ 1 -, /* mNumStagesSmemSfA */ 4 -, /* mNumStagesSmemSfB */ 4 -, /* mNumStagesTmemSfA */ 1 -, /* mNumStagesTmemSfB */ 4 +, /* mNumStagesMmaAcrossWorkTile */ 2 +, /* mNumStagesSmemSfA */ 5 +, /* mNumStagesSmemSfB */ 5 +, /* mNumStagesTmemSfA */ 5 +, /* mNumStagesTmemSfB */ 5 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -37813,13 +38782,13 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfBlockSizeC */ 16 , /* mSfLayoutA */ trtllm::gen::SfLayout(3) , /* mSfLayoutB */ trtllm::gen::SfLayout(0) -, /* mSfLayoutC */ trtllm::gen::SfLayout(3) +, /* mSfLayoutC */ trtllm::gen::SfLayout(1) , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) , /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 256 +, /* mTileK */ 512 , /* mTileM */ 128 -, /* mTileN */ 256 +, /* mTileN */ 16 , /* mTileScheduler */ gemm::TileScheduler(1) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 @@ -37827,19 +38796,19 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseDeepSeekFp8 */ 0 , /* mUseFlexibleClusterDims */ 0 , /* mUseHoistTryWaitForCustomMmaSchedule */ 0 -, /* mUseMaxTmemOverlap */ 1 +, /* mUseMaxTmemOverlap */ 0 , /* mUsePerTokenSfA */ 0 , /* mUsePerTokenSfB */ 0 , /* mUseShuffledMatrix */ 1 , /* mUseTmaStore */ 1 , /* mUseTwoTmaLoadWarps */ 1 , /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 0 +, /* mUseUnrollLoop2xForMma */ 1 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 256 +, /* mValidK */ 1024 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(1) +, /* mActType */ gemmGatedAct::ActType(0) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} @@ -37851,21 +38820,21 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mIsUniformNumTokensPerBatch */ 0 , /* mNumBatches */ 128 , /* mNumRegsPerThreadLoadA */ 0 -, /* mNumRegsPerThreadLoadB */ 32 +, /* mNumRegsPerThreadLoadB */ 0 , /* mNumRegsPerThreadLoadSfA */ 0 -, /* mNumRegsPerThreadLoadSfB */ 40 +, /* mNumRegsPerThreadLoadSfB */ 0 , /* mNumTokens */ 2 , /* mNumWarpsLoadA */ 0 -, /* mNumWarpsLoadB */ 8 +, /* mNumWarpsLoadB */ 0 , /* mNumWarpsLoadSfA */ 0 -, /* mNumWarpsLoadSfB */ 4 +, /* mNumWarpsLoadSfB */ 0 , /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(3)} +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len, 172696, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f", 768, "210a5300774b5d1bdcbc8e1184964a9346fa83049ceea15e07b871b3444e0e67", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 203672, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 640, "2e94157e76c0c9887560acf58688bca3f4c9a25f11455399a3c7c62515a82717", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -37881,19 +38850,19 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEltwiseActType */ gemm::EltwiseActType(2) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 , /* mEpilogueLdtmBits */ 256 , /* mEpilogueTileM */ 128 -, /* mEpilogueTileN */ 64 +, /* mEpilogueTileN */ 16 , /* mFallbackClusterDimX */ 1 , /* mFallbackClusterDimY */ 1 , /* mFallbackClusterDimZ */ 1 , /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) , /* mFuseLoadSfTask */ 0 -, /* mFuseUtccpWithUtcmma */ 1 +, /* mFuseUtccpWithUtcmma */ 0 , /* mGridTriggerSecondaryA */ 0 , /* mGridTriggerSecondaryB */ 1 , /* mGridWaitForPrimaryEarlyExit */ 1 @@ -37901,7 +38870,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 256 +, /* mK */ 1024 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) @@ -37909,26 +38878,26 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) , /* mMmaM */ 256 -, /* mMmaN */ 256 -, /* mMmaTileK */ 256 +, /* mMmaN */ 16 +, /* mMmaTileK */ 512 , /* mMockAllReduce */ 0 , /* mN */ 256 -, /* mNumEpilogueWarps */ 8 +, /* mNumEpilogueWarps */ 4 , /* mNumRegsCastAWarps */ 0 , /* mNumRegsCopySfLdsSttm */ 0 , /* mNumRegsCopySparsityInfo */ 0 -, /* mNumRegsPerThreadEpilogueWarp */ 144 -, /* mNumRegsPerThreadNonEpilogueWarp */ 88 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 4 -, /* mNumStagesB */ 4 -, /* mNumStagesMma */ 1 +, /* mNumStagesA */ 5 +, /* mNumStagesB */ 5 +, /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 -, /* mNumStagesMmaAcrossWorkTile */ 1 -, /* mNumStagesSmemSfA */ 4 -, /* mNumStagesSmemSfB */ 4 -, /* mNumStagesTmemSfA */ 1 -, /* mNumStagesTmemSfB */ 4 +, /* mNumStagesMmaAcrossWorkTile */ 2 +, /* mNumStagesSmemSfA */ 5 +, /* mNumStagesSmemSfB */ 5 +, /* mNumStagesTmemSfA */ 5 +, /* mNumStagesTmemSfB */ 5 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -37939,13 +38908,13 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfBlockSizeC */ 16 , /* mSfLayoutA */ trtllm::gen::SfLayout(3) , /* mSfLayoutB */ trtllm::gen::SfLayout(0) -, /* mSfLayoutC */ trtllm::gen::SfLayout(3) +, /* mSfLayoutC */ trtllm::gen::SfLayout(1) , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) , /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 256 +, /* mTileK */ 512 , /* mTileM */ 128 -, /* mTileN */ 256 +, /* mTileN */ 16 , /* mTileScheduler */ gemm::TileScheduler(1) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 @@ -37953,17 +38922,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseDeepSeekFp8 */ 0 , /* mUseFlexibleClusterDims */ 0 , /* mUseHoistTryWaitForCustomMmaSchedule */ 0 -, /* mUseMaxTmemOverlap */ 1 +, /* mUseMaxTmemOverlap */ 0 , /* mUsePerTokenSfA */ 0 , /* mUsePerTokenSfB */ 0 , /* mUseShuffledMatrix */ 1 , /* mUseTmaStore */ 1 , /* mUseTwoTmaLoadWarps */ 1 , /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 0 +, /* mUseUnrollLoop2xForMma */ 1 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 256 +, /* mValidK */ 1024 , /* mWorldSize */ 1 , /* mActType */ gemmGatedAct::ActType(0) , /* mClampBeforeAct */ 1 @@ -37971,27 +38940,27 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 1 +, /* mFusedAct */ 0 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 , /* mNumBatches */ 128 , /* mNumRegsPerThreadLoadA */ 0 -, /* mNumRegsPerThreadLoadB */ 32 +, /* mNumRegsPerThreadLoadB */ 0 , /* mNumRegsPerThreadLoadSfA */ 0 -, /* mNumRegsPerThreadLoadSfB */ 40 +, /* mNumRegsPerThreadLoadSfB */ 0 , /* mNumTokens */ 2 , /* mNumWarpsLoadA */ 0 -, /* mNumWarpsLoadB */ 8 +, /* mNumWarpsLoadB */ 0 , /* mNumWarpsLoadSfA */ 0 -, /* mNumWarpsLoadSfB */ 4 +, /* mNumWarpsLoadSfB */ 0 , /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(3)} +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_geGlu_lbW4_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_geGlu_lbW4_dynB_sm100f_cubin_len, 168600, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_geGlu_lbW4_dynB_sm100f", 384, "754e25643de5a5bba6cce4516cd4a9744a7460eaadedfd6a3df40b763f4966bf", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 203672, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 640, "44c17d20d16eec211c8aa5225073168da917f019553605bea6a6a772da3edb54", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -38007,19 +38976,19 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEltwiseActType */ gemm::EltwiseActType(3) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 , /* mEpilogueLdtmBits */ 256 , /* mEpilogueTileM */ 128 -, /* mEpilogueTileN */ 64 +, /* mEpilogueTileN */ 16 , /* mFallbackClusterDimX */ 1 , /* mFallbackClusterDimY */ 1 , /* mFallbackClusterDimZ */ 1 , /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) -, /* mFuseLoadSfTask */ 1 -, /* mFuseUtccpWithUtcmma */ 1 +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 0 , /* mGridTriggerSecondaryA */ 0 , /* mGridTriggerSecondaryB */ 1 , /* mGridWaitForPrimaryEarlyExit */ 1 @@ -38027,7 +38996,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 256 +, /* mK */ 1024 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) @@ -38035,26 +39004,26 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) , /* mMmaM */ 256 -, /* mMmaN */ 256 -, /* mMmaTileK */ 256 +, /* mMmaN */ 16 +, /* mMmaTileK */ 512 , /* mMockAllReduce */ 0 , /* mN */ 256 , /* mNumEpilogueWarps */ 4 , /* mNumRegsCastAWarps */ 0 , /* mNumRegsCopySfLdsSttm */ 0 , /* mNumRegsCopySparsityInfo */ 0 -, /* mNumRegsPerThreadEpilogueWarp */ 232 -, /* mNumRegsPerThreadNonEpilogueWarp */ 88 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 4 -, /* mNumStagesB */ 4 -, /* mNumStagesMma */ 1 +, /* mNumStagesA */ 5 +, /* mNumStagesB */ 5 +, /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 -, /* mNumStagesMmaAcrossWorkTile */ 1 -, /* mNumStagesSmemSfA */ 4 -, /* mNumStagesSmemSfB */ 4 -, /* mNumStagesTmemSfA */ 1 -, /* mNumStagesTmemSfB */ 4 +, /* mNumStagesMmaAcrossWorkTile */ 2 +, /* mNumStagesSmemSfA */ 5 +, /* mNumStagesSmemSfB */ 5 +, /* mNumStagesTmemSfA */ 5 +, /* mNumStagesTmemSfB */ 5 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -38065,13 +39034,13 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfBlockSizeC */ 16 , /* mSfLayoutA */ trtllm::gen::SfLayout(3) , /* mSfLayoutB */ trtllm::gen::SfLayout(0) -, /* mSfLayoutC */ trtllm::gen::SfLayout(3) +, /* mSfLayoutC */ trtllm::gen::SfLayout(1) , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) , /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 256 +, /* mTileK */ 512 , /* mTileM */ 128 -, /* mTileN */ 256 +, /* mTileN */ 16 , /* mTileScheduler */ gemm::TileScheduler(1) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 @@ -38079,45 +39048,45 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseDeepSeekFp8 */ 0 , /* mUseFlexibleClusterDims */ 0 , /* mUseHoistTryWaitForCustomMmaSchedule */ 0 -, /* mUseMaxTmemOverlap */ 1 +, /* mUseMaxTmemOverlap */ 0 , /* mUsePerTokenSfA */ 0 , /* mUsePerTokenSfB */ 0 , /* mUseShuffledMatrix */ 1 , /* mUseTmaStore */ 1 , /* mUseTwoTmaLoadWarps */ 1 , /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 0 +, /* mUseUnrollLoop2xForMma */ 1 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 256 +, /* mValidK */ 1024 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(1) +, /* mActType */ gemmGatedAct::ActType(3) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 1 +, /* mFusedAct */ 0 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 , /* mNumBatches */ 128 , /* mNumRegsPerThreadLoadA */ 0 -, /* mNumRegsPerThreadLoadB */ 160 +, /* mNumRegsPerThreadLoadB */ 0 , /* mNumRegsPerThreadLoadSfA */ 0 , /* mNumRegsPerThreadLoadSfB */ 0 , /* mNumTokens */ 2 , /* mNumWarpsLoadA */ 0 -, /* mNumWarpsLoadB */ 4 +, /* mNumWarpsLoadB */ 0 , /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 -, /* mRouteImpl */ batchedGemm::RouteImpl(1) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_dynB_sm100f_cubin_len, 168600, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_dynB_sm100f", 384, "54766d60c6b1f2826639722e0e8bf981ff8df71116db358dd7ef2d767802b03e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 203288, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f", 640, "5748ad6871d89335003913288d097935158a9cf103e77a0ebbb89c41d9f7453b", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -38139,13 +39108,13 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mEpilogueLdtmDps */ 16 , /* mEpilogueLdtmBits */ 256 , /* mEpilogueTileM */ 128 -, /* mEpilogueTileN */ 64 +, /* mEpilogueTileN */ 16 , /* mFallbackClusterDimX */ 1 , /* mFallbackClusterDimY */ 1 , /* mFallbackClusterDimZ */ 1 , /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) -, /* mFuseLoadSfTask */ 1 -, /* mFuseUtccpWithUtcmma */ 1 +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 0 , /* mGridTriggerSecondaryA */ 0 , /* mGridTriggerSecondaryB */ 1 , /* mGridWaitForPrimaryEarlyExit */ 1 @@ -38153,7 +39122,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 256 +, /* mK */ 1024 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) @@ -38161,26 +39130,26 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) , /* mMmaM */ 256 -, /* mMmaN */ 256 -, /* mMmaTileK */ 256 +, /* mMmaN */ 16 +, /* mMmaTileK */ 512 , /* mMockAllReduce */ 0 , /* mN */ 256 , /* mNumEpilogueWarps */ 4 , /* mNumRegsCastAWarps */ 0 , /* mNumRegsCopySfLdsSttm */ 0 , /* mNumRegsCopySparsityInfo */ 0 -, /* mNumRegsPerThreadEpilogueWarp */ 232 -, /* mNumRegsPerThreadNonEpilogueWarp */ 88 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 4 -, /* mNumStagesB */ 4 +, /* mNumStagesA */ 5 +, /* mNumStagesB */ 5 , /* mNumStagesMma */ 1 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 1 -, /* mNumStagesSmemSfA */ 4 -, /* mNumStagesSmemSfB */ 4 -, /* mNumStagesTmemSfA */ 1 -, /* mNumStagesTmemSfB */ 4 +, /* mNumStagesSmemSfA */ 5 +, /* mNumStagesSmemSfB */ 5 +, /* mNumStagesTmemSfA */ 5 +, /* mNumStagesTmemSfB */ 5 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -38191,33 +39160,33 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfBlockSizeC */ 16 , /* mSfLayoutA */ trtllm::gen::SfLayout(3) , /* mSfLayoutB */ trtllm::gen::SfLayout(0) -, /* mSfLayoutC */ trtllm::gen::SfLayout(3) +, /* mSfLayoutC */ trtllm::gen::SfLayout(1) , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) , /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 256 +, /* mTileK */ 512 , /* mTileM */ 128 -, /* mTileN */ 256 -, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTileN */ 16 +, /* mTileScheduler */ gemm::TileScheduler(0) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 , /* mUseCustomMmaSchedule */ 1 , /* mUseDeepSeekFp8 */ 0 , /* mUseFlexibleClusterDims */ 0 , /* mUseHoistTryWaitForCustomMmaSchedule */ 0 -, /* mUseMaxTmemOverlap */ 1 +, /* mUseMaxTmemOverlap */ 0 , /* mUsePerTokenSfA */ 0 , /* mUsePerTokenSfB */ 0 , /* mUseShuffledMatrix */ 1 , /* mUseTmaStore */ 1 , /* mUseTwoTmaLoadWarps */ 1 , /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 0 +, /* mUseUnrollLoop2xForMma */ 1 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 256 +, /* mValidK */ 1024 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(0) +, /* mActType */ gemmGatedAct::ActType(1) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} @@ -38229,21 +39198,21 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mIsUniformNumTokensPerBatch */ 0 , /* mNumBatches */ 128 , /* mNumRegsPerThreadLoadA */ 0 -, /* mNumRegsPerThreadLoadB */ 160 +, /* mNumRegsPerThreadLoadB */ 0 , /* mNumRegsPerThreadLoadSfA */ 0 , /* mNumRegsPerThreadLoadSfB */ 0 , /* mNumTokens */ 2 , /* mNumWarpsLoadA */ 0 -, /* mNumWarpsLoadB */ 4 +, /* mNumWarpsLoadB */ 0 , /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 -, /* mRouteImpl */ batchedGemm::RouteImpl(1) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_relu2_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_lbW8_lsfbW4_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_relu2_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_lbW8_lsfbW4_dynB_sm100f_cubin_len, 172696, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_relu2_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_lbW8_lsfbW4_dynB_sm100f", 768, "6ed22eb842af4c21ccf604ac0bba7feb5aa9dee683e79380c9f6b51d89062da4", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 203288, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f", 640, "f2032055da48516e0d5af186f67d8f0852902fdad1d24e91c7878e02317747b4", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -38259,19 +39228,19 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(2) +, /* mEltwiseActType */ gemm::EltwiseActType(0) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 , /* mEpilogueLdtmBits */ 256 , /* mEpilogueTileM */ 128 -, /* mEpilogueTileN */ 64 +, /* mEpilogueTileN */ 16 , /* mFallbackClusterDimX */ 1 , /* mFallbackClusterDimY */ 1 , /* mFallbackClusterDimZ */ 1 , /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) , /* mFuseLoadSfTask */ 0 -, /* mFuseUtccpWithUtcmma */ 1 +, /* mFuseUtccpWithUtcmma */ 0 , /* mGridTriggerSecondaryA */ 0 , /* mGridTriggerSecondaryB */ 1 , /* mGridWaitForPrimaryEarlyExit */ 1 @@ -38279,7 +39248,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 256 +, /* mK */ 1024 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) @@ -38287,26 +39256,26 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) , /* mMmaM */ 256 -, /* mMmaN */ 256 -, /* mMmaTileK */ 256 +, /* mMmaN */ 16 +, /* mMmaTileK */ 512 , /* mMockAllReduce */ 0 , /* mN */ 256 -, /* mNumEpilogueWarps */ 8 +, /* mNumEpilogueWarps */ 4 , /* mNumRegsCastAWarps */ 0 , /* mNumRegsCopySfLdsSttm */ 0 , /* mNumRegsCopySparsityInfo */ 0 -, /* mNumRegsPerThreadEpilogueWarp */ 144 -, /* mNumRegsPerThreadNonEpilogueWarp */ 88 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 4 -, /* mNumStagesB */ 4 +, /* mNumStagesA */ 5 +, /* mNumStagesB */ 5 , /* mNumStagesMma */ 1 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 1 -, /* mNumStagesSmemSfA */ 4 -, /* mNumStagesSmemSfB */ 4 -, /* mNumStagesTmemSfA */ 1 -, /* mNumStagesTmemSfB */ 4 +, /* mNumStagesSmemSfA */ 5 +, /* mNumStagesSmemSfB */ 5 +, /* mNumStagesTmemSfA */ 5 +, /* mNumStagesTmemSfB */ 5 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -38317,59 +39286,59 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfBlockSizeC */ 16 , /* mSfLayoutA */ trtllm::gen::SfLayout(3) , /* mSfLayoutB */ trtllm::gen::SfLayout(0) -, /* mSfLayoutC */ trtllm::gen::SfLayout(3) +, /* mSfLayoutC */ trtllm::gen::SfLayout(1) , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) , /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 256 +, /* mTileK */ 512 , /* mTileM */ 128 -, /* mTileN */ 256 -, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTileN */ 16 +, /* mTileScheduler */ gemm::TileScheduler(0) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 , /* mUseCustomMmaSchedule */ 1 , /* mUseDeepSeekFp8 */ 0 , /* mUseFlexibleClusterDims */ 0 , /* mUseHoistTryWaitForCustomMmaSchedule */ 0 -, /* mUseMaxTmemOverlap */ 1 +, /* mUseMaxTmemOverlap */ 0 , /* mUsePerTokenSfA */ 0 , /* mUsePerTokenSfB */ 0 , /* mUseShuffledMatrix */ 1 , /* mUseTmaStore */ 1 , /* mUseTwoTmaLoadWarps */ 1 , /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 0 +, /* mUseUnrollLoop2xForMma */ 1 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 256 +, /* mValidK */ 1024 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(3) +, /* mActType */ gemmGatedAct::ActType(2) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 0 +, /* mFusedAct */ 1 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 , /* mNumBatches */ 128 , /* mNumRegsPerThreadLoadA */ 0 -, /* mNumRegsPerThreadLoadB */ 32 +, /* mNumRegsPerThreadLoadB */ 0 , /* mNumRegsPerThreadLoadSfA */ 0 -, /* mNumRegsPerThreadLoadSfB */ 40 +, /* mNumRegsPerThreadLoadSfB */ 0 , /* mNumTokens */ 2 , /* mNumWarpsLoadA */ 0 -, /* mNumWarpsLoadB */ 8 +, /* mNumWarpsLoadB */ 0 , /* mNumWarpsLoadSfA */ 0 -, /* mNumWarpsLoadSfB */ 4 +, /* mNumWarpsLoadSfB */ 0 , /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(3)} +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_relu2_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_relu2_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_dynB_sm100f_cubin_len, 168600, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_relu2_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_dynB_sm100f", 384, "a275d63ffe577a954792672c7061a25560a982806973924ebf623e51f42ce86f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 203288, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "9ab02a15eec9435853f3616db114c5d0908112f8d98a79bb1ed5a080bbf25b49", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -38385,19 +39354,19 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(2) +, /* mEltwiseActType */ gemm::EltwiseActType(0) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 , /* mEpilogueLdtmBits */ 256 , /* mEpilogueTileM */ 128 -, /* mEpilogueTileN */ 64 +, /* mEpilogueTileN */ 16 , /* mFallbackClusterDimX */ 1 , /* mFallbackClusterDimY */ 1 , /* mFallbackClusterDimZ */ 1 , /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) -, /* mFuseLoadSfTask */ 1 -, /* mFuseUtccpWithUtcmma */ 1 +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 0 , /* mGridTriggerSecondaryA */ 0 , /* mGridTriggerSecondaryB */ 1 , /* mGridWaitForPrimaryEarlyExit */ 1 @@ -38405,7 +39374,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 256 +, /* mK */ 1024 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) @@ -38413,26 +39382,26 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) , /* mMmaM */ 256 -, /* mMmaN */ 256 -, /* mMmaTileK */ 256 +, /* mMmaN */ 16 +, /* mMmaTileK */ 512 , /* mMockAllReduce */ 0 , /* mN */ 256 , /* mNumEpilogueWarps */ 4 , /* mNumRegsCastAWarps */ 0 , /* mNumRegsCopySfLdsSttm */ 0 , /* mNumRegsCopySparsityInfo */ 0 -, /* mNumRegsPerThreadEpilogueWarp */ 232 -, /* mNumRegsPerThreadNonEpilogueWarp */ 88 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 4 -, /* mNumStagesB */ 4 +, /* mNumStagesA */ 5 +, /* mNumStagesB */ 5 , /* mNumStagesMma */ 1 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 1 -, /* mNumStagesSmemSfA */ 4 -, /* mNumStagesSmemSfB */ 4 -, /* mNumStagesTmemSfA */ 1 -, /* mNumStagesTmemSfB */ 4 +, /* mNumStagesSmemSfA */ 5 +, /* mNumStagesSmemSfB */ 5 +, /* mNumStagesTmemSfA */ 5 +, /* mNumStagesTmemSfB */ 5 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -38443,59 +39412,59 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfBlockSizeC */ 16 , /* mSfLayoutA */ trtllm::gen::SfLayout(3) , /* mSfLayoutB */ trtllm::gen::SfLayout(0) -, /* mSfLayoutC */ trtllm::gen::SfLayout(3) +, /* mSfLayoutC */ trtllm::gen::SfLayout(1) , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) , /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 256 +, /* mTileK */ 512 , /* mTileM */ 128 -, /* mTileN */ 256 -, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTileN */ 16 +, /* mTileScheduler */ gemm::TileScheduler(0) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 , /* mUseCustomMmaSchedule */ 1 , /* mUseDeepSeekFp8 */ 0 , /* mUseFlexibleClusterDims */ 0 , /* mUseHoistTryWaitForCustomMmaSchedule */ 0 -, /* mUseMaxTmemOverlap */ 1 +, /* mUseMaxTmemOverlap */ 0 , /* mUsePerTokenSfA */ 0 , /* mUsePerTokenSfB */ 0 , /* mUseShuffledMatrix */ 1 , /* mUseTmaStore */ 1 , /* mUseTwoTmaLoadWarps */ 1 , /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 0 +, /* mUseUnrollLoop2xForMma */ 1 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 256 +, /* mValidK */ 1024 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(3) +, /* mActType */ gemmGatedAct::ActType(0) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 0 +, /* mFusedAct */ 1 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 , /* mNumBatches */ 128 , /* mNumRegsPerThreadLoadA */ 0 -, /* mNumRegsPerThreadLoadB */ 160 +, /* mNumRegsPerThreadLoadB */ 0 , /* mNumRegsPerThreadLoadSfA */ 0 , /* mNumRegsPerThreadLoadSfB */ 0 , /* mNumTokens */ 2 , /* mNumWarpsLoadA */ 0 -, /* mNumWarpsLoadB */ 4 +, /* mNumWarpsLoadB */ 0 , /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 -, /* mRouteImpl */ batchedGemm::RouteImpl(1) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_silu_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_lbW8_lsfbW4_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_silu_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_lbW8_lsfbW4_dynB_sm100f_cubin_len, 172696, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_silu_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_lbW8_lsfbW4_dynB_sm100f", 768, "9edb847256c8ba18cad06eb83d7317d98470992688f6dc9c8c740d4f81771a58", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 203288, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 640, "90654e37289b63a865aa03aa465c46f90ddac963e4eba8e5dbc257e3bd30397f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -38511,19 +39480,19 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(3) +, /* mEltwiseActType */ gemm::EltwiseActType(2) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 , /* mEpilogueLdtmBits */ 256 , /* mEpilogueTileM */ 128 -, /* mEpilogueTileN */ 64 +, /* mEpilogueTileN */ 16 , /* mFallbackClusterDimX */ 1 , /* mFallbackClusterDimY */ 1 , /* mFallbackClusterDimZ */ 1 , /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) , /* mFuseLoadSfTask */ 0 -, /* mFuseUtccpWithUtcmma */ 1 +, /* mFuseUtccpWithUtcmma */ 0 , /* mGridTriggerSecondaryA */ 0 , /* mGridTriggerSecondaryB */ 1 , /* mGridWaitForPrimaryEarlyExit */ 1 @@ -38531,7 +39500,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 256 +, /* mK */ 1024 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) @@ -38539,26 +39508,26 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) , /* mMmaM */ 256 -, /* mMmaN */ 256 -, /* mMmaTileK */ 256 +, /* mMmaN */ 16 +, /* mMmaTileK */ 512 , /* mMockAllReduce */ 0 , /* mN */ 256 -, /* mNumEpilogueWarps */ 8 +, /* mNumEpilogueWarps */ 4 , /* mNumRegsCastAWarps */ 0 , /* mNumRegsCopySfLdsSttm */ 0 , /* mNumRegsCopySparsityInfo */ 0 -, /* mNumRegsPerThreadEpilogueWarp */ 144 -, /* mNumRegsPerThreadNonEpilogueWarp */ 88 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 4 -, /* mNumStagesB */ 4 +, /* mNumStagesA */ 5 +, /* mNumStagesB */ 5 , /* mNumStagesMma */ 1 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 1 -, /* mNumStagesSmemSfA */ 4 -, /* mNumStagesSmemSfB */ 4 -, /* mNumStagesTmemSfA */ 1 -, /* mNumStagesTmemSfB */ 4 +, /* mNumStagesSmemSfA */ 5 +, /* mNumStagesSmemSfB */ 5 +, /* mNumStagesTmemSfA */ 5 +, /* mNumStagesTmemSfB */ 5 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -38569,33 +39538,33 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfBlockSizeC */ 16 , /* mSfLayoutA */ trtllm::gen::SfLayout(3) , /* mSfLayoutB */ trtllm::gen::SfLayout(0) -, /* mSfLayoutC */ trtllm::gen::SfLayout(3) +, /* mSfLayoutC */ trtllm::gen::SfLayout(1) , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) , /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 256 +, /* mTileK */ 512 , /* mTileM */ 128 -, /* mTileN */ 256 -, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTileN */ 16 +, /* mTileScheduler */ gemm::TileScheduler(0) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 , /* mUseCustomMmaSchedule */ 1 , /* mUseDeepSeekFp8 */ 0 , /* mUseFlexibleClusterDims */ 0 , /* mUseHoistTryWaitForCustomMmaSchedule */ 0 -, /* mUseMaxTmemOverlap */ 1 +, /* mUseMaxTmemOverlap */ 0 , /* mUsePerTokenSfA */ 0 , /* mUsePerTokenSfB */ 0 , /* mUseShuffledMatrix */ 1 , /* mUseTmaStore */ 1 , /* mUseTwoTmaLoadWarps */ 1 , /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 0 +, /* mUseUnrollLoop2xForMma */ 1 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 256 +, /* mValidK */ 1024 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(3) +, /* mActType */ gemmGatedAct::ActType(0) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} @@ -38607,21 +39576,21 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mIsUniformNumTokensPerBatch */ 0 , /* mNumBatches */ 128 , /* mNumRegsPerThreadLoadA */ 0 -, /* mNumRegsPerThreadLoadB */ 32 +, /* mNumRegsPerThreadLoadB */ 0 , /* mNumRegsPerThreadLoadSfA */ 0 -, /* mNumRegsPerThreadLoadSfB */ 40 +, /* mNumRegsPerThreadLoadSfB */ 0 , /* mNumTokens */ 2 , /* mNumWarpsLoadA */ 0 -, /* mNumWarpsLoadB */ 8 +, /* mNumWarpsLoadB */ 0 , /* mNumWarpsLoadSfA */ 0 -, /* mNumWarpsLoadSfB */ 4 +, /* mNumWarpsLoadSfB */ 0 , /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(3)} +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_silu_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_silu_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_dynB_sm100f_cubin_len, 168600, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_silu_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_dynB_sm100f", 384, "644b8da7b105b67fdb91a25c3b7778db1cf317daa3ab18deacbbab5cbf418b62", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 203288, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 640, "5db7e8bcf771c2ba241fd4e58a94a3b536b7e7ee9d24d9901b0cd34721e664b7", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -38643,12 +39612,138 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mEpilogueLdtmDps */ 16 , /* mEpilogueLdtmBits */ 256 , /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 16 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 0 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 1024 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 256 +, /* mMmaN */ 16 +, /* mMmaTileK */ 512 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 5 +, /* mNumStagesB */ 5 +, /* mNumStagesMma */ 1 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 1 +, /* mNumStagesSmemSfA */ 5 +, /* mNumStagesSmemSfB */ 5 +, /* mNumStagesTmemSfA */ 5 +, /* mNumStagesTmemSfB */ 5 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(3) +, /* mSfLayoutB */ trtllm::gen::SfLayout(0) +, /* mSfLayoutC */ trtllm::gen::SfLayout(1) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 512 +, /* mTileM */ 128 +, /* mTileN */ 16 +, /* mTileScheduler */ gemm::TileScheduler(0) +, /* mTransposeMmaOutput */ 1 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 0 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 1 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 1024 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(3) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 0 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 0 +, /* mNumRegsPerThreadLoadSfA */ 0 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 0 +, /* mNumWarpsLoadSfA */ 0 +, /* mNumWarpsLoadSfB */ 0 +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_geGlu_lbW8_lsfbW4_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_geGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len, 172696, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_geGlu_lbW8_lsfbW4_dynB_sm100f", 768, "cc52ac986188e4e3a7f2230799ba7812f9221381507d552e30fc1ad1cb532219", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(1) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 2 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 1 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 16 +, /* mEpilogueLdtmBits */ 256 +, /* mEpilogueTileM */ 128 , /* mEpilogueTileN */ 64 , /* mFallbackClusterDimX */ 1 , /* mFallbackClusterDimY */ 1 , /* mFallbackClusterDimZ */ 1 , /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) -, /* mFuseLoadSfTask */ 1 +, /* mFuseLoadSfTask */ 0 , /* mFuseUtccpWithUtcmma */ 1 , /* mGridTriggerSecondaryA */ 0 , /* mGridTriggerSecondaryB */ 1 @@ -38669,11 +39764,11 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mMmaTileK */ 256 , /* mMockAllReduce */ 0 , /* mN */ 256 -, /* mNumEpilogueWarps */ 4 +, /* mNumEpilogueWarps */ 8 , /* mNumRegsCastAWarps */ 0 , /* mNumRegsCopySfLdsSttm */ 0 , /* mNumRegsCopySparsityInfo */ 0 -, /* mNumRegsPerThreadEpilogueWarp */ 232 +, /* mNumRegsPerThreadEpilogueWarp */ 144 , /* mNumRegsPerThreadNonEpilogueWarp */ 88 , /* mNumSlicesForSplitK */ 1 , /* mNumStagesA */ 4 @@ -38721,41 +39816,41 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mValidN */ 256 , /* mValidK */ 256 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(3) +, /* mActType */ gemmGatedAct::ActType(1) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 0 +, /* mFusedAct */ 1 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 , /* mNumBatches */ 128 , /* mNumRegsPerThreadLoadA */ 0 -, /* mNumRegsPerThreadLoadB */ 160 +, /* mNumRegsPerThreadLoadB */ 32 , /* mNumRegsPerThreadLoadSfA */ 0 -, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumRegsPerThreadLoadSfB */ 40 , /* mNumTokens */ 2 , /* mNumWarpsLoadA */ 0 -, /* mNumWarpsLoadB */ 4 +, /* mNumWarpsLoadB */ 8 , /* mNumWarpsLoadSfA */ 0 -, /* mNumWarpsLoadSfB */ 0 -, /* mRouteImpl */ batchedGemm::RouteImpl(1) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mNumWarpsLoadSfB */ 4 +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(3)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_geGlu_laW8_lsfaW4_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_geGlu_laW8_lsfaW4_dynB_sm100f_cubin_len, 203416, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_geGlu_laW8_lsfaW4_dynB_sm100f", 640, "a86c8dee26a642c62a097020e049304873a22aaec872ecb6c85f96ba5596e562", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len, 172696, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f", 768, "907ca5091217f4c4aef9dce891a266e8457c856f9f1f97a8adba2aaadb627b8d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) -, /* mBiasType */ gemm::BiasType(2) +, /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 , /* mClusterDimX */ 2 , /* mClusterDimY */ 1 , /* mClusterDimZ */ 1 -, /* mCtaSwizzleType */ gemm::CtaSwizzleType(1) +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) , /* mDtypeAcc */ trtllm::gen::Dtype(1056776) , /* mDtypeA */ trtllm::gen::Dtype(17826818) , /* mDtypeB */ trtllm::gen::Dtype(17826818) @@ -38766,8 +39861,8 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mEltwiseActType */ gemm::EltwiseActType(0) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 -, /* mEpilogueLdtmDps */ 32 -, /* mEpilogueLdtmBits */ 32 +, /* mEpilogueLdtmDps */ 16 +, /* mEpilogueLdtmBits */ 256 , /* mEpilogueTileM */ 128 , /* mEpilogueTileN */ 64 , /* mFallbackClusterDimX */ 1 @@ -38795,22 +39890,22 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mMmaTileK */ 256 , /* mMockAllReduce */ 0 , /* mN */ 256 -, /* mNumEpilogueWarps */ 4 +, /* mNumEpilogueWarps */ 8 , /* mNumRegsCastAWarps */ 0 , /* mNumRegsCopySfLdsSttm */ 0 , /* mNumRegsCopySparsityInfo */ 0 -, /* mNumRegsPerThreadEpilogueWarp */ 256 -, /* mNumRegsPerThreadNonEpilogueWarp */ 104 +, /* mNumRegsPerThreadEpilogueWarp */ 144 +, /* mNumRegsPerThreadNonEpilogueWarp */ 88 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 5 -, /* mNumStagesB */ 5 +, /* mNumStagesA */ 4 +, /* mNumStagesB */ 4 , /* mNumStagesMma */ 1 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 1 -, /* mNumStagesSmemSfA */ 5 -, /* mNumStagesSmemSfB */ 5 -, /* mNumStagesTmemSfA */ 5 -, /* mNumStagesTmemSfB */ 1 +, /* mNumStagesSmemSfA */ 4 +, /* mNumStagesSmemSfB */ 4 +, /* mNumStagesTmemSfA */ 1 +, /* mNumStagesTmemSfB */ 4 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -38819,8 +39914,8 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfBlockSizeA */ 16 , /* mSfBlockSizeB */ 16 , /* mSfBlockSizeC */ 16 -, /* mSfLayoutA */ trtllm::gen::SfLayout(0) -, /* mSfLayoutB */ trtllm::gen::SfLayout(3) +, /* mSfLayoutA */ trtllm::gen::SfLayout(3) +, /* mSfLayoutB */ trtllm::gen::SfLayout(0) , /* mSfLayoutC */ trtllm::gen::SfLayout(3) , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) @@ -38829,7 +39924,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mTileM */ 128 , /* mTileN */ 256 , /* mTileScheduler */ gemm::TileScheduler(1) -, /* mTransposeMmaOutput */ 0 +, /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 , /* mUseCustomMmaSchedule */ 1 , /* mUseDeepSeekFp8 */ 0 @@ -38847,41 +39942,41 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mValidN */ 256 , /* mValidK */ 256 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(1) +, /* mActType */ gemmGatedAct::ActType(2) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} -, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(0) +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 , /* mFusedAct */ 1 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 , /* mNumBatches */ 128 -, /* mNumRegsPerThreadLoadA */ 32 -, /* mNumRegsPerThreadLoadB */ 0 -, /* mNumRegsPerThreadLoadSfA */ 40 -, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 32 +, /* mNumRegsPerThreadLoadSfA */ 0 +, /* mNumRegsPerThreadLoadSfB */ 40 , /* mNumTokens */ 2 -, /* mNumWarpsLoadA */ 8 -, /* mNumWarpsLoadB */ 0 -, /* mNumWarpsLoadSfA */ 4 -, /* mNumWarpsLoadSfB */ 0 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 8 +, /* mNumWarpsLoadSfA */ 0 +, /* mNumWarpsLoadSfB */ 4 , /* mRouteImpl */ batchedGemm::RouteImpl(2) , /* mRouteSfsImpl */ {batchedGemm::RouteImpl(3)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_swiGlu_laW8_lsfaW4_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_swiGlu_laW8_lsfaW4_dynB_sm100f_cubin_len, 203416, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_swiGlu_laW8_lsfaW4_dynB_sm100f", 640, "d116020528832ac1471b8ece21a3989029f00bf9c954c8290a1239cdfaefc502", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len, 172696, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f", 768, "bc8216d8549b992f4fcf48775d9a6398a4fe54d141e3371f5ff17cf54e6dadd3", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) -, /* mBiasType */ gemm::BiasType(2) +, /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 , /* mClusterDimX */ 2 , /* mClusterDimY */ 1 , /* mClusterDimZ */ 1 -, /* mCtaSwizzleType */ gemm::CtaSwizzleType(1) +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) , /* mDtypeAcc */ trtllm::gen::Dtype(1056776) , /* mDtypeA */ trtllm::gen::Dtype(17826818) , /* mDtypeB */ trtllm::gen::Dtype(17826818) @@ -38892,8 +39987,8 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mEltwiseActType */ gemm::EltwiseActType(0) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 -, /* mEpilogueLdtmDps */ 32 -, /* mEpilogueLdtmBits */ 32 +, /* mEpilogueLdtmDps */ 16 +, /* mEpilogueLdtmBits */ 256 , /* mEpilogueTileM */ 128 , /* mEpilogueTileN */ 64 , /* mFallbackClusterDimX */ 1 @@ -38921,22 +40016,22 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mMmaTileK */ 256 , /* mMockAllReduce */ 0 , /* mN */ 256 -, /* mNumEpilogueWarps */ 4 +, /* mNumEpilogueWarps */ 8 , /* mNumRegsCastAWarps */ 0 , /* mNumRegsCopySfLdsSttm */ 0 , /* mNumRegsCopySparsityInfo */ 0 -, /* mNumRegsPerThreadEpilogueWarp */ 256 -, /* mNumRegsPerThreadNonEpilogueWarp */ 104 +, /* mNumRegsPerThreadEpilogueWarp */ 144 +, /* mNumRegsPerThreadNonEpilogueWarp */ 88 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 5 -, /* mNumStagesB */ 5 +, /* mNumStagesA */ 4 +, /* mNumStagesB */ 4 , /* mNumStagesMma */ 1 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 1 -, /* mNumStagesSmemSfA */ 5 -, /* mNumStagesSmemSfB */ 5 -, /* mNumStagesTmemSfA */ 5 -, /* mNumStagesTmemSfB */ 1 +, /* mNumStagesSmemSfA */ 4 +, /* mNumStagesSmemSfB */ 4 +, /* mNumStagesTmemSfA */ 1 +, /* mNumStagesTmemSfB */ 4 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -38945,8 +40040,8 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfBlockSizeA */ 16 , /* mSfBlockSizeB */ 16 , /* mSfBlockSizeC */ 16 -, /* mSfLayoutA */ trtllm::gen::SfLayout(0) -, /* mSfLayoutB */ trtllm::gen::SfLayout(3) +, /* mSfLayoutA */ trtllm::gen::SfLayout(3) +, /* mSfLayoutB */ trtllm::gen::SfLayout(0) , /* mSfLayoutC */ trtllm::gen::SfLayout(3) , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) @@ -38955,7 +40050,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mTileM */ 128 , /* mTileN */ 256 , /* mTileScheduler */ gemm::TileScheduler(1) -, /* mTransposeMmaOutput */ 0 +, /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 , /* mUseCustomMmaSchedule */ 1 , /* mUseDeepSeekFp8 */ 0 @@ -38977,37 +40072,37 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} -, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(0) +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 , /* mFusedAct */ 1 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 , /* mNumBatches */ 128 -, /* mNumRegsPerThreadLoadA */ 32 -, /* mNumRegsPerThreadLoadB */ 0 -, /* mNumRegsPerThreadLoadSfA */ 40 -, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 32 +, /* mNumRegsPerThreadLoadSfA */ 0 +, /* mNumRegsPerThreadLoadSfB */ 40 , /* mNumTokens */ 2 -, /* mNumWarpsLoadA */ 8 -, /* mNumWarpsLoadB */ 0 -, /* mNumWarpsLoadSfA */ 4 -, /* mNumWarpsLoadSfB */ 0 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 8 +, /* mNumWarpsLoadSfA */ 0 +, /* mNumWarpsLoadSfB */ 4 , /* mRouteImpl */ batchedGemm::RouteImpl(2) , /* mRouteSfsImpl */ {batchedGemm::RouteImpl(3)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_relu2_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_laW8_lsfaW4_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_relu2_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_laW8_lsfaW4_dynB_sm100f_cubin_len, 203416, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_relu2_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_laW8_lsfaW4_dynB_sm100f", 640, "0e233531b1ab78c69085cc5c28a8aa376609729d512bd5201dd82d507f1256fe", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_geGlu_lbW4_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_geGlu_lbW4_dynB_sm100f_cubin_len, 168600, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_geGlu_lbW4_dynB_sm100f", 384, "a1e1d735d80b0258c822e9aaa05796ae7e07d41cb5a1d4a2114f6c550ce6baef", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) -, /* mBiasType */ gemm::BiasType(2) +, /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 , /* mClusterDimX */ 2 , /* mClusterDimY */ 1 , /* mClusterDimZ */ 1 -, /* mCtaSwizzleType */ gemm::CtaSwizzleType(1) +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) , /* mDtypeAcc */ trtllm::gen::Dtype(1056776) , /* mDtypeA */ trtllm::gen::Dtype(17826818) , /* mDtypeB */ trtllm::gen::Dtype(17826818) @@ -39015,18 +40110,18 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(2) +, /* mEltwiseActType */ gemm::EltwiseActType(0) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 -, /* mEpilogueLdtmDps */ 32 -, /* mEpilogueLdtmBits */ 32 +, /* mEpilogueLdtmDps */ 16 +, /* mEpilogueLdtmBits */ 256 , /* mEpilogueTileM */ 128 , /* mEpilogueTileN */ 64 , /* mFallbackClusterDimX */ 1 , /* mFallbackClusterDimY */ 1 , /* mFallbackClusterDimZ */ 1 , /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) -, /* mFuseLoadSfTask */ 0 +, /* mFuseLoadSfTask */ 1 , /* mFuseUtccpWithUtcmma */ 1 , /* mGridTriggerSecondaryA */ 0 , /* mGridTriggerSecondaryB */ 1 @@ -39051,18 +40146,18 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCastAWarps */ 0 , /* mNumRegsCopySfLdsSttm */ 0 , /* mNumRegsCopySparsityInfo */ 0 -, /* mNumRegsPerThreadEpilogueWarp */ 256 -, /* mNumRegsPerThreadNonEpilogueWarp */ 104 +, /* mNumRegsPerThreadEpilogueWarp */ 232 +, /* mNumRegsPerThreadNonEpilogueWarp */ 88 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 5 -, /* mNumStagesB */ 5 +, /* mNumStagesA */ 4 +, /* mNumStagesB */ 4 , /* mNumStagesMma */ 1 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 1 -, /* mNumStagesSmemSfA */ 5 -, /* mNumStagesSmemSfB */ 5 -, /* mNumStagesTmemSfA */ 5 -, /* mNumStagesTmemSfB */ 1 +, /* mNumStagesSmemSfA */ 4 +, /* mNumStagesSmemSfB */ 4 +, /* mNumStagesTmemSfA */ 1 +, /* mNumStagesTmemSfB */ 4 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -39071,8 +40166,8 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfBlockSizeA */ 16 , /* mSfBlockSizeB */ 16 , /* mSfBlockSizeC */ 16 -, /* mSfLayoutA */ trtllm::gen::SfLayout(0) -, /* mSfLayoutB */ trtllm::gen::SfLayout(3) +, /* mSfLayoutA */ trtllm::gen::SfLayout(3) +, /* mSfLayoutB */ trtllm::gen::SfLayout(0) , /* mSfLayoutC */ trtllm::gen::SfLayout(3) , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) @@ -39081,7 +40176,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mTileM */ 128 , /* mTileN */ 256 , /* mTileScheduler */ gemm::TileScheduler(1) -, /* mTransposeMmaOutput */ 0 +, /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 , /* mUseCustomMmaSchedule */ 1 , /* mUseDeepSeekFp8 */ 0 @@ -39099,41 +40194,41 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mValidN */ 256 , /* mValidK */ 256 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(3) +, /* mActType */ gemmGatedAct::ActType(1) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} -, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(0) +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 0 +, /* mFusedAct */ 1 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 , /* mNumBatches */ 128 -, /* mNumRegsPerThreadLoadA */ 32 -, /* mNumRegsPerThreadLoadB */ 0 -, /* mNumRegsPerThreadLoadSfA */ 40 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 160 +, /* mNumRegsPerThreadLoadSfA */ 0 , /* mNumRegsPerThreadLoadSfB */ 0 , /* mNumTokens */ 2 -, /* mNumWarpsLoadA */ 8 -, /* mNumWarpsLoadB */ 0 -, /* mNumWarpsLoadSfA */ 4 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 4 +, /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 -, /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(3)} +, /* mRouteImpl */ batchedGemm::RouteImpl(1) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_silu_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_laW8_lsfaW4_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_silu_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_laW8_lsfaW4_dynB_sm100f_cubin_len, 203416, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_silu_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_laW8_lsfaW4_dynB_sm100f", 640, "ea3b2a39ddf253f7d9d7d9d57cceb030d32665ca5d4cf02bbfd64bc5fc727e8c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_siTuGlu_lbW4_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_siTuGlu_lbW4_dynB_sm100f_cubin_len, 168600, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_siTuGlu_lbW4_dynB_sm100f", 384, "e3df7605750b2c3307106a910b88d7833844561a89469fd01d371b2e2a775b46", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) -, /* mBiasType */ gemm::BiasType(2) +, /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 , /* mClusterDimX */ 2 , /* mClusterDimY */ 1 , /* mClusterDimZ */ 1 -, /* mCtaSwizzleType */ gemm::CtaSwizzleType(1) +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) , /* mDtypeAcc */ trtllm::gen::Dtype(1056776) , /* mDtypeA */ trtllm::gen::Dtype(17826818) , /* mDtypeB */ trtllm::gen::Dtype(17826818) @@ -39141,18 +40236,18 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(3) +, /* mEltwiseActType */ gemm::EltwiseActType(0) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 -, /* mEpilogueLdtmDps */ 32 -, /* mEpilogueLdtmBits */ 32 +, /* mEpilogueLdtmDps */ 16 +, /* mEpilogueLdtmBits */ 256 , /* mEpilogueTileM */ 128 , /* mEpilogueTileN */ 64 , /* mFallbackClusterDimX */ 1 , /* mFallbackClusterDimY */ 1 , /* mFallbackClusterDimZ */ 1 , /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) -, /* mFuseLoadSfTask */ 0 +, /* mFuseLoadSfTask */ 1 , /* mFuseUtccpWithUtcmma */ 1 , /* mGridTriggerSecondaryA */ 0 , /* mGridTriggerSecondaryB */ 1 @@ -39177,18 +40272,18 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCastAWarps */ 0 , /* mNumRegsCopySfLdsSttm */ 0 , /* mNumRegsCopySparsityInfo */ 0 -, /* mNumRegsPerThreadEpilogueWarp */ 256 -, /* mNumRegsPerThreadNonEpilogueWarp */ 104 +, /* mNumRegsPerThreadEpilogueWarp */ 232 +, /* mNumRegsPerThreadNonEpilogueWarp */ 88 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 5 -, /* mNumStagesB */ 5 +, /* mNumStagesA */ 4 +, /* mNumStagesB */ 4 , /* mNumStagesMma */ 1 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 1 -, /* mNumStagesSmemSfA */ 5 -, /* mNumStagesSmemSfB */ 5 -, /* mNumStagesTmemSfA */ 5 -, /* mNumStagesTmemSfB */ 1 +, /* mNumStagesSmemSfA */ 4 +, /* mNumStagesSmemSfB */ 4 +, /* mNumStagesTmemSfA */ 1 +, /* mNumStagesTmemSfB */ 4 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -39197,8 +40292,8 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfBlockSizeA */ 16 , /* mSfBlockSizeB */ 16 , /* mSfBlockSizeC */ 16 -, /* mSfLayoutA */ trtllm::gen::SfLayout(0) -, /* mSfLayoutB */ trtllm::gen::SfLayout(3) +, /* mSfLayoutA */ trtllm::gen::SfLayout(3) +, /* mSfLayoutB */ trtllm::gen::SfLayout(0) , /* mSfLayoutC */ trtllm::gen::SfLayout(3) , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) @@ -39207,7 +40302,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mTileM */ 128 , /* mTileN */ 256 , /* mTileScheduler */ gemm::TileScheduler(1) -, /* mTransposeMmaOutput */ 0 +, /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 , /* mUseCustomMmaSchedule */ 1 , /* mUseDeepSeekFp8 */ 0 @@ -39225,38 +40320,38 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mValidN */ 256 , /* mValidK */ 256 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(3) +, /* mActType */ gemmGatedAct::ActType(2) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} -, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(0) +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 0 +, /* mFusedAct */ 1 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 , /* mNumBatches */ 128 -, /* mNumRegsPerThreadLoadA */ 32 -, /* mNumRegsPerThreadLoadB */ 0 -, /* mNumRegsPerThreadLoadSfA */ 40 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 160 +, /* mNumRegsPerThreadLoadSfA */ 0 , /* mNumRegsPerThreadLoadSfB */ 0 , /* mNumTokens */ 2 -, /* mNumWarpsLoadA */ 8 -, /* mNumWarpsLoadB */ 0 -, /* mNumWarpsLoadSfA */ 4 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 4 +, /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 -, /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(3)} +, /* mRouteImpl */ batchedGemm::RouteImpl(1) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 215184, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f", 640, "3fa46629bbcf794937dcfed33fdd12b19a1d15ef0c61d18ba5e827eccb52bfd5", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_dynB_sm100f_cubin_len, 168600, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_dynB_sm100f", 384, "e0e5eff777711ebfa80079c16b26d09d4595886b000065b272e0126603f766c5", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 -, /* mClusterDimX */ 1 +, /* mClusterDimX */ 2 , /* mClusterDimY */ 1 , /* mClusterDimZ */ 1 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) @@ -39273,13 +40368,13 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mEpilogueLdtmDps */ 16 , /* mEpilogueLdtmBits */ 256 , /* mEpilogueTileM */ 128 -, /* mEpilogueTileN */ 32 +, /* mEpilogueTileN */ 64 , /* mFallbackClusterDimX */ 1 , /* mFallbackClusterDimY */ 1 , /* mFallbackClusterDimZ */ 1 , /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) -, /* mFuseLoadSfTask */ 0 -, /* mFuseUtccpWithUtcmma */ 0 +, /* mFuseLoadSfTask */ 1 +, /* mFuseUtccpWithUtcmma */ 1 , /* mGridTriggerSecondaryA */ 0 , /* mGridTriggerSecondaryB */ 1 , /* mGridWaitForPrimaryEarlyExit */ 1 @@ -39294,8 +40389,8 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mM */ 256 , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) -, /* mMmaM */ 128 -, /* mMmaN */ 32 +, /* mMmaM */ 256 +, /* mMmaN */ 256 , /* mMmaTileK */ 256 , /* mMockAllReduce */ 0 , /* mN */ 256 @@ -39303,18 +40398,18 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCastAWarps */ 0 , /* mNumRegsCopySfLdsSttm */ 0 , /* mNumRegsCopySparsityInfo */ 0 -, /* mNumRegsPerThreadEpilogueWarp */ 0 -, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 232 +, /* mNumRegsPerThreadNonEpilogueWarp */ 88 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 9 -, /* mNumStagesB */ 9 -, /* mNumStagesMma */ 2 +, /* mNumStagesA */ 4 +, /* mNumStagesB */ 4 +, /* mNumStagesMma */ 1 , /* mNumStagesMmaWithinWorkTile */ 1 -, /* mNumStagesMmaAcrossWorkTile */ 2 -, /* mNumStagesSmemSfA */ 9 -, /* mNumStagesSmemSfB */ 9 -, /* mNumStagesTmemSfA */ 9 -, /* mNumStagesTmemSfB */ 9 +, /* mNumStagesMmaAcrossWorkTile */ 1 +, /* mNumStagesSmemSfA */ 4 +, /* mNumStagesSmemSfB */ 4 +, /* mNumStagesTmemSfA */ 1 +, /* mNumStagesTmemSfB */ 4 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -39325,13 +40420,13 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfBlockSizeC */ 16 , /* mSfLayoutA */ trtllm::gen::SfLayout(3) , /* mSfLayoutB */ trtllm::gen::SfLayout(0) -, /* mSfLayoutC */ trtllm::gen::SfLayout(1) +, /* mSfLayoutC */ trtllm::gen::SfLayout(3) , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) , /* mSplitK */ gemm::SplitK(0) , /* mTileK */ 256 , /* mTileM */ 128 -, /* mTileN */ 32 +, /* mTileN */ 256 , /* mTileScheduler */ gemm::TileScheduler(1) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 @@ -39339,7 +40434,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseDeepSeekFp8 */ 0 , /* mUseFlexibleClusterDims */ 0 , /* mUseHoistTryWaitForCustomMmaSchedule */ 0 -, /* mUseMaxTmemOverlap */ 0 +, /* mUseMaxTmemOverlap */ 1 , /* mUsePerTokenSfA */ 0 , /* mUsePerTokenSfB */ 0 , /* mUseShuffledMatrix */ 1 @@ -39351,7 +40446,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mValidN */ 256 , /* mValidK */ 256 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(1) +, /* mActType */ gemmGatedAct::ActType(0) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} @@ -39363,26 +40458,4814 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mIsUniformNumTokensPerBatch */ 0 , /* mNumBatches */ 128 , /* mNumRegsPerThreadLoadA */ 0 -, /* mNumRegsPerThreadLoadB */ 0 +, /* mNumRegsPerThreadLoadB */ 160 , /* mNumRegsPerThreadLoadSfA */ 0 , /* mNumRegsPerThreadLoadSfB */ 0 , /* mNumTokens */ 2 , /* mNumWarpsLoadA */ 0 -, /* mNumWarpsLoadB */ 0 +, /* mNumWarpsLoadB */ 4 , /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 -, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteImpl */ batchedGemm::RouteImpl(1) , /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 215184, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "e31bc1891ac54cc71a79de912b1642d0cf3e6db14218108041d5cb0e68ff4bc8", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_relu2_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_lbW8_lsfbW4_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_relu2_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_lbW8_lsfbW4_dynB_sm100f_cubin_len, 172696, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_relu2_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_lbW8_lsfbW4_dynB_sm100f", 768, "8d9e787da8d00b6a185ab4933518796068a88463d5a6f8a5dae0011bf8d124f1", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 -, /* mClusterDimX */ 1 +, /* mClusterDimX */ 2 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 1 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(2) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 16 +, /* mEpilogueLdtmBits */ 256 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 64 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 1 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 256 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 256 +, /* mMmaN */ 256 +, /* mMmaTileK */ 256 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 8 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 144 +, /* mNumRegsPerThreadNonEpilogueWarp */ 88 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 4 +, /* mNumStagesB */ 4 +, /* mNumStagesMma */ 1 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 1 +, /* mNumStagesSmemSfA */ 4 +, /* mNumStagesSmemSfB */ 4 +, /* mNumStagesTmemSfA */ 1 +, /* mNumStagesTmemSfB */ 4 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(3) +, /* mSfLayoutB */ trtllm::gen::SfLayout(0) +, /* mSfLayoutC */ trtllm::gen::SfLayout(3) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 256 +, /* mTileM */ 128 +, /* mTileN */ 256 +, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTransposeMmaOutput */ 1 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 1 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 0 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 256 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(3) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 0 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 32 +, /* mNumRegsPerThreadLoadSfA */ 0 +, /* mNumRegsPerThreadLoadSfB */ 40 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 8 +, /* mNumWarpsLoadSfA */ 0 +, /* mNumWarpsLoadSfB */ 4 +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(3)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_relu2_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_relu2_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_dynB_sm100f_cubin_len, 168600, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_relu2_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_dynB_sm100f", 384, "6bc0606888a728a8d4b0011f8181456dc5300db1faaa8d8fe7b530419372658e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(1) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 2 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 1 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(2) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 16 +, /* mEpilogueLdtmBits */ 256 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 64 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 1 +, /* mFuseUtccpWithUtcmma */ 1 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 256 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 256 +, /* mMmaN */ 256 +, /* mMmaTileK */ 256 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 232 +, /* mNumRegsPerThreadNonEpilogueWarp */ 88 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 4 +, /* mNumStagesB */ 4 +, /* mNumStagesMma */ 1 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 1 +, /* mNumStagesSmemSfA */ 4 +, /* mNumStagesSmemSfB */ 4 +, /* mNumStagesTmemSfA */ 1 +, /* mNumStagesTmemSfB */ 4 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(3) +, /* mSfLayoutB */ trtllm::gen::SfLayout(0) +, /* mSfLayoutC */ trtllm::gen::SfLayout(3) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 256 +, /* mTileM */ 128 +, /* mTileN */ 256 +, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTransposeMmaOutput */ 1 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 1 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 0 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 256 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(3) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 0 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 160 +, /* mNumRegsPerThreadLoadSfA */ 0 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 4 +, /* mNumWarpsLoadSfA */ 0 +, /* mNumWarpsLoadSfB */ 0 +, /* mRouteImpl */ batchedGemm::RouteImpl(1) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_silu_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_lbW8_lsfbW4_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_silu_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_lbW8_lsfbW4_dynB_sm100f_cubin_len, 172696, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_silu_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_lbW8_lsfbW4_dynB_sm100f", 768, "7907e23f4499fdf7ceeb5e25519628dfd4f0e483332ea39eaa4db184e5f00d1f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(1) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 2 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 1 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(3) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 16 +, /* mEpilogueLdtmBits */ 256 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 64 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 1 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 256 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 256 +, /* mMmaN */ 256 +, /* mMmaTileK */ 256 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 8 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 144 +, /* mNumRegsPerThreadNonEpilogueWarp */ 88 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 4 +, /* mNumStagesB */ 4 +, /* mNumStagesMma */ 1 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 1 +, /* mNumStagesSmemSfA */ 4 +, /* mNumStagesSmemSfB */ 4 +, /* mNumStagesTmemSfA */ 1 +, /* mNumStagesTmemSfB */ 4 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(3) +, /* mSfLayoutB */ trtllm::gen::SfLayout(0) +, /* mSfLayoutC */ trtllm::gen::SfLayout(3) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 256 +, /* mTileM */ 128 +, /* mTileN */ 256 +, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTransposeMmaOutput */ 1 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 1 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 0 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 256 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(3) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 0 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 32 +, /* mNumRegsPerThreadLoadSfA */ 0 +, /* mNumRegsPerThreadLoadSfB */ 40 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 8 +, /* mNumWarpsLoadSfA */ 0 +, /* mNumWarpsLoadSfB */ 4 +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(3)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_silu_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_silu_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_dynB_sm100f_cubin_len, 168600, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_silu_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_dynB_sm100f", 384, "86eac8fc35d88b78842de7df743dbe7405dbf31a3d31b743b91dfd2eafd4c0ea", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(1) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 2 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 1 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(3) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 16 +, /* mEpilogueLdtmBits */ 256 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 64 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 1 +, /* mFuseUtccpWithUtcmma */ 1 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 256 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 256 +, /* mMmaN */ 256 +, /* mMmaTileK */ 256 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 232 +, /* mNumRegsPerThreadNonEpilogueWarp */ 88 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 4 +, /* mNumStagesB */ 4 +, /* mNumStagesMma */ 1 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 1 +, /* mNumStagesSmemSfA */ 4 +, /* mNumStagesSmemSfB */ 4 +, /* mNumStagesTmemSfA */ 1 +, /* mNumStagesTmemSfB */ 4 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(3) +, /* mSfLayoutB */ trtllm::gen::SfLayout(0) +, /* mSfLayoutC */ trtllm::gen::SfLayout(3) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 256 +, /* mTileM */ 128 +, /* mTileN */ 256 +, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTransposeMmaOutput */ 1 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 1 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 0 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 256 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(3) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 0 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 160 +, /* mNumRegsPerThreadLoadSfA */ 0 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 4 +, /* mNumWarpsLoadSfA */ 0 +, /* mNumWarpsLoadSfB */ 0 +, /* mRouteImpl */ batchedGemm::RouteImpl(1) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_geGlu_laW8_lsfaW4_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_geGlu_laW8_lsfaW4_dynB_sm100f_cubin_len, 203416, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_geGlu_laW8_lsfaW4_dynB_sm100f", 640, "059b8f94e347331297f5a6cc8978aa4c55dddb2273c9066141bc3270e9d6c4fc", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(2) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 2 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 1 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(1) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 32 +, /* mEpilogueLdtmBits */ 32 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 64 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 1 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 256 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 256 +, /* mMmaN */ 256 +, /* mMmaTileK */ 256 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 256 +, /* mNumRegsPerThreadNonEpilogueWarp */ 104 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 5 +, /* mNumStagesB */ 5 +, /* mNumStagesMma */ 1 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 1 +, /* mNumStagesSmemSfA */ 5 +, /* mNumStagesSmemSfB */ 5 +, /* mNumStagesTmemSfA */ 5 +, /* mNumStagesTmemSfB */ 1 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(0) +, /* mSfLayoutB */ trtllm::gen::SfLayout(3) +, /* mSfLayoutC */ trtllm::gen::SfLayout(3) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 256 +, /* mTileM */ 128 +, /* mTileN */ 256 +, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTransposeMmaOutput */ 0 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 1 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 0 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 256 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(1) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(0) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 1 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 32 +, /* mNumRegsPerThreadLoadB */ 0 +, /* mNumRegsPerThreadLoadSfA */ 40 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 8 +, /* mNumWarpsLoadB */ 0 +, /* mNumWarpsLoadSfA */ 4 +, /* mNumWarpsLoadSfB */ 0 +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(3)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_siTuGlu_laW8_lsfaW4_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_siTuGlu_laW8_lsfaW4_dynB_sm100f_cubin_len, 203416, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_siTuGlu_laW8_lsfaW4_dynB_sm100f", 640, "e4a3b816236858604acfcc262fd52fe9318df3d3b09561fd05b6edd34d0571a6", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(2) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 2 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 1 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(1) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 32 +, /* mEpilogueLdtmBits */ 32 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 64 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 1 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 256 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 256 +, /* mMmaN */ 256 +, /* mMmaTileK */ 256 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 256 +, /* mNumRegsPerThreadNonEpilogueWarp */ 104 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 5 +, /* mNumStagesB */ 5 +, /* mNumStagesMma */ 1 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 1 +, /* mNumStagesSmemSfA */ 5 +, /* mNumStagesSmemSfB */ 5 +, /* mNumStagesTmemSfA */ 5 +, /* mNumStagesTmemSfB */ 1 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(0) +, /* mSfLayoutB */ trtllm::gen::SfLayout(3) +, /* mSfLayoutC */ trtllm::gen::SfLayout(3) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 256 +, /* mTileM */ 128 +, /* mTileN */ 256 +, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTransposeMmaOutput */ 0 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 1 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 0 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 256 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(2) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(0) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 1 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 32 +, /* mNumRegsPerThreadLoadB */ 0 +, /* mNumRegsPerThreadLoadSfA */ 40 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 8 +, /* mNumWarpsLoadB */ 0 +, /* mNumWarpsLoadSfA */ 4 +, /* mNumWarpsLoadSfB */ 0 +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(3)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_swiGlu_laW8_lsfaW4_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_swiGlu_laW8_lsfaW4_dynB_sm100f_cubin_len, 203416, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_swiGlu_laW8_lsfaW4_dynB_sm100f", 640, "2834fdfe94bb0299bcd345a8618b800be85054e857f2b04b671b06d61a608efa", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(2) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 2 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 1 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(1) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 32 +, /* mEpilogueLdtmBits */ 32 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 64 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 1 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 256 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 256 +, /* mMmaN */ 256 +, /* mMmaTileK */ 256 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 256 +, /* mNumRegsPerThreadNonEpilogueWarp */ 104 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 5 +, /* mNumStagesB */ 5 +, /* mNumStagesMma */ 1 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 1 +, /* mNumStagesSmemSfA */ 5 +, /* mNumStagesSmemSfB */ 5 +, /* mNumStagesTmemSfA */ 5 +, /* mNumStagesTmemSfB */ 1 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(0) +, /* mSfLayoutB */ trtllm::gen::SfLayout(3) +, /* mSfLayoutC */ trtllm::gen::SfLayout(3) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 256 +, /* mTileM */ 128 +, /* mTileN */ 256 +, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTransposeMmaOutput */ 0 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 1 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 0 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 256 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(0) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(0) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 1 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 32 +, /* mNumRegsPerThreadLoadB */ 0 +, /* mNumRegsPerThreadLoadSfA */ 40 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 8 +, /* mNumWarpsLoadB */ 0 +, /* mNumWarpsLoadSfA */ 4 +, /* mNumWarpsLoadSfB */ 0 +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(3)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_relu2_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_laW8_lsfaW4_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_relu2_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_laW8_lsfaW4_dynB_sm100f_cubin_len, 203416, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_relu2_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_laW8_lsfaW4_dynB_sm100f", 640, "3a8973aba519da24d9eb407f0e9a0e81e69341cc50e6b49deb8cdefadbfba96e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(2) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 2 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 1 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(1) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(2) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 32 +, /* mEpilogueLdtmBits */ 32 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 64 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 1 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 256 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 256 +, /* mMmaN */ 256 +, /* mMmaTileK */ 256 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 256 +, /* mNumRegsPerThreadNonEpilogueWarp */ 104 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 5 +, /* mNumStagesB */ 5 +, /* mNumStagesMma */ 1 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 1 +, /* mNumStagesSmemSfA */ 5 +, /* mNumStagesSmemSfB */ 5 +, /* mNumStagesTmemSfA */ 5 +, /* mNumStagesTmemSfB */ 1 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(0) +, /* mSfLayoutB */ trtllm::gen::SfLayout(3) +, /* mSfLayoutC */ trtllm::gen::SfLayout(3) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 256 +, /* mTileM */ 128 +, /* mTileN */ 256 +, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTransposeMmaOutput */ 0 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 1 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 0 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 256 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(3) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(0) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 0 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 32 +, /* mNumRegsPerThreadLoadB */ 0 +, /* mNumRegsPerThreadLoadSfA */ 40 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 8 +, /* mNumWarpsLoadB */ 0 +, /* mNumWarpsLoadSfA */ 4 +, /* mNumWarpsLoadSfB */ 0 +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(3)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_silu_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_laW8_lsfaW4_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_silu_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_laW8_lsfaW4_dynB_sm100f_cubin_len, 203416, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_silu_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_laW8_lsfaW4_dynB_sm100f", 640, "d48a965569b6669a54e8e9351b96dff71ec9b815074dfe4a3bdc444f2fd931f4", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(2) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 2 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 1 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(1) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(3) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 32 +, /* mEpilogueLdtmBits */ 32 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 64 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 1 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 256 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 256 +, /* mMmaN */ 256 +, /* mMmaTileK */ 256 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 256 +, /* mNumRegsPerThreadNonEpilogueWarp */ 104 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 5 +, /* mNumStagesB */ 5 +, /* mNumStagesMma */ 1 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 1 +, /* mNumStagesSmemSfA */ 5 +, /* mNumStagesSmemSfB */ 5 +, /* mNumStagesTmemSfA */ 5 +, /* mNumStagesTmemSfB */ 1 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(0) +, /* mSfLayoutB */ trtllm::gen::SfLayout(3) +, /* mSfLayoutC */ trtllm::gen::SfLayout(3) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 256 +, /* mTileM */ 128 +, /* mTileN */ 256 +, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTransposeMmaOutput */ 0 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 1 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 0 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 256 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(3) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(0) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 0 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 32 +, /* mNumRegsPerThreadLoadB */ 0 +, /* mNumRegsPerThreadLoadSfA */ 40 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 8 +, /* mNumWarpsLoadB */ 0 +, /* mNumWarpsLoadSfA */ 4 +, /* mNumWarpsLoadSfB */ 0 +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(3)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 215184, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f", 640, "b5a541e52e952caa06bf43abb0adad3559e61933294231492374bc7ef72a2d40", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(1) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 1 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 1 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 16 +, /* mEpilogueLdtmBits */ 256 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 32 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 0 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 256 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 128 +, /* mMmaN */ 32 +, /* mMmaTileK */ 256 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 9 +, /* mNumStagesB */ 9 +, /* mNumStagesMma */ 2 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 2 +, /* mNumStagesSmemSfA */ 9 +, /* mNumStagesSmemSfB */ 9 +, /* mNumStagesTmemSfA */ 9 +, /* mNumStagesTmemSfB */ 9 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(3) +, /* mSfLayoutB */ trtllm::gen::SfLayout(0) +, /* mSfLayoutC */ trtllm::gen::SfLayout(1) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 256 +, /* mTileM */ 128 +, /* mTileN */ 32 +, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTransposeMmaOutput */ 1 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 0 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 0 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 256 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(1) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 1 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 0 +, /* mNumRegsPerThreadLoadSfA */ 0 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 0 +, /* mNumWarpsLoadSfA */ 0 +, /* mNumWarpsLoadSfB */ 0 +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 215184, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 640, "fd8e5a2dba4140fb52aa683a8df9586e0f9f016b21ea366a93a84ca63a1d045d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(1) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 1 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 1 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 16 +, /* mEpilogueLdtmBits */ 256 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 32 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 0 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 256 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 128 +, /* mMmaN */ 32 +, /* mMmaTileK */ 256 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 9 +, /* mNumStagesB */ 9 +, /* mNumStagesMma */ 2 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 2 +, /* mNumStagesSmemSfA */ 9 +, /* mNumStagesSmemSfB */ 9 +, /* mNumStagesTmemSfA */ 9 +, /* mNumStagesTmemSfB */ 9 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(3) +, /* mSfLayoutB */ trtllm::gen::SfLayout(0) +, /* mSfLayoutC */ trtllm::gen::SfLayout(1) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 256 +, /* mTileM */ 128 +, /* mTileN */ 32 +, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTransposeMmaOutput */ 1 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 0 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 0 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 256 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(2) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 1 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 0 +, /* mNumRegsPerThreadLoadSfA */ 0 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 0 +, /* mNumWarpsLoadSfA */ 0 +, /* mNumWarpsLoadSfB */ 0 +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 215184, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "bc6b7f5dbde0343a4223b3746f3be28449026a44c642b8698168d12abdc11c8b", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(1) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 1 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 1 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 16 +, /* mEpilogueLdtmBits */ 256 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 32 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 0 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 256 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 128 +, /* mMmaN */ 32 +, /* mMmaTileK */ 256 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 9 +, /* mNumStagesB */ 9 +, /* mNumStagesMma */ 2 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 2 +, /* mNumStagesSmemSfA */ 9 +, /* mNumStagesSmemSfB */ 9 +, /* mNumStagesTmemSfA */ 9 +, /* mNumStagesTmemSfB */ 9 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(3) +, /* mSfLayoutB */ trtllm::gen::SfLayout(0) +, /* mSfLayoutC */ trtllm::gen::SfLayout(1) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 256 +, /* mTileM */ 128 +, /* mTileN */ 32 +, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTransposeMmaOutput */ 1 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 0 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 0 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 256 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(0) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 1 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 0 +, /* mNumRegsPerThreadLoadSfA */ 0 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 0 +, /* mNumWarpsLoadSfA */ 0 +, /* mNumWarpsLoadSfB */ 0 +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 215184, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 640, "36abe1403e1dc902b39d79b868e9e74b1fdce4890b951cedbb7550ba94887bb9", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(1) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 1 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 1 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(2) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 16 +, /* mEpilogueLdtmBits */ 256 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 32 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 0 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 256 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 128 +, /* mMmaN */ 32 +, /* mMmaTileK */ 256 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 9 +, /* mNumStagesB */ 9 +, /* mNumStagesMma */ 2 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 2 +, /* mNumStagesSmemSfA */ 9 +, /* mNumStagesSmemSfB */ 9 +, /* mNumStagesTmemSfA */ 9 +, /* mNumStagesTmemSfB */ 9 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(3) +, /* mSfLayoutB */ trtllm::gen::SfLayout(0) +, /* mSfLayoutC */ trtllm::gen::SfLayout(1) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 256 +, /* mTileM */ 128 +, /* mTileN */ 32 +, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTransposeMmaOutput */ 1 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 0 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 0 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 256 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(0) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 0 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 0 +, /* mNumRegsPerThreadLoadSfA */ 0 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 0 +, /* mNumWarpsLoadSfA */ 0 +, /* mNumWarpsLoadSfB */ 0 +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 215184, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 640, "5cd622397e2e5f90bea1b3ff1022c005adc1b3e0eac92c4240b3449121af3e07", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(1) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 1 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 1 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(3) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 16 +, /* mEpilogueLdtmBits */ 256 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 32 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 0 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 256 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 128 +, /* mMmaN */ 32 +, /* mMmaTileK */ 256 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 9 +, /* mNumStagesB */ 9 +, /* mNumStagesMma */ 2 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 2 +, /* mNumStagesSmemSfA */ 9 +, /* mNumStagesSmemSfB */ 9 +, /* mNumStagesTmemSfA */ 9 +, /* mNumStagesTmemSfB */ 9 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(3) +, /* mSfLayoutB */ trtllm::gen::SfLayout(0) +, /* mSfLayoutC */ trtllm::gen::SfLayout(1) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 256 +, /* mTileM */ 128 +, /* mTileN */ 32 +, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTransposeMmaOutput */ 1 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 0 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 0 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 256 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(3) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 0 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 0 +, /* mNumRegsPerThreadLoadSfA */ 0 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 0 +, /* mNumWarpsLoadSfA */ 0 +, /* mNumWarpsLoadSfB */ 0 +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 215056, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f", 640, "1cd69bd4386a41b36981358a17785b94b44fe184d0ad07bb6417e95263867a8d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(1) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 1 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 1 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 16 +, /* mEpilogueLdtmBits */ 256 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 32 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 0 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 256 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 128 +, /* mMmaN */ 32 +, /* mMmaTileK */ 256 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 9 +, /* mNumStagesB */ 9 +, /* mNumStagesMma */ 1 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 1 +, /* mNumStagesSmemSfA */ 9 +, /* mNumStagesSmemSfB */ 9 +, /* mNumStagesTmemSfA */ 9 +, /* mNumStagesTmemSfB */ 9 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(3) +, /* mSfLayoutB */ trtllm::gen::SfLayout(0) +, /* mSfLayoutC */ trtllm::gen::SfLayout(1) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 256 +, /* mTileM */ 128 +, /* mTileN */ 32 +, /* mTileScheduler */ gemm::TileScheduler(0) +, /* mTransposeMmaOutput */ 1 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 0 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 0 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 256 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(1) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 1 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 0 +, /* mNumRegsPerThreadLoadSfA */ 0 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 0 +, /* mNumWarpsLoadSfA */ 0 +, /* mNumWarpsLoadSfB */ 0 +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 215056, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 640, "1a1b4cf4a34b20dfd0fd33d5bc93beb41afc944a9dd660945384cfa10986091f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(1) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 1 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 1 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 16 +, /* mEpilogueLdtmBits */ 256 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 32 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 0 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 256 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 128 +, /* mMmaN */ 32 +, /* mMmaTileK */ 256 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 9 +, /* mNumStagesB */ 9 +, /* mNumStagesMma */ 1 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 1 +, /* mNumStagesSmemSfA */ 9 +, /* mNumStagesSmemSfB */ 9 +, /* mNumStagesTmemSfA */ 9 +, /* mNumStagesTmemSfB */ 9 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(3) +, /* mSfLayoutB */ trtllm::gen::SfLayout(0) +, /* mSfLayoutC */ trtllm::gen::SfLayout(1) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 256 +, /* mTileM */ 128 +, /* mTileN */ 32 +, /* mTileScheduler */ gemm::TileScheduler(0) +, /* mTransposeMmaOutput */ 1 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 0 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 0 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 256 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(2) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 1 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 0 +, /* mNumRegsPerThreadLoadSfA */ 0 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 0 +, /* mNumWarpsLoadSfA */ 0 +, /* mNumWarpsLoadSfB */ 0 +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 215056, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "df3560a258e84322007362086336e26aec8d4e1660b5eeafd25e95cc4f5db0b4", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(1) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 1 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 1 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 16 +, /* mEpilogueLdtmBits */ 256 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 32 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 0 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 256 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 128 +, /* mMmaN */ 32 +, /* mMmaTileK */ 256 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 9 +, /* mNumStagesB */ 9 +, /* mNumStagesMma */ 1 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 1 +, /* mNumStagesSmemSfA */ 9 +, /* mNumStagesSmemSfB */ 9 +, /* mNumStagesTmemSfA */ 9 +, /* mNumStagesTmemSfB */ 9 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(3) +, /* mSfLayoutB */ trtllm::gen::SfLayout(0) +, /* mSfLayoutC */ trtllm::gen::SfLayout(1) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 256 +, /* mTileM */ 128 +, /* mTileN */ 32 +, /* mTileScheduler */ gemm::TileScheduler(0) +, /* mTransposeMmaOutput */ 1 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 0 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 0 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 256 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(0) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 1 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 0 +, /* mNumRegsPerThreadLoadSfA */ 0 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 0 +, /* mNumWarpsLoadSfA */ 0 +, /* mNumWarpsLoadSfB */ 0 +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 215056, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 640, "59e9a58dc2cde22da3fe387b46d3d6a2d37e3d9094823b29b47429d049f804aa", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(1) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 1 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 1 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(2) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 16 +, /* mEpilogueLdtmBits */ 256 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 32 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 0 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 256 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 128 +, /* mMmaN */ 32 +, /* mMmaTileK */ 256 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 9 +, /* mNumStagesB */ 9 +, /* mNumStagesMma */ 1 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 1 +, /* mNumStagesSmemSfA */ 9 +, /* mNumStagesSmemSfB */ 9 +, /* mNumStagesTmemSfA */ 9 +, /* mNumStagesTmemSfB */ 9 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(3) +, /* mSfLayoutB */ trtllm::gen::SfLayout(0) +, /* mSfLayoutC */ trtllm::gen::SfLayout(1) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 256 +, /* mTileM */ 128 +, /* mTileN */ 32 +, /* mTileScheduler */ gemm::TileScheduler(0) +, /* mTransposeMmaOutput */ 1 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 0 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 0 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 256 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(0) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 0 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 0 +, /* mNumRegsPerThreadLoadSfA */ 0 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 0 +, /* mNumWarpsLoadSfA */ 0 +, /* mNumWarpsLoadSfB */ 0 +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 215056, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 640, "21193e2dd64366ef5ef083d72006b000204613ae8b3b7378e62dfcfcee53fadc", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(1) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 1 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 1 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(3) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 16 +, /* mEpilogueLdtmBits */ 256 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 32 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 0 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 256 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 128 +, /* mMmaN */ 32 +, /* mMmaTileK */ 256 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 9 +, /* mNumStagesB */ 9 +, /* mNumStagesMma */ 1 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 1 +, /* mNumStagesSmemSfA */ 9 +, /* mNumStagesSmemSfB */ 9 +, /* mNumStagesTmemSfA */ 9 +, /* mNumStagesTmemSfB */ 9 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(3) +, /* mSfLayoutB */ trtllm::gen::SfLayout(0) +, /* mSfLayoutC */ trtllm::gen::SfLayout(1) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 256 +, /* mTileM */ 128 +, /* mTileN */ 32 +, /* mTileScheduler */ gemm::TileScheduler(0) +, /* mTransposeMmaOutput */ 1 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 0 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 0 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 256 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(3) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 0 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 0 +, /* mNumRegsPerThreadLoadSfA */ 0 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 0 +, /* mNumWarpsLoadSfA */ 0 +, /* mNumWarpsLoadSfB */ 0 +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 215184, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f", 640, "575ecb72818ab019adfae6a6aa07bd9da9268d3ecbaceb7f7257362ec109d278", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(1) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 1 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 1 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 16 +, /* mEpilogueLdtmBits */ 256 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 32 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 0 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 512 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 128 +, /* mMmaN */ 32 +, /* mMmaTileK */ 256 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 9 +, /* mNumStagesB */ 9 +, /* mNumStagesMma */ 2 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 2 +, /* mNumStagesSmemSfA */ 9 +, /* mNumStagesSmemSfB */ 9 +, /* mNumStagesTmemSfA */ 9 +, /* mNumStagesTmemSfB */ 9 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(3) +, /* mSfLayoutB */ trtllm::gen::SfLayout(0) +, /* mSfLayoutC */ trtllm::gen::SfLayout(1) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 256 +, /* mTileM */ 128 +, /* mTileN */ 32 +, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTransposeMmaOutput */ 1 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 0 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 1 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 512 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(1) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 1 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 0 +, /* mNumRegsPerThreadLoadSfA */ 0 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 0 +, /* mNumWarpsLoadSfA */ 0 +, /* mNumWarpsLoadSfB */ 0 +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 215184, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 640, "003da4694e188c7e3f769596e2a3ed307b9237ed7530cb9a98b2c90ed0b66b40", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(1) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 1 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 1 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 16 +, /* mEpilogueLdtmBits */ 256 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 32 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 0 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 512 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 128 +, /* mMmaN */ 32 +, /* mMmaTileK */ 256 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 9 +, /* mNumStagesB */ 9 +, /* mNumStagesMma */ 2 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 2 +, /* mNumStagesSmemSfA */ 9 +, /* mNumStagesSmemSfB */ 9 +, /* mNumStagesTmemSfA */ 9 +, /* mNumStagesTmemSfB */ 9 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(3) +, /* mSfLayoutB */ trtllm::gen::SfLayout(0) +, /* mSfLayoutC */ trtllm::gen::SfLayout(1) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 256 +, /* mTileM */ 128 +, /* mTileN */ 32 +, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTransposeMmaOutput */ 1 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 0 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 1 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 512 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(2) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 1 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 0 +, /* mNumRegsPerThreadLoadSfA */ 0 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 0 +, /* mNumWarpsLoadSfA */ 0 +, /* mNumWarpsLoadSfB */ 0 +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 215184, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "73c57ef627a8ab379908ada24284d419bf00dce7289ee7db5033361145d9e06f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(1) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 1 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 1 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 16 +, /* mEpilogueLdtmBits */ 256 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 32 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 0 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 512 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 128 +, /* mMmaN */ 32 +, /* mMmaTileK */ 256 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 9 +, /* mNumStagesB */ 9 +, /* mNumStagesMma */ 2 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 2 +, /* mNumStagesSmemSfA */ 9 +, /* mNumStagesSmemSfB */ 9 +, /* mNumStagesTmemSfA */ 9 +, /* mNumStagesTmemSfB */ 9 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(3) +, /* mSfLayoutB */ trtllm::gen::SfLayout(0) +, /* mSfLayoutC */ trtllm::gen::SfLayout(1) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 256 +, /* mTileM */ 128 +, /* mTileN */ 32 +, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTransposeMmaOutput */ 1 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 0 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 1 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 512 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(0) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 1 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 0 +, /* mNumRegsPerThreadLoadSfA */ 0 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 0 +, /* mNumWarpsLoadSfA */ 0 +, /* mNumWarpsLoadSfB */ 0 +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 215184, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 640, "73c952a871d8c4e9fcf415450949d479530adfe71ab3789ae321be76eb2456df", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(1) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 1 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 1 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(2) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 16 +, /* mEpilogueLdtmBits */ 256 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 32 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 0 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 512 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 128 +, /* mMmaN */ 32 +, /* mMmaTileK */ 256 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 9 +, /* mNumStagesB */ 9 +, /* mNumStagesMma */ 2 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 2 +, /* mNumStagesSmemSfA */ 9 +, /* mNumStagesSmemSfB */ 9 +, /* mNumStagesTmemSfA */ 9 +, /* mNumStagesTmemSfB */ 9 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(3) +, /* mSfLayoutB */ trtllm::gen::SfLayout(0) +, /* mSfLayoutC */ trtllm::gen::SfLayout(1) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 256 +, /* mTileM */ 128 +, /* mTileN */ 32 +, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTransposeMmaOutput */ 1 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 0 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 1 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 512 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(0) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 0 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 0 +, /* mNumRegsPerThreadLoadSfA */ 0 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 0 +, /* mNumWarpsLoadSfA */ 0 +, /* mNumWarpsLoadSfB */ 0 +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 215184, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 640, "e91b932ccdd1099353f0d019b5d6e46580e2f06c57f6c4b1e88fb8a0fd8f9d04", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(1) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 1 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 1 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(3) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 16 +, /* mEpilogueLdtmBits */ 256 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 32 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 0 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 512 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 128 +, /* mMmaN */ 32 +, /* mMmaTileK */ 256 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 9 +, /* mNumStagesB */ 9 +, /* mNumStagesMma */ 2 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 2 +, /* mNumStagesSmemSfA */ 9 +, /* mNumStagesSmemSfB */ 9 +, /* mNumStagesTmemSfA */ 9 +, /* mNumStagesTmemSfB */ 9 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(3) +, /* mSfLayoutB */ trtllm::gen::SfLayout(0) +, /* mSfLayoutC */ trtllm::gen::SfLayout(1) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 256 +, /* mTileM */ 128 +, /* mTileN */ 32 +, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTransposeMmaOutput */ 1 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 0 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 1 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 512 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(3) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 0 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 0 +, /* mNumRegsPerThreadLoadSfA */ 0 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 0 +, /* mNumWarpsLoadSfA */ 0 +, /* mNumWarpsLoadSfB */ 0 +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 215056, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f", 640, "e1d416090e54a181f1937c931fd2d184fbfa564546448612e6457eb9becd0feb", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(1) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 1 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 1 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 16 +, /* mEpilogueLdtmBits */ 256 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 32 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 0 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 512 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 128 +, /* mMmaN */ 32 +, /* mMmaTileK */ 256 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 9 +, /* mNumStagesB */ 9 +, /* mNumStagesMma */ 1 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 1 +, /* mNumStagesSmemSfA */ 9 +, /* mNumStagesSmemSfB */ 9 +, /* mNumStagesTmemSfA */ 9 +, /* mNumStagesTmemSfB */ 9 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(3) +, /* mSfLayoutB */ trtllm::gen::SfLayout(0) +, /* mSfLayoutC */ trtllm::gen::SfLayout(1) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 256 +, /* mTileM */ 128 +, /* mTileN */ 32 +, /* mTileScheduler */ gemm::TileScheduler(0) +, /* mTransposeMmaOutput */ 1 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 0 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 1 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 512 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(1) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 1 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 0 +, /* mNumRegsPerThreadLoadSfA */ 0 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 0 +, /* mNumWarpsLoadSfA */ 0 +, /* mNumWarpsLoadSfB */ 0 +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 215056, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 640, "59fa3c52d56d2b2126d886d141ceed692b49f7bb6013114bf2575135649ba131", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(1) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 1 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 1 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 16 +, /* mEpilogueLdtmBits */ 256 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 32 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 0 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 512 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 128 +, /* mMmaN */ 32 +, /* mMmaTileK */ 256 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 9 +, /* mNumStagesB */ 9 +, /* mNumStagesMma */ 1 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 1 +, /* mNumStagesSmemSfA */ 9 +, /* mNumStagesSmemSfB */ 9 +, /* mNumStagesTmemSfA */ 9 +, /* mNumStagesTmemSfB */ 9 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(3) +, /* mSfLayoutB */ trtllm::gen::SfLayout(0) +, /* mSfLayoutC */ trtllm::gen::SfLayout(1) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 256 +, /* mTileM */ 128 +, /* mTileN */ 32 +, /* mTileScheduler */ gemm::TileScheduler(0) +, /* mTransposeMmaOutput */ 1 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 0 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 1 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 512 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(2) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 1 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 0 +, /* mNumRegsPerThreadLoadSfA */ 0 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 0 +, /* mNumWarpsLoadSfA */ 0 +, /* mNumWarpsLoadSfB */ 0 +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 215056, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "fd2262b0b25f20f697f53e05de7fc8952cccd86d778c71fb6578c83099d8d567", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(1) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 1 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 1 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 16 +, /* mEpilogueLdtmBits */ 256 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 32 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 0 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 512 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 128 +, /* mMmaN */ 32 +, /* mMmaTileK */ 256 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 9 +, /* mNumStagesB */ 9 +, /* mNumStagesMma */ 1 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 1 +, /* mNumStagesSmemSfA */ 9 +, /* mNumStagesSmemSfB */ 9 +, /* mNumStagesTmemSfA */ 9 +, /* mNumStagesTmemSfB */ 9 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(3) +, /* mSfLayoutB */ trtllm::gen::SfLayout(0) +, /* mSfLayoutC */ trtllm::gen::SfLayout(1) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 256 +, /* mTileM */ 128 +, /* mTileN */ 32 +, /* mTileScheduler */ gemm::TileScheduler(0) +, /* mTransposeMmaOutput */ 1 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 0 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 1 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 512 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(0) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 1 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 0 +, /* mNumRegsPerThreadLoadSfA */ 0 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 0 +, /* mNumWarpsLoadSfA */ 0 +, /* mNumWarpsLoadSfB */ 0 +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 215056, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 640, "fb9b41906e4af29e923e8d5f44f46f8f5f37f8155cfc2be9b7f2cdc23873f048", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(1) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 1 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 1 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(2) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 16 +, /* mEpilogueLdtmBits */ 256 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 32 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 0 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 512 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 128 +, /* mMmaN */ 32 +, /* mMmaTileK */ 256 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 9 +, /* mNumStagesB */ 9 +, /* mNumStagesMma */ 1 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 1 +, /* mNumStagesSmemSfA */ 9 +, /* mNumStagesSmemSfB */ 9 +, /* mNumStagesTmemSfA */ 9 +, /* mNumStagesTmemSfB */ 9 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(3) +, /* mSfLayoutB */ trtllm::gen::SfLayout(0) +, /* mSfLayoutC */ trtllm::gen::SfLayout(1) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 256 +, /* mTileM */ 128 +, /* mTileN */ 32 +, /* mTileScheduler */ gemm::TileScheduler(0) +, /* mTransposeMmaOutput */ 1 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 0 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 1 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 512 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(0) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 0 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 0 +, /* mNumRegsPerThreadLoadSfA */ 0 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 0 +, /* mNumWarpsLoadSfA */ 0 +, /* mNumWarpsLoadSfB */ 0 +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 215056, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 640, "fd1921e06d7998435224e6f1e4084ed44e115a34047ef7f429b317a88743891c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(1) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 1 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 1 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(3) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 16 +, /* mEpilogueLdtmBits */ 256 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 32 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 0 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 512 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 128 +, /* mMmaN */ 32 +, /* mMmaTileK */ 256 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 9 +, /* mNumStagesB */ 9 +, /* mNumStagesMma */ 1 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 1 +, /* mNumStagesSmemSfA */ 9 +, /* mNumStagesSmemSfB */ 9 +, /* mNumStagesTmemSfA */ 9 +, /* mNumStagesTmemSfB */ 9 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(3) +, /* mSfLayoutB */ trtllm::gen::SfLayout(0) +, /* mSfLayoutC */ trtllm::gen::SfLayout(1) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 256 +, /* mTileM */ 128 +, /* mTileN */ 32 +, /* mTileScheduler */ gemm::TileScheduler(0) +, /* mTransposeMmaOutput */ 1 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 0 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 1 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 512 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(3) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 0 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 0 +, /* mNumRegsPerThreadLoadSfA */ 0 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 0 +, /* mNumWarpsLoadSfA */ 0 +, /* mNumWarpsLoadSfB */ 0 +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 175768, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f", 896, "4bca825f6b8b189ddeed4d9c2d0a0dfed6b69b1d61b4e7fafa5e8e62b57c6aeb", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(1) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 1 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 2 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 16 +, /* mEpilogueLdtmBits */ 256 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 32 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 0 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 1024 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 128 +, /* mMmaN */ 32 +, /* mMmaTileK */ 512 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumSlicesForSplitK */ 2 +, /* mNumStagesA */ 3 +, /* mNumStagesB */ 3 +, /* mNumStagesMma */ 2 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 2 +, /* mNumStagesSmemSfA */ 3 +, /* mNumStagesSmemSfB */ 3 +, /* mNumStagesTmemSfA */ 3 +, /* mNumStagesTmemSfB */ 3 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(3) +, /* mSfLayoutB */ trtllm::gen::SfLayout(0) +, /* mSfLayoutC */ trtllm::gen::SfLayout(1) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(2) +, /* mTileK */ 512 +, /* mTileM */ 128 +, /* mTileN */ 32 +, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTransposeMmaOutput */ 1 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 0 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 0 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 1024 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(1) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 1 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 0 +, /* mNumRegsPerThreadLoadSfA */ 0 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 0 +, /* mNumWarpsLoadSfA */ 0 +, /* mNumWarpsLoadSfB */ 0 +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 175768, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f", 896, "47caddf1cac2d4f76600cdd2303caf861f50d378d7f7c873458abe4e40f2faf7", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(1) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 1 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 2 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 16 +, /* mEpilogueLdtmBits */ 256 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 32 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 0 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 1024 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 128 +, /* mMmaN */ 32 +, /* mMmaTileK */ 512 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumSlicesForSplitK */ 2 +, /* mNumStagesA */ 3 +, /* mNumStagesB */ 3 +, /* mNumStagesMma */ 2 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 2 +, /* mNumStagesSmemSfA */ 3 +, /* mNumStagesSmemSfB */ 3 +, /* mNumStagesTmemSfA */ 3 +, /* mNumStagesTmemSfB */ 3 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(3) +, /* mSfLayoutB */ trtllm::gen::SfLayout(0) +, /* mSfLayoutC */ trtllm::gen::SfLayout(1) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(2) +, /* mTileK */ 512 +, /* mTileM */ 128 +, /* mTileN */ 32 +, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTransposeMmaOutput */ 1 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 0 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 0 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 1024 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(2) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 1 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 0 +, /* mNumRegsPerThreadLoadSfA */ 0 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 0 +, /* mNumWarpsLoadSfA */ 0 +, /* mNumWarpsLoadSfB */ 0 +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 175768, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 896, "23d9ac27cbcec7626d8c8d33ecdd521ba3fdb035084dee8c4a18d0c23a685d63", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(1) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 1 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 2 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(2) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 16 +, /* mEpilogueLdtmBits */ 256 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 32 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 0 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 1024 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 128 +, /* mMmaN */ 32 +, /* mMmaTileK */ 512 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumSlicesForSplitK */ 2 +, /* mNumStagesA */ 3 +, /* mNumStagesB */ 3 +, /* mNumStagesMma */ 2 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 2 +, /* mNumStagesSmemSfA */ 3 +, /* mNumStagesSmemSfB */ 3 +, /* mNumStagesTmemSfA */ 3 +, /* mNumStagesTmemSfB */ 3 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(3) +, /* mSfLayoutB */ trtllm::gen::SfLayout(0) +, /* mSfLayoutC */ trtllm::gen::SfLayout(1) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(2) +, /* mTileK */ 512 +, /* mTileM */ 128 +, /* mTileN */ 32 +, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTransposeMmaOutput */ 1 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 0 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 0 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 1024 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(3) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 0 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 0 +, /* mNumRegsPerThreadLoadSfA */ 0 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 0 +, /* mNumWarpsLoadSfA */ 0 +, /* mNumWarpsLoadSfB */ 0 +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 175768, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 896, "9a28cc57b2d7786f26abb5435b6eb1621d0eeed9b43ee33de1fe9a7c25025687", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(1) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 1 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 2 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(3) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 16 +, /* mEpilogueLdtmBits */ 256 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 32 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 0 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 1024 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 128 +, /* mMmaN */ 32 +, /* mMmaTileK */ 512 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumSlicesForSplitK */ 2 +, /* mNumStagesA */ 3 +, /* mNumStagesB */ 3 +, /* mNumStagesMma */ 2 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 2 +, /* mNumStagesSmemSfA */ 3 +, /* mNumStagesSmemSfB */ 3 +, /* mNumStagesTmemSfA */ 3 +, /* mNumStagesTmemSfB */ 3 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(3) +, /* mSfLayoutB */ trtllm::gen::SfLayout(0) +, /* mSfLayoutC */ trtllm::gen::SfLayout(1) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(2) +, /* mTileK */ 512 +, /* mTileM */ 128 +, /* mTileN */ 32 +, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTransposeMmaOutput */ 1 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 0 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 0 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 1024 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(3) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 0 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 0 +, /* mNumRegsPerThreadLoadSfA */ 0 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 0 +, /* mNumWarpsLoadSfA */ 0 +, /* mNumWarpsLoadSfB */ 0 +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 216984, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f", 896, "6530171e2c121f1127e54585fe8e1656539667aa1b848d1be9421010ac29ef5f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(1) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 2 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 1 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 16 +, /* mEpilogueLdtmBits */ 256 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 32 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 0 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 512 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 256 +, /* mMmaN */ 32 +, /* mMmaTileK */ 512 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 5 +, /* mNumStagesB */ 5 +, /* mNumStagesMma */ 2 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 2 +, /* mNumStagesSmemSfA */ 5 +, /* mNumStagesSmemSfB */ 5 +, /* mNumStagesTmemSfA */ 5 +, /* mNumStagesTmemSfB */ 5 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(3) +, /* mSfLayoutB */ trtllm::gen::SfLayout(0) +, /* mSfLayoutC */ trtllm::gen::SfLayout(1) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 512 +, /* mTileM */ 128 +, /* mTileN */ 32 +, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTransposeMmaOutput */ 1 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 0 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 0 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 512 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(1) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 1 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 0 +, /* mNumRegsPerThreadLoadSfA */ 0 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 0 +, /* mNumWarpsLoadSfA */ 0 +, /* mNumWarpsLoadSfB */ 0 +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 216984, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f", 896, "660041962fd2353a91328a01818648cd496d8c650c2575c8708478caf4c8f1a6", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(1) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 2 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 1 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 16 +, /* mEpilogueLdtmBits */ 256 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 32 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 0 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 512 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 256 +, /* mMmaN */ 32 +, /* mMmaTileK */ 512 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 5 +, /* mNumStagesB */ 5 +, /* mNumStagesMma */ 2 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 2 +, /* mNumStagesSmemSfA */ 5 +, /* mNumStagesSmemSfB */ 5 +, /* mNumStagesTmemSfA */ 5 +, /* mNumStagesTmemSfB */ 5 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(3) +, /* mSfLayoutB */ trtllm::gen::SfLayout(0) +, /* mSfLayoutC */ trtllm::gen::SfLayout(1) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 512 +, /* mTileM */ 128 +, /* mTileN */ 32 +, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTransposeMmaOutput */ 1 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 0 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 0 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 512 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(2) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 1 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 0 +, /* mNumRegsPerThreadLoadSfA */ 0 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 0 +, /* mNumWarpsLoadSfA */ 0 +, /* mNumWarpsLoadSfB */ 0 +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 216984, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 896, "54c7e013453d8eef4bd20e4f4edf15a5a41bc5ea90fcf417fff9af44999ba318", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(1) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 2 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 1 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 16 +, /* mEpilogueLdtmBits */ 256 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 32 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 0 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 512 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 256 +, /* mMmaN */ 32 +, /* mMmaTileK */ 512 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 5 +, /* mNumStagesB */ 5 +, /* mNumStagesMma */ 2 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 2 +, /* mNumStagesSmemSfA */ 5 +, /* mNumStagesSmemSfB */ 5 +, /* mNumStagesTmemSfA */ 5 +, /* mNumStagesTmemSfB */ 5 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(3) +, /* mSfLayoutB */ trtllm::gen::SfLayout(0) +, /* mSfLayoutC */ trtllm::gen::SfLayout(1) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 512 +, /* mTileM */ 128 +, /* mTileN */ 32 +, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTransposeMmaOutput */ 1 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 0 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 0 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 512 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(0) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 1 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 0 +, /* mNumRegsPerThreadLoadSfA */ 0 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 0 +, /* mNumWarpsLoadSfA */ 0 +, /* mNumWarpsLoadSfB */ 0 +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 216984, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 896, "309e7cd370a0b2b2737b6436410e868ef3c4fdc760c59faa1a9e317ef47c85c7", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(1) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 2 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 1 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(2) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 16 +, /* mEpilogueLdtmBits */ 256 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 32 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 0 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 512 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 256 +, /* mMmaN */ 32 +, /* mMmaTileK */ 512 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 5 +, /* mNumStagesB */ 5 +, /* mNumStagesMma */ 2 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 2 +, /* mNumStagesSmemSfA */ 5 +, /* mNumStagesSmemSfB */ 5 +, /* mNumStagesTmemSfA */ 5 +, /* mNumStagesTmemSfB */ 5 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(3) +, /* mSfLayoutB */ trtllm::gen::SfLayout(0) +, /* mSfLayoutC */ trtllm::gen::SfLayout(1) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 512 +, /* mTileM */ 128 +, /* mTileN */ 32 +, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTransposeMmaOutput */ 1 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 0 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 0 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 512 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(0) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 0 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 0 +, /* mNumRegsPerThreadLoadSfA */ 0 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 0 +, /* mNumWarpsLoadSfA */ 0 +, /* mNumWarpsLoadSfB */ 0 +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 216984, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 896, "267c71d56ae764f52c9bd598c9bc3835b43229c301caa1f9369a67276ab84e10", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(1) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 2 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 1 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(3) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 16 +, /* mEpilogueLdtmBits */ 256 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 32 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 0 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 512 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 256 +, /* mMmaN */ 32 +, /* mMmaTileK */ 512 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 5 +, /* mNumStagesB */ 5 +, /* mNumStagesMma */ 2 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 2 +, /* mNumStagesSmemSfA */ 5 +, /* mNumStagesSmemSfB */ 5 +, /* mNumStagesTmemSfA */ 5 +, /* mNumStagesTmemSfB */ 5 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(3) +, /* mSfLayoutB */ trtllm::gen::SfLayout(0) +, /* mSfLayoutC */ trtllm::gen::SfLayout(1) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 512 +, /* mTileM */ 128 +, /* mTileN */ 32 +, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTransposeMmaOutput */ 1 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 0 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 0 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 512 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(3) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 0 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 0 +, /* mNumRegsPerThreadLoadSfA */ 0 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 0 +, /* mNumWarpsLoadSfA */ 0 +, /* mNumWarpsLoadSfB */ 0 +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 216600, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f", 768, "6bfb750d67a6850337d0873751c6b782d295c01ffe4536a1c9b29560a69e314b", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(1) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 2 , /* mClusterDimY */ 1 , /* mClusterDimZ */ 1 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) @@ -39413,16 +45296,16 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 256 +, /* mK */ 512 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) , /* mM */ 256 , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) -, /* mMmaM */ 128 +, /* mMmaM */ 256 , /* mMmaN */ 32 -, /* mMmaTileK */ 256 +, /* mMmaTileK */ 512 , /* mMockAllReduce */ 0 , /* mN */ 256 , /* mNumEpilogueWarps */ 4 @@ -39432,15 +45315,15 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 9 -, /* mNumStagesB */ 9 -, /* mNumStagesMma */ 2 +, /* mNumStagesA */ 5 +, /* mNumStagesB */ 5 +, /* mNumStagesMma */ 1 , /* mNumStagesMmaWithinWorkTile */ 1 -, /* mNumStagesMmaAcrossWorkTile */ 2 -, /* mNumStagesSmemSfA */ 9 -, /* mNumStagesSmemSfB */ 9 -, /* mNumStagesTmemSfA */ 9 -, /* mNumStagesTmemSfB */ 9 +, /* mNumStagesMmaAcrossWorkTile */ 1 +, /* mNumStagesSmemSfA */ 5 +, /* mNumStagesSmemSfB */ 5 +, /* mNumStagesTmemSfA */ 5 +, /* mNumStagesTmemSfB */ 5 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -39455,10 +45338,10 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) , /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 256 +, /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 32 -, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTileScheduler */ gemm::TileScheduler(0) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 , /* mUseCustomMmaSchedule */ 1 @@ -39475,9 +45358,9 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 256 +, /* mValidK */ 512 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(0) +, /* mActType */ gemmGatedAct::ActType(1) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} @@ -39498,143 +45381,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 , /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} -, /* mPrefetchB */ 0 -, /* mUseCMultiCast */ 0 -, /* mUseTmaOobOpt */ 1 - }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 215184, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 640, "29a4121f549f323fee4a2a3bf79f54f8aa72d2a3000368d7b714f38882eee224", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) -, /* mBiasDtype */ trtllm::gen::Dtype(1056776) -, /* mBiasType */ gemm::BiasType(1) -, /* mBlockK */ -1 -, /* mClcFastDrain */ 1 -, /* mClusterDimX */ 1 -, /* mClusterDimY */ 1 -, /* mClusterDimZ */ 1 -, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) -, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) -, /* mDtypeA */ trtllm::gen::Dtype(17826818) -, /* mDtypeB */ trtllm::gen::Dtype(17826818) -, /* mDtypeC */ trtllm::gen::Dtype(17826818) -, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) -, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) -, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(2) -, /* mEnablesEarlyExit */ 1 -, /* mEnablesDelayedEarlyExit */ 0 -, /* mEpilogueLdtmDps */ 16 -, /* mEpilogueLdtmBits */ 256 -, /* mEpilogueTileM */ 128 -, /* mEpilogueTileN */ 32 -, /* mFallbackClusterDimX */ 1 -, /* mFallbackClusterDimY */ 1 -, /* mFallbackClusterDimZ */ 1 -, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) -, /* mFuseLoadSfTask */ 0 -, /* mFuseUtccpWithUtcmma */ 0 -, /* mGridTriggerSecondaryA */ 0 -, /* mGridTriggerSecondaryB */ 1 -, /* mGridWaitForPrimaryEarlyExit */ 1 -, /* mGridWaitForPrimaryA */ 0 -, /* mGridWaitForPrimaryB */ 1 -, /* mHoistLoadTaskInit */ 1 -, /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 256 -, /* mKernelTraits */ {} -, /* mLayoutA */ gemm::MatrixLayout(0) -, /* mLayoutB */ gemm::MatrixLayout(0) -, /* mM */ 256 -, /* mMmaK */ 64 -, /* mMmaKind */ trtllm::gen::MmaKind(4) -, /* mMmaM */ 128 -, /* mMmaN */ 32 -, /* mMmaTileK */ 256 -, /* mMockAllReduce */ 0 -, /* mN */ 256 -, /* mNumEpilogueWarps */ 4 -, /* mNumRegsCastAWarps */ 0 -, /* mNumRegsCopySfLdsSttm */ 0 -, /* mNumRegsCopySparsityInfo */ 0 -, /* mNumRegsPerThreadEpilogueWarp */ 0 -, /* mNumRegsPerThreadNonEpilogueWarp */ 0 -, /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 9 -, /* mNumStagesB */ 9 -, /* mNumStagesMma */ 2 -, /* mNumStagesMmaWithinWorkTile */ 1 -, /* mNumStagesMmaAcrossWorkTile */ 2 -, /* mNumStagesSmemSfA */ 9 -, /* mNumStagesSmemSfB */ 9 -, /* mNumStagesTmemSfA */ 9 -, /* mNumStagesTmemSfB */ 9 -, /* mNumStagesWorkId */ 3 -, /* mOutputDebugTensors */ 0 -, /* mPatchF2fp */ 0 -, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) -, /* mSchedHostTask */ gemm::SchedHostTask(0) -, /* mSfBlockSizeA */ 16 -, /* mSfBlockSizeB */ 16 -, /* mSfBlockSizeC */ 16 -, /* mSfLayoutA */ trtllm::gen::SfLayout(3) -, /* mSfLayoutB */ trtllm::gen::SfLayout(0) -, /* mSfLayoutC */ trtllm::gen::SfLayout(1) -, /* mSfReshapeFactor */ 1 -, /* mSparsityA */ trtllm::gen::Sparsity(0) -, /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 256 -, /* mTileM */ 128 -, /* mTileN */ 32 -, /* mTileScheduler */ gemm::TileScheduler(1) -, /* mTransposeMmaOutput */ 1 -, /* mUseCustomizedMma3xNvFp4 */ 0 -, /* mUseCustomMmaSchedule */ 1 -, /* mUseDeepSeekFp8 */ 0 -, /* mUseFlexibleClusterDims */ 0 -, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 -, /* mUseMaxTmemOverlap */ 0 -, /* mUsePerTokenSfA */ 0 -, /* mUsePerTokenSfB */ 0 -, /* mUseShuffledMatrix */ 1 -, /* mUseTmaStore */ 1 -, /* mUseTwoTmaLoadWarps */ 1 -, /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 0 -, /* mValidM */ 256 -, /* mValidN */ 256 -, /* mValidK */ 256 -, /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(0) -, /* mClampBeforeAct */ 1 -, /* mBatchedM */ {} -, /* mBatchedN */ {} -, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) -, /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 0 -, /* mGridWaitForPrimaryRouting */ 1 -, /* mIsStaticBatch */ 0 -, /* mIsUniformNumTokensPerBatch */ 0 -, /* mNumBatches */ 128 -, /* mNumRegsPerThreadLoadA */ 0 -, /* mNumRegsPerThreadLoadB */ 0 -, /* mNumRegsPerThreadLoadSfA */ 0 -, /* mNumRegsPerThreadLoadSfB */ 0 -, /* mNumTokens */ 2 -, /* mNumWarpsLoadA */ 0 -, /* mNumWarpsLoadB */ 0 -, /* mNumWarpsLoadSfA */ 0 -, /* mNumWarpsLoadSfB */ 0 -, /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 215184, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 640, "b443a38efe304820dadb35a3a5523a068487ee6d02fd5cd7cf563747b220b645", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 216600, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f", 768, "ef17f89bb87a6f9304b77ee8569f19db6fc2aa79f43bee1b55810d1328f4c88b", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 -, /* mClusterDimX */ 1 +, /* mClusterDimX */ 2 , /* mClusterDimY */ 1 , /* mClusterDimZ */ 1 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) @@ -39645,7 +45402,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(3) +, /* mEltwiseActType */ gemm::EltwiseActType(0) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 @@ -39665,16 +45422,16 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 256 +, /* mK */ 512 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) , /* mM */ 256 , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) -, /* mMmaM */ 128 +, /* mMmaM */ 256 , /* mMmaN */ 32 -, /* mMmaTileK */ 256 +, /* mMmaTileK */ 512 , /* mMockAllReduce */ 0 , /* mN */ 256 , /* mNumEpilogueWarps */ 4 @@ -39684,15 +45441,15 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 9 -, /* mNumStagesB */ 9 -, /* mNumStagesMma */ 2 +, /* mNumStagesA */ 5 +, /* mNumStagesB */ 5 +, /* mNumStagesMma */ 1 , /* mNumStagesMmaWithinWorkTile */ 1 -, /* mNumStagesMmaAcrossWorkTile */ 2 -, /* mNumStagesSmemSfA */ 9 -, /* mNumStagesSmemSfB */ 9 -, /* mNumStagesTmemSfA */ 9 -, /* mNumStagesTmemSfB */ 9 +, /* mNumStagesMmaAcrossWorkTile */ 1 +, /* mNumStagesSmemSfA */ 5 +, /* mNumStagesSmemSfB */ 5 +, /* mNumStagesTmemSfA */ 5 +, /* mNumStagesTmemSfB */ 5 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -39707,10 +45464,10 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) , /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 256 +, /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 32 -, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTileScheduler */ gemm::TileScheduler(0) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 , /* mUseCustomMmaSchedule */ 1 @@ -39727,15 +45484,15 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 256 +, /* mValidK */ 512 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(3) +, /* mActType */ gemmGatedAct::ActType(2) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 0 +, /* mFusedAct */ 1 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 @@ -39750,17 +45507,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 , /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 215056, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f", 640, "1cf3a4604b0cfcc3c259db26c34eab0606cc9703bdd554298e8b69f6bf22eec4", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 216600, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 768, "75f4b8c897a4f3ab6488356d8ed2e199fbbf82c2c67ed91d684b383d24c83b4a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 -, /* mClusterDimX */ 1 +, /* mClusterDimX */ 2 , /* mClusterDimY */ 1 , /* mClusterDimZ */ 1 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) @@ -39791,16 +45548,16 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 256 +, /* mK */ 512 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) , /* mM */ 256 , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) -, /* mMmaM */ 128 +, /* mMmaM */ 256 , /* mMmaN */ 32 -, /* mMmaTileK */ 256 +, /* mMmaTileK */ 512 , /* mMockAllReduce */ 0 , /* mN */ 256 , /* mNumEpilogueWarps */ 4 @@ -39810,15 +45567,15 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 9 -, /* mNumStagesB */ 9 +, /* mNumStagesA */ 5 +, /* mNumStagesB */ 5 , /* mNumStagesMma */ 1 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 1 -, /* mNumStagesSmemSfA */ 9 -, /* mNumStagesSmemSfB */ 9 -, /* mNumStagesTmemSfA */ 9 -, /* mNumStagesTmemSfB */ 9 +, /* mNumStagesSmemSfA */ 5 +, /* mNumStagesSmemSfB */ 5 +, /* mNumStagesTmemSfA */ 5 +, /* mNumStagesTmemSfB */ 5 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -39833,7 +45590,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) , /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 256 +, /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 32 , /* mTileScheduler */ gemm::TileScheduler(0) @@ -39853,9 +45610,9 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 256 +, /* mValidK */ 512 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(1) +, /* mActType */ gemmGatedAct::ActType(0) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} @@ -39876,17 +45633,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 , /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 215056, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "f6b9c68880afe76421606c898c10500ae3e2468ee15b2bc382b4cf2995acc280", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 216600, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 768, "f1cd0c5fd4b88c95e28c6bf22fa3ade995edf1e02d7420c1e581f0aba41e4e1a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 -, /* mClusterDimX */ 1 +, /* mClusterDimX */ 2 , /* mClusterDimY */ 1 , /* mClusterDimZ */ 1 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) @@ -39897,7 +45654,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEltwiseActType */ gemm::EltwiseActType(2) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 @@ -39917,16 +45674,16 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 256 +, /* mK */ 512 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) , /* mM */ 256 , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) -, /* mMmaM */ 128 +, /* mMmaM */ 256 , /* mMmaN */ 32 -, /* mMmaTileK */ 256 +, /* mMmaTileK */ 512 , /* mMockAllReduce */ 0 , /* mN */ 256 , /* mNumEpilogueWarps */ 4 @@ -39936,15 +45693,15 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 9 -, /* mNumStagesB */ 9 +, /* mNumStagesA */ 5 +, /* mNumStagesB */ 5 , /* mNumStagesMma */ 1 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 1 -, /* mNumStagesSmemSfA */ 9 -, /* mNumStagesSmemSfB */ 9 -, /* mNumStagesTmemSfA */ 9 -, /* mNumStagesTmemSfB */ 9 +, /* mNumStagesSmemSfA */ 5 +, /* mNumStagesSmemSfB */ 5 +, /* mNumStagesTmemSfA */ 5 +, /* mNumStagesTmemSfB */ 5 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -39959,7 +45716,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) , /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 256 +, /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 32 , /* mTileScheduler */ gemm::TileScheduler(0) @@ -39979,7 +45736,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 256 +, /* mValidK */ 512 , /* mWorldSize */ 1 , /* mActType */ gemmGatedAct::ActType(0) , /* mClampBeforeAct */ 1 @@ -39987,7 +45744,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 1 +, /* mFusedAct */ 0 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 @@ -40002,17 +45759,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 , /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 215056, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 640, "f93cd33d47aafcccfd5979a2ec457ec545bcc7dc12cf7d5bac5f4ba2c58e02b5", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 216600, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 768, "7c22fc0d64bd1d607e440b2ba3acee08251689791db55fa108c4c56cc7ad3595", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 -, /* mClusterDimX */ 1 +, /* mClusterDimX */ 2 , /* mClusterDimY */ 1 , /* mClusterDimZ */ 1 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) @@ -40023,7 +45780,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(2) +, /* mEltwiseActType */ gemm::EltwiseActType(3) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 @@ -40043,16 +45800,16 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 256 +, /* mK */ 512 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) , /* mM */ 256 , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) -, /* mMmaM */ 128 +, /* mMmaM */ 256 , /* mMmaN */ 32 -, /* mMmaTileK */ 256 +, /* mMmaTileK */ 512 , /* mMockAllReduce */ 0 , /* mN */ 256 , /* mNumEpilogueWarps */ 4 @@ -40062,15 +45819,15 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 9 -, /* mNumStagesB */ 9 +, /* mNumStagesA */ 5 +, /* mNumStagesB */ 5 , /* mNumStagesMma */ 1 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 1 -, /* mNumStagesSmemSfA */ 9 -, /* mNumStagesSmemSfB */ 9 -, /* mNumStagesTmemSfA */ 9 -, /* mNumStagesTmemSfB */ 9 +, /* mNumStagesSmemSfA */ 5 +, /* mNumStagesSmemSfB */ 5 +, /* mNumStagesTmemSfA */ 5 +, /* mNumStagesTmemSfB */ 5 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -40085,7 +45842,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) , /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 256 +, /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 32 , /* mTileScheduler */ gemm::TileScheduler(0) @@ -40105,9 +45862,9 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 256 +, /* mValidK */ 512 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(0) +, /* mActType */ gemmGatedAct::ActType(3) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} @@ -40128,17 +45885,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 , /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 215056, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 640, "e25b71344195ba8d447662915f1591ea0b0ecb0b1ea02f11823ef84fcf407ba5", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 216984, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f", 896, "bb3f74f70b39df4a43f38caada7025de1d959df78b65cda2c3570d4a3ec5387a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 -, /* mClusterDimX */ 1 +, /* mClusterDimX */ 2 , /* mClusterDimY */ 1 , /* mClusterDimZ */ 1 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) @@ -40149,7 +45906,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(3) +, /* mEltwiseActType */ gemm::EltwiseActType(0) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 @@ -40169,16 +45926,16 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 256 +, /* mK */ 1024 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) , /* mM */ 256 , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) -, /* mMmaM */ 128 +, /* mMmaM */ 256 , /* mMmaN */ 32 -, /* mMmaTileK */ 256 +, /* mMmaTileK */ 512 , /* mMockAllReduce */ 0 , /* mN */ 256 , /* mNumEpilogueWarps */ 4 @@ -40188,15 +45945,15 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 9 -, /* mNumStagesB */ 9 -, /* mNumStagesMma */ 1 +, /* mNumStagesA */ 5 +, /* mNumStagesB */ 5 +, /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 -, /* mNumStagesMmaAcrossWorkTile */ 1 -, /* mNumStagesSmemSfA */ 9 -, /* mNumStagesSmemSfB */ 9 -, /* mNumStagesTmemSfA */ 9 -, /* mNumStagesTmemSfB */ 9 +, /* mNumStagesMmaAcrossWorkTile */ 2 +, /* mNumStagesSmemSfA */ 5 +, /* mNumStagesSmemSfB */ 5 +, /* mNumStagesTmemSfA */ 5 +, /* mNumStagesTmemSfB */ 5 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -40211,10 +45968,10 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) , /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 256 +, /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 32 -, /* mTileScheduler */ gemm::TileScheduler(0) +, /* mTileScheduler */ gemm::TileScheduler(1) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 , /* mUseCustomMmaSchedule */ 1 @@ -40228,18 +45985,18 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseTmaStore */ 1 , /* mUseTwoTmaLoadWarps */ 1 , /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 0 +, /* mUseUnrollLoop2xForMma */ 1 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 256 +, /* mValidK */ 1024 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(3) +, /* mActType */ gemmGatedAct::ActType(1) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 0 +, /* mFusedAct */ 1 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 @@ -40254,17 +46011,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 , /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 215184, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f", 640, "432cc5c5d5c2e6164b46d917fda5e06514b9d62b8a14e31102cdbb9c19e7077a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 216984, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f", 896, "d6ed20e9c8a8e7e0eb1e04218d841ce3e09ab79aff962bbb8b0954f009302fc7", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 -, /* mClusterDimX */ 1 +, /* mClusterDimX */ 2 , /* mClusterDimY */ 1 , /* mClusterDimZ */ 1 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) @@ -40295,16 +46052,16 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 512 +, /* mK */ 1024 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) , /* mM */ 256 , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) -, /* mMmaM */ 128 +, /* mMmaM */ 256 , /* mMmaN */ 32 -, /* mMmaTileK */ 256 +, /* mMmaTileK */ 512 , /* mMockAllReduce */ 0 , /* mN */ 256 , /* mNumEpilogueWarps */ 4 @@ -40314,15 +46071,15 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 9 -, /* mNumStagesB */ 9 +, /* mNumStagesA */ 5 +, /* mNumStagesB */ 5 , /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 2 -, /* mNumStagesSmemSfA */ 9 -, /* mNumStagesSmemSfB */ 9 -, /* mNumStagesTmemSfA */ 9 -, /* mNumStagesTmemSfB */ 9 +, /* mNumStagesSmemSfA */ 5 +, /* mNumStagesSmemSfB */ 5 +, /* mNumStagesTmemSfA */ 5 +, /* mNumStagesTmemSfB */ 5 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -40337,7 +46094,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) , /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 256 +, /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 32 , /* mTileScheduler */ gemm::TileScheduler(1) @@ -40357,9 +46114,9 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseUnrollLoop2xForMma */ 1 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 512 +, /* mValidK */ 1024 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(1) +, /* mActType */ gemmGatedAct::ActType(2) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} @@ -40380,17 +46137,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 , /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 215184, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "6cc28befcb46faad5bb875dae905e03bbc2335052a8c71531aa06b292b4871ae", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 216984, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 896, "33fa4babba7ef1fb33c14958324b690588c6d56cddb67f2db1f43a530f99e6a8", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 -, /* mClusterDimX */ 1 +, /* mClusterDimX */ 2 , /* mClusterDimY */ 1 , /* mClusterDimZ */ 1 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) @@ -40421,16 +46178,16 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 512 +, /* mK */ 1024 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) , /* mM */ 256 , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) -, /* mMmaM */ 128 +, /* mMmaM */ 256 , /* mMmaN */ 32 -, /* mMmaTileK */ 256 +, /* mMmaTileK */ 512 , /* mMockAllReduce */ 0 , /* mN */ 256 , /* mNumEpilogueWarps */ 4 @@ -40440,15 +46197,15 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 9 -, /* mNumStagesB */ 9 +, /* mNumStagesA */ 5 +, /* mNumStagesB */ 5 , /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 2 -, /* mNumStagesSmemSfA */ 9 -, /* mNumStagesSmemSfB */ 9 -, /* mNumStagesTmemSfA */ 9 -, /* mNumStagesTmemSfB */ 9 +, /* mNumStagesSmemSfA */ 5 +, /* mNumStagesSmemSfB */ 5 +, /* mNumStagesTmemSfA */ 5 +, /* mNumStagesTmemSfB */ 5 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -40463,7 +46220,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) , /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 256 +, /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 32 , /* mTileScheduler */ gemm::TileScheduler(1) @@ -40483,7 +46240,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseUnrollLoop2xForMma */ 1 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 512 +, /* mValidK */ 1024 , /* mWorldSize */ 1 , /* mActType */ gemmGatedAct::ActType(0) , /* mClampBeforeAct */ 1 @@ -40506,17 +46263,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 , /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 215184, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 640, "f333d0dce2e1704471c5890f277d62cce2f38120a8d5a4a3434c1addcd1bfb8e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 216984, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 896, "974102794677920e192651d68fd7459239745d1ec90a6fc432db3460c82b098d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 -, /* mClusterDimX */ 1 +, /* mClusterDimX */ 2 , /* mClusterDimY */ 1 , /* mClusterDimZ */ 1 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) @@ -40547,16 +46304,16 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 512 +, /* mK */ 1024 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) , /* mM */ 256 , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) -, /* mMmaM */ 128 +, /* mMmaM */ 256 , /* mMmaN */ 32 -, /* mMmaTileK */ 256 +, /* mMmaTileK */ 512 , /* mMockAllReduce */ 0 , /* mN */ 256 , /* mNumEpilogueWarps */ 4 @@ -40566,15 +46323,15 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 9 -, /* mNumStagesB */ 9 +, /* mNumStagesA */ 5 +, /* mNumStagesB */ 5 , /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 2 -, /* mNumStagesSmemSfA */ 9 -, /* mNumStagesSmemSfB */ 9 -, /* mNumStagesTmemSfA */ 9 -, /* mNumStagesTmemSfB */ 9 +, /* mNumStagesSmemSfA */ 5 +, /* mNumStagesSmemSfB */ 5 +, /* mNumStagesTmemSfA */ 5 +, /* mNumStagesTmemSfB */ 5 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -40589,7 +46346,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) , /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 256 +, /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 32 , /* mTileScheduler */ gemm::TileScheduler(1) @@ -40609,7 +46366,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseUnrollLoop2xForMma */ 1 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 512 +, /* mValidK */ 1024 , /* mWorldSize */ 1 , /* mActType */ gemmGatedAct::ActType(0) , /* mClampBeforeAct */ 1 @@ -40632,17 +46389,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 , /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 215184, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 640, "2b53fadc2d3706885737e4940c44ae3e996a2cff734ec20de580f14dbd5fb992", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 216984, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 896, "d211ce07be7a178e37586159270ebb0e75c1605bb4f5b51346a0e0781dd66268", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 -, /* mClusterDimX */ 1 +, /* mClusterDimX */ 2 , /* mClusterDimY */ 1 , /* mClusterDimZ */ 1 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) @@ -40673,16 +46430,16 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 512 +, /* mK */ 1024 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) , /* mM */ 256 , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) -, /* mMmaM */ 128 +, /* mMmaM */ 256 , /* mMmaN */ 32 -, /* mMmaTileK */ 256 +, /* mMmaTileK */ 512 , /* mMockAllReduce */ 0 , /* mN */ 256 , /* mNumEpilogueWarps */ 4 @@ -40692,15 +46449,15 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 9 -, /* mNumStagesB */ 9 +, /* mNumStagesA */ 5 +, /* mNumStagesB */ 5 , /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 2 -, /* mNumStagesSmemSfA */ 9 -, /* mNumStagesSmemSfB */ 9 -, /* mNumStagesTmemSfA */ 9 -, /* mNumStagesTmemSfB */ 9 +, /* mNumStagesSmemSfA */ 5 +, /* mNumStagesSmemSfB */ 5 +, /* mNumStagesTmemSfA */ 5 +, /* mNumStagesTmemSfB */ 5 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -40715,7 +46472,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) , /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 256 +, /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 32 , /* mTileScheduler */ gemm::TileScheduler(1) @@ -40735,7 +46492,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseUnrollLoop2xForMma */ 1 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 512 +, /* mValidK */ 1024 , /* mWorldSize */ 1 , /* mActType */ gemmGatedAct::ActType(3) , /* mClampBeforeAct */ 1 @@ -40758,17 +46515,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 , /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 215056, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f", 640, "757e46a221011bceeea41ab9c950e766f4dada1eeb2a1327c7c0cbab53431e8b", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 216600, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f", 768, "4cafff06d11d5a1929d386be4206bdf044daee0597893c7829c63e1380b2bc6b", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 -, /* mClusterDimX */ 1 +, /* mClusterDimX */ 2 , /* mClusterDimY */ 1 , /* mClusterDimZ */ 1 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) @@ -40799,16 +46556,16 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 512 +, /* mK */ 1024 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) , /* mM */ 256 , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) -, /* mMmaM */ 128 +, /* mMmaM */ 256 , /* mMmaN */ 32 -, /* mMmaTileK */ 256 +, /* mMmaTileK */ 512 , /* mMockAllReduce */ 0 , /* mN */ 256 , /* mNumEpilogueWarps */ 4 @@ -40818,15 +46575,15 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 9 -, /* mNumStagesB */ 9 +, /* mNumStagesA */ 5 +, /* mNumStagesB */ 5 , /* mNumStagesMma */ 1 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 1 -, /* mNumStagesSmemSfA */ 9 -, /* mNumStagesSmemSfB */ 9 -, /* mNumStagesTmemSfA */ 9 -, /* mNumStagesTmemSfB */ 9 +, /* mNumStagesSmemSfA */ 5 +, /* mNumStagesSmemSfB */ 5 +, /* mNumStagesTmemSfA */ 5 +, /* mNumStagesTmemSfB */ 5 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -40841,7 +46598,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) , /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 256 +, /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 32 , /* mTileScheduler */ gemm::TileScheduler(0) @@ -40861,7 +46618,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseUnrollLoop2xForMma */ 1 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 512 +, /* mValidK */ 1024 , /* mWorldSize */ 1 , /* mActType */ gemmGatedAct::ActType(1) , /* mClampBeforeAct */ 1 @@ -40884,17 +46641,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 , /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 215056, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "41c968f3fbd4acd910b9b3f64405860cd5b4a3bf927f28bd62626227de2dc89d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 216600, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f", 768, "65eb6fe8a2738ebd685aa22b7252ddf43750037eb90608d9e6e63276baa5c101", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 -, /* mClusterDimX */ 1 +, /* mClusterDimX */ 2 , /* mClusterDimY */ 1 , /* mClusterDimZ */ 1 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) @@ -40925,16 +46682,16 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 512 +, /* mK */ 1024 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) , /* mM */ 256 , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) -, /* mMmaM */ 128 +, /* mMmaM */ 256 , /* mMmaN */ 32 -, /* mMmaTileK */ 256 +, /* mMmaTileK */ 512 , /* mMockAllReduce */ 0 , /* mN */ 256 , /* mNumEpilogueWarps */ 4 @@ -40944,15 +46701,15 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 9 -, /* mNumStagesB */ 9 +, /* mNumStagesA */ 5 +, /* mNumStagesB */ 5 , /* mNumStagesMma */ 1 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 1 -, /* mNumStagesSmemSfA */ 9 -, /* mNumStagesSmemSfB */ 9 -, /* mNumStagesTmemSfA */ 9 -, /* mNumStagesTmemSfB */ 9 +, /* mNumStagesSmemSfA */ 5 +, /* mNumStagesSmemSfB */ 5 +, /* mNumStagesTmemSfA */ 5 +, /* mNumStagesTmemSfB */ 5 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -40967,7 +46724,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) , /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 256 +, /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 32 , /* mTileScheduler */ gemm::TileScheduler(0) @@ -40987,7 +46744,133 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseUnrollLoop2xForMma */ 1 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 512 +, /* mValidK */ 1024 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(2) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 1 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 0 +, /* mNumRegsPerThreadLoadSfA */ 0 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 0 +, /* mNumWarpsLoadSfA */ 0 +, /* mNumWarpsLoadSfB */ 0 +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 216600, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 768, "8caa15144249729e6b6276726d2be74c4c071bd88f4c38fc8bb86c65c8181eaa", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(1) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 2 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 1 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 16 +, /* mEpilogueLdtmBits */ 256 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 32 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 0 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 1024 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 256 +, /* mMmaN */ 32 +, /* mMmaTileK */ 512 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 5 +, /* mNumStagesB */ 5 +, /* mNumStagesMma */ 1 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 1 +, /* mNumStagesSmemSfA */ 5 +, /* mNumStagesSmemSfB */ 5 +, /* mNumStagesTmemSfA */ 5 +, /* mNumStagesTmemSfB */ 5 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(3) +, /* mSfLayoutB */ trtllm::gen::SfLayout(0) +, /* mSfLayoutC */ trtllm::gen::SfLayout(1) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 512 +, /* mTileM */ 128 +, /* mTileN */ 32 +, /* mTileScheduler */ gemm::TileScheduler(0) +, /* mTransposeMmaOutput */ 1 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 0 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 1 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 1024 , /* mWorldSize */ 1 , /* mActType */ gemmGatedAct::ActType(0) , /* mClampBeforeAct */ 1 @@ -41010,17 +46893,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 , /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 215056, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 640, "0da1fb14827844701f93372b22d1f1847e730ba8f41bb0863acacc9b9b74baf2", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 216600, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 768, "1548e1d741fe34bf9f762e0df6df9b81fbca299fa806c9f6298c7420df44fa7e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 -, /* mClusterDimX */ 1 +, /* mClusterDimX */ 2 , /* mClusterDimY */ 1 , /* mClusterDimZ */ 1 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) @@ -41051,16 +46934,16 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 512 +, /* mK */ 1024 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) , /* mM */ 256 , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) -, /* mMmaM */ 128 +, /* mMmaM */ 256 , /* mMmaN */ 32 -, /* mMmaTileK */ 256 +, /* mMmaTileK */ 512 , /* mMockAllReduce */ 0 , /* mN */ 256 , /* mNumEpilogueWarps */ 4 @@ -41070,15 +46953,15 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 9 -, /* mNumStagesB */ 9 +, /* mNumStagesA */ 5 +, /* mNumStagesB */ 5 , /* mNumStagesMma */ 1 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 1 -, /* mNumStagesSmemSfA */ 9 -, /* mNumStagesSmemSfB */ 9 -, /* mNumStagesTmemSfA */ 9 -, /* mNumStagesTmemSfB */ 9 +, /* mNumStagesSmemSfA */ 5 +, /* mNumStagesSmemSfB */ 5 +, /* mNumStagesTmemSfA */ 5 +, /* mNumStagesTmemSfB */ 5 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -41093,7 +46976,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) , /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 256 +, /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 32 , /* mTileScheduler */ gemm::TileScheduler(0) @@ -41113,7 +46996,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseUnrollLoop2xForMma */ 1 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 512 +, /* mValidK */ 1024 , /* mWorldSize */ 1 , /* mActType */ gemmGatedAct::ActType(0) , /* mClampBeforeAct */ 1 @@ -41136,17 +47019,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 , /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 215056, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 640, "f9b27ab662a0bb14e48d66481075bebfe9e35a6b1c82e78de576292fc9bb1876", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 216600, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 768, "cee632d2d27c1f4b1fdd43ecf552df87d5d07415c77bc400bdb1d1ed1b3192c1", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 -, /* mClusterDimX */ 1 +, /* mClusterDimX */ 2 , /* mClusterDimY */ 1 , /* mClusterDimZ */ 1 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) @@ -41177,16 +47060,16 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 512 +, /* mK */ 1024 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) , /* mM */ 256 , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) -, /* mMmaM */ 128 +, /* mMmaM */ 256 , /* mMmaN */ 32 -, /* mMmaTileK */ 256 +, /* mMmaTileK */ 512 , /* mMockAllReduce */ 0 , /* mN */ 256 , /* mNumEpilogueWarps */ 4 @@ -41196,15 +47079,15 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 9 -, /* mNumStagesB */ 9 +, /* mNumStagesA */ 5 +, /* mNumStagesB */ 5 , /* mNumStagesMma */ 1 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 1 -, /* mNumStagesSmemSfA */ 9 -, /* mNumStagesSmemSfB */ 9 -, /* mNumStagesTmemSfA */ 9 -, /* mNumStagesTmemSfB */ 9 +, /* mNumStagesSmemSfA */ 5 +, /* mNumStagesSmemSfB */ 5 +, /* mNumStagesTmemSfA */ 5 +, /* mNumStagesTmemSfB */ 5 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -41219,7 +47102,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) , /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 256 +, /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 32 , /* mTileScheduler */ gemm::TileScheduler(0) @@ -41239,7 +47122,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseUnrollLoop2xForMma */ 1 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 512 +, /* mValidK */ 1024 , /* mWorldSize */ 1 , /* mActType */ gemmGatedAct::ActType(3) , /* mClampBeforeAct */ 1 @@ -41262,19 +47145,19 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 , /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 175768, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f", 896, "a0f856b937942695890b99f63190a45d749d83a3a3515a69e807f87798277bb6", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 151448, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f", 640, "d92e699a22799e47fb7cad428ba6411e16715d174f96ce50653cec1a42996b3e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 -, /* mClusterDimX */ 1 +, /* mClusterDimX */ 2 , /* mClusterDimY */ 1 -, /* mClusterDimZ */ 2 +, /* mClusterDimZ */ 1 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) , /* mDtypeAcc */ trtllm::gen::Dtype(1056776) , /* mDtypeA */ trtllm::gen::Dtype(17826818) @@ -41289,7 +47172,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mEpilogueLdtmDps */ 16 , /* mEpilogueLdtmBits */ 256 , /* mEpilogueTileM */ 128 -, /* mEpilogueTileN */ 32 +, /* mEpilogueTileN */ 64 , /* mFallbackClusterDimX */ 1 , /* mFallbackClusterDimY */ 1 , /* mFallbackClusterDimZ */ 1 @@ -41303,16 +47186,16 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 1024 +, /* mK */ 256 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) , /* mM */ 256 , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) -, /* mMmaM */ 128 -, /* mMmaN */ 32 -, /* mMmaTileK */ 512 +, /* mMmaM */ 256 +, /* mMmaN */ 64 +, /* mMmaTileK */ 256 , /* mMockAllReduce */ 0 , /* mN */ 256 , /* mNumEpilogueWarps */ 4 @@ -41320,17 +47203,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySfLdsSttm */ 0 , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 -, /* mNumRegsPerThreadNonEpilogueWarp */ 0 -, /* mNumSlicesForSplitK */ 2 -, /* mNumStagesA */ 3 -, /* mNumStagesB */ 3 +, /* mNumRegsPerThreadNonEpilogueWarp */ 48 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 6 +, /* mNumStagesB */ 6 , /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 2 -, /* mNumStagesSmemSfA */ 3 -, /* mNumStagesSmemSfB */ 3 -, /* mNumStagesTmemSfA */ 3 -, /* mNumStagesTmemSfB */ 3 +, /* mNumStagesSmemSfA */ 6 +, /* mNumStagesSmemSfB */ 6 +, /* mNumStagesTmemSfA */ 6 +, /* mNumStagesTmemSfB */ 6 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -41344,10 +47227,10 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfLayoutC */ trtllm::gen::SfLayout(1) , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) -, /* mSplitK */ gemm::SplitK(2) -, /* mTileK */ 512 +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 256 , /* mTileM */ 128 -, /* mTileN */ 32 +, /* mTileN */ 64 , /* mTileScheduler */ gemm::TileScheduler(1) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 @@ -41365,7 +47248,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 1024 +, /* mValidK */ 256 , /* mWorldSize */ 1 , /* mActType */ gemmGatedAct::ActType(1) , /* mClampBeforeAct */ 1 @@ -41388,19 +47271,19 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 , /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 175768, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 896, "a79089c8dafb298893b4af40453ba24c0c395179dfc6103089e3d04ab9d0e955", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 151448, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 640, "004480694e248a536a1e505a7f0c955c14f1bfb4afd4de433b26d2ebf8146212", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 -, /* mClusterDimX */ 1 +, /* mClusterDimX */ 2 , /* mClusterDimY */ 1 -, /* mClusterDimZ */ 2 +, /* mClusterDimZ */ 1 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) , /* mDtypeAcc */ trtllm::gen::Dtype(1056776) , /* mDtypeA */ trtllm::gen::Dtype(17826818) @@ -41409,13 +47292,13 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(2) +, /* mEltwiseActType */ gemm::EltwiseActType(0) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 , /* mEpilogueLdtmBits */ 256 , /* mEpilogueTileM */ 128 -, /* mEpilogueTileN */ 32 +, /* mEpilogueTileN */ 64 , /* mFallbackClusterDimX */ 1 , /* mFallbackClusterDimY */ 1 , /* mFallbackClusterDimZ */ 1 @@ -41429,16 +47312,16 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 1024 +, /* mK */ 256 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) , /* mM */ 256 , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) -, /* mMmaM */ 128 -, /* mMmaN */ 32 -, /* mMmaTileK */ 512 +, /* mMmaM */ 256 +, /* mMmaN */ 64 +, /* mMmaTileK */ 256 , /* mMockAllReduce */ 0 , /* mN */ 256 , /* mNumEpilogueWarps */ 4 @@ -41446,17 +47329,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySfLdsSttm */ 0 , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 -, /* mNumRegsPerThreadNonEpilogueWarp */ 0 -, /* mNumSlicesForSplitK */ 2 -, /* mNumStagesA */ 3 -, /* mNumStagesB */ 3 +, /* mNumRegsPerThreadNonEpilogueWarp */ 48 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 6 +, /* mNumStagesB */ 6 , /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 2 -, /* mNumStagesSmemSfA */ 3 -, /* mNumStagesSmemSfB */ 3 -, /* mNumStagesTmemSfA */ 3 -, /* mNumStagesTmemSfB */ 3 +, /* mNumStagesSmemSfA */ 6 +, /* mNumStagesSmemSfB */ 6 +, /* mNumStagesTmemSfA */ 6 +, /* mNumStagesTmemSfB */ 6 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -41470,10 +47353,10 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfLayoutC */ trtllm::gen::SfLayout(1) , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) -, /* mSplitK */ gemm::SplitK(2) -, /* mTileK */ 512 +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 256 , /* mTileM */ 128 -, /* mTileN */ 32 +, /* mTileN */ 64 , /* mTileScheduler */ gemm::TileScheduler(1) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 @@ -41491,15 +47374,15 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 1024 +, /* mValidK */ 256 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(3) +, /* mActType */ gemmGatedAct::ActType(2) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 0 +, /* mFusedAct */ 1 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 @@ -41514,19 +47397,19 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 , /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 175768, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 896, "b89ff6ddd24f8d90e97639991c0d1058679efa4913802c5accdb661ea59d818b", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 151448, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "18d5de73b962847085182ae0d05cf1124b43a7bd33d3ee3535d69b600cec6a61", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 -, /* mClusterDimX */ 1 +, /* mClusterDimX */ 2 , /* mClusterDimY */ 1 -, /* mClusterDimZ */ 2 +, /* mClusterDimZ */ 1 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) , /* mDtypeAcc */ trtllm::gen::Dtype(1056776) , /* mDtypeA */ trtllm::gen::Dtype(17826818) @@ -41535,13 +47418,13 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(3) +, /* mEltwiseActType */ gemm::EltwiseActType(0) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 , /* mEpilogueLdtmBits */ 256 , /* mEpilogueTileM */ 128 -, /* mEpilogueTileN */ 32 +, /* mEpilogueTileN */ 64 , /* mFallbackClusterDimX */ 1 , /* mFallbackClusterDimY */ 1 , /* mFallbackClusterDimZ */ 1 @@ -41555,16 +47438,16 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 1024 +, /* mK */ 256 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) , /* mM */ 256 , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) -, /* mMmaM */ 128 -, /* mMmaN */ 32 -, /* mMmaTileK */ 512 +, /* mMmaM */ 256 +, /* mMmaN */ 64 +, /* mMmaTileK */ 256 , /* mMockAllReduce */ 0 , /* mN */ 256 , /* mNumEpilogueWarps */ 4 @@ -41572,17 +47455,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySfLdsSttm */ 0 , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 -, /* mNumRegsPerThreadNonEpilogueWarp */ 0 -, /* mNumSlicesForSplitK */ 2 -, /* mNumStagesA */ 3 -, /* mNumStagesB */ 3 +, /* mNumRegsPerThreadNonEpilogueWarp */ 48 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 6 +, /* mNumStagesB */ 6 , /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 2 -, /* mNumStagesSmemSfA */ 3 -, /* mNumStagesSmemSfB */ 3 -, /* mNumStagesTmemSfA */ 3 -, /* mNumStagesTmemSfB */ 3 +, /* mNumStagesSmemSfA */ 6 +, /* mNumStagesSmemSfB */ 6 +, /* mNumStagesTmemSfA */ 6 +, /* mNumStagesTmemSfB */ 6 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -41596,10 +47479,10 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfLayoutC */ trtllm::gen::SfLayout(1) , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) -, /* mSplitK */ gemm::SplitK(2) -, /* mTileK */ 512 +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 256 , /* mTileM */ 128 -, /* mTileN */ 32 +, /* mTileN */ 64 , /* mTileScheduler */ gemm::TileScheduler(1) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 @@ -41617,7 +47500,133 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 1024 +, /* mValidK */ 256 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(0) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 1 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 0 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 128 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 0 +, /* mNumRegsPerThreadLoadSfA */ 0 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 2 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 0 +, /* mNumWarpsLoadSfA */ 0 +, /* mNumWarpsLoadSfB */ 0 +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 151448, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 640, "9ea72d46d9e35e2fe1e751b742884becf603cdb58e803076c3f98a997759732a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(1) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 2 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 1 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(2) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 16 +, /* mEpilogueLdtmBits */ 256 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 64 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 0 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 256 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 256 +, /* mMmaN */ 64 +, /* mMmaTileK */ 256 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 48 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 6 +, /* mNumStagesB */ 6 +, /* mNumStagesMma */ 2 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 2 +, /* mNumStagesSmemSfA */ 6 +, /* mNumStagesSmemSfB */ 6 +, /* mNumStagesTmemSfA */ 6 +, /* mNumStagesTmemSfB */ 6 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(3) +, /* mSfLayoutB */ trtllm::gen::SfLayout(0) +, /* mSfLayoutC */ trtllm::gen::SfLayout(1) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 256 +, /* mTileM */ 128 +, /* mTileN */ 64 +, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTransposeMmaOutput */ 1 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 0 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 0 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 256 , /* mWorldSize */ 1 , /* mActType */ gemmGatedAct::ActType(3) , /* mClampBeforeAct */ 1 @@ -41640,12 +47649,12 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 , /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 216984, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f", 896, "1db4be9e0ebe9c6b7dfe23880a2a16115aee8614dc5b11197c474c5fc5cfa5db", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 151448, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 640, "9b07d1b1afd7c1d06bf39ae82e6e8df139d77d04c5f762967dca25cda944a16c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -41661,13 +47670,13 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEltwiseActType */ gemm::EltwiseActType(3) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 , /* mEpilogueLdtmBits */ 256 , /* mEpilogueTileM */ 128 -, /* mEpilogueTileN */ 32 +, /* mEpilogueTileN */ 64 , /* mFallbackClusterDimX */ 1 , /* mFallbackClusterDimY */ 1 , /* mFallbackClusterDimZ */ 1 @@ -41681,7 +47690,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 512 +, /* mK */ 256 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) @@ -41689,8 +47698,8 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) , /* mMmaM */ 256 -, /* mMmaN */ 32 -, /* mMmaTileK */ 512 +, /* mMmaN */ 64 +, /* mMmaTileK */ 256 , /* mMockAllReduce */ 0 , /* mN */ 256 , /* mNumEpilogueWarps */ 4 @@ -41698,17 +47707,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySfLdsSttm */ 0 , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 -, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 48 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 5 -, /* mNumStagesB */ 5 +, /* mNumStagesA */ 6 +, /* mNumStagesB */ 6 , /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 2 -, /* mNumStagesSmemSfA */ 5 -, /* mNumStagesSmemSfB */ 5 -, /* mNumStagesTmemSfA */ 5 -, /* mNumStagesTmemSfB */ 5 +, /* mNumStagesSmemSfA */ 6 +, /* mNumStagesSmemSfB */ 6 +, /* mNumStagesTmemSfA */ 6 +, /* mNumStagesTmemSfB */ 6 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -41723,9 +47732,9 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) , /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 512 +, /* mTileK */ 256 , /* mTileM */ 128 -, /* mTileN */ 32 +, /* mTileN */ 64 , /* mTileScheduler */ gemm::TileScheduler(1) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 @@ -41743,15 +47752,15 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 512 +, /* mValidK */ 256 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(1) +, /* mActType */ gemmGatedAct::ActType(3) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 1 +, /* mFusedAct */ 0 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 @@ -41766,12 +47775,12 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 , /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 216984, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 896, "85a869d77ed64db2ad22f80f0dc7f78a25def7a7cb68ef8885d8e5986efdbb41", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 151448, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f", 640, "efa96932ef02f9517ccd8f96c21b048897e8d82d3ba7d3ea34173ffd1768799f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -41793,7 +47802,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mEpilogueLdtmDps */ 16 , /* mEpilogueLdtmBits */ 256 , /* mEpilogueTileM */ 128 -, /* mEpilogueTileN */ 32 +, /* mEpilogueTileN */ 64 , /* mFallbackClusterDimX */ 1 , /* mFallbackClusterDimY */ 1 , /* mFallbackClusterDimZ */ 1 @@ -41815,8 +47824,8 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) , /* mMmaM */ 256 -, /* mMmaN */ 32 -, /* mMmaTileK */ 512 +, /* mMmaN */ 64 +, /* mMmaTileK */ 256 , /* mMockAllReduce */ 0 , /* mN */ 256 , /* mNumEpilogueWarps */ 4 @@ -41824,17 +47833,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySfLdsSttm */ 0 , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 -, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 48 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 5 -, /* mNumStagesB */ 5 +, /* mNumStagesA */ 6 +, /* mNumStagesB */ 6 , /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 2 -, /* mNumStagesSmemSfA */ 5 -, /* mNumStagesSmemSfB */ 5 -, /* mNumStagesTmemSfA */ 5 -, /* mNumStagesTmemSfB */ 5 +, /* mNumStagesSmemSfA */ 6 +, /* mNumStagesSmemSfB */ 6 +, /* mNumStagesTmemSfA */ 6 +, /* mNumStagesTmemSfB */ 6 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -41849,9 +47858,9 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) , /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 512 +, /* mTileK */ 256 , /* mTileM */ 128 -, /* mTileN */ 32 +, /* mTileN */ 64 , /* mTileScheduler */ gemm::TileScheduler(1) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 @@ -41866,12 +47875,12 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseTmaStore */ 1 , /* mUseTwoTmaLoadWarps */ 1 , /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 0 +, /* mUseUnrollLoop2xForMma */ 1 , /* mValidM */ 256 , /* mValidN */ 256 , /* mValidK */ 512 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(0) +, /* mActType */ gemmGatedAct::ActType(1) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} @@ -41892,12 +47901,12 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 , /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 216984, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 896, "5590da7579c4042a2d76a446db9246d70f0048ac98e188e730c8b9365962fd87", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 151448, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 640, "f5f5421abc1e90f334c65f29d583b3ecc22113e8b1bedde5d50e4c7d28f22acb", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -41913,13 +47922,13 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(2) +, /* mEltwiseActType */ gemm::EltwiseActType(0) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 , /* mEpilogueLdtmBits */ 256 , /* mEpilogueTileM */ 128 -, /* mEpilogueTileN */ 32 +, /* mEpilogueTileN */ 64 , /* mFallbackClusterDimX */ 1 , /* mFallbackClusterDimY */ 1 , /* mFallbackClusterDimZ */ 1 @@ -41941,8 +47950,8 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) , /* mMmaM */ 256 -, /* mMmaN */ 32 -, /* mMmaTileK */ 512 +, /* mMmaN */ 64 +, /* mMmaTileK */ 256 , /* mMockAllReduce */ 0 , /* mN */ 256 , /* mNumEpilogueWarps */ 4 @@ -41950,17 +47959,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySfLdsSttm */ 0 , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 -, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 48 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 5 -, /* mNumStagesB */ 5 +, /* mNumStagesA */ 6 +, /* mNumStagesB */ 6 , /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 2 -, /* mNumStagesSmemSfA */ 5 -, /* mNumStagesSmemSfB */ 5 -, /* mNumStagesTmemSfA */ 5 -, /* mNumStagesTmemSfB */ 5 +, /* mNumStagesSmemSfA */ 6 +, /* mNumStagesSmemSfB */ 6 +, /* mNumStagesTmemSfA */ 6 +, /* mNumStagesTmemSfB */ 6 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -41975,9 +47984,9 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) , /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 512 +, /* mTileK */ 256 , /* mTileM */ 128 -, /* mTileN */ 32 +, /* mTileN */ 64 , /* mTileScheduler */ gemm::TileScheduler(1) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 @@ -41992,18 +48001,18 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseTmaStore */ 1 , /* mUseTwoTmaLoadWarps */ 1 , /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 0 +, /* mUseUnrollLoop2xForMma */ 1 , /* mValidM */ 256 , /* mValidN */ 256 , /* mValidK */ 512 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(0) +, /* mActType */ gemmGatedAct::ActType(2) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 0 +, /* mFusedAct */ 1 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 @@ -42018,12 +48027,12 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 , /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 216984, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 896, "29bfcacbe0656a1b44c9edebf04a10740d4eaeb6be3f55cf9b11997bb2a506f6", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 151448, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "3670a762a818d8539bbb1526e87ed3ba1ee2eb4f6348778d8a0638b11fc11193", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -42039,13 +48048,13 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(3) +, /* mEltwiseActType */ gemm::EltwiseActType(0) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 , /* mEpilogueLdtmBits */ 256 , /* mEpilogueTileM */ 128 -, /* mEpilogueTileN */ 32 +, /* mEpilogueTileN */ 64 , /* mFallbackClusterDimX */ 1 , /* mFallbackClusterDimY */ 1 , /* mFallbackClusterDimZ */ 1 @@ -42067,8 +48076,8 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) , /* mMmaM */ 256 -, /* mMmaN */ 32 -, /* mMmaTileK */ 512 +, /* mMmaN */ 64 +, /* mMmaTileK */ 256 , /* mMockAllReduce */ 0 , /* mN */ 256 , /* mNumEpilogueWarps */ 4 @@ -42076,17 +48085,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySfLdsSttm */ 0 , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 -, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 48 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 5 -, /* mNumStagesB */ 5 +, /* mNumStagesA */ 6 +, /* mNumStagesB */ 6 , /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 2 -, /* mNumStagesSmemSfA */ 5 -, /* mNumStagesSmemSfB */ 5 -, /* mNumStagesTmemSfA */ 5 -, /* mNumStagesTmemSfB */ 5 +, /* mNumStagesSmemSfA */ 6 +, /* mNumStagesSmemSfB */ 6 +, /* mNumStagesTmemSfA */ 6 +, /* mNumStagesTmemSfB */ 6 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -42101,9 +48110,9 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) , /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 512 +, /* mTileK */ 256 , /* mTileM */ 128 -, /* mTileN */ 32 +, /* mTileN */ 64 , /* mTileScheduler */ gemm::TileScheduler(1) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 @@ -42118,138 +48127,12 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseTmaStore */ 1 , /* mUseTwoTmaLoadWarps */ 1 , /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 0 -, /* mValidM */ 256 -, /* mValidN */ 256 -, /* mValidK */ 512 -, /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(3) -, /* mClampBeforeAct */ 1 -, /* mBatchedM */ {} -, /* mBatchedN */ {} -, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) -, /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 0 -, /* mGridWaitForPrimaryRouting */ 1 -, /* mIsStaticBatch */ 0 -, /* mIsUniformNumTokensPerBatch */ 0 -, /* mNumBatches */ 128 -, /* mNumRegsPerThreadLoadA */ 0 -, /* mNumRegsPerThreadLoadB */ 0 -, /* mNumRegsPerThreadLoadSfA */ 0 -, /* mNumRegsPerThreadLoadSfB */ 0 -, /* mNumTokens */ 2 -, /* mNumWarpsLoadA */ 0 -, /* mNumWarpsLoadB */ 0 -, /* mNumWarpsLoadSfA */ 0 -, /* mNumWarpsLoadSfB */ 0 -, /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} -, /* mPrefetchB */ 0 -, /* mUseCMultiCast */ 0 -, /* mUseTmaOobOpt */ 1 - }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 216600, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f", 768, "99e879b8ed8c8bb5fd1719f2bdb3472893778a978525ad31f5fc6a596f7cf6e4", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) -, /* mBiasDtype */ trtllm::gen::Dtype(1056776) -, /* mBiasType */ gemm::BiasType(1) -, /* mBlockK */ -1 -, /* mClcFastDrain */ 1 -, /* mClusterDimX */ 2 -, /* mClusterDimY */ 1 -, /* mClusterDimZ */ 1 -, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) -, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) -, /* mDtypeA */ trtllm::gen::Dtype(17826818) -, /* mDtypeB */ trtllm::gen::Dtype(17826818) -, /* mDtypeC */ trtllm::gen::Dtype(17826818) -, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) -, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) -, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(0) -, /* mEnablesEarlyExit */ 1 -, /* mEnablesDelayedEarlyExit */ 0 -, /* mEpilogueLdtmDps */ 16 -, /* mEpilogueLdtmBits */ 256 -, /* mEpilogueTileM */ 128 -, /* mEpilogueTileN */ 32 -, /* mFallbackClusterDimX */ 1 -, /* mFallbackClusterDimY */ 1 -, /* mFallbackClusterDimZ */ 1 -, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) -, /* mFuseLoadSfTask */ 0 -, /* mFuseUtccpWithUtcmma */ 0 -, /* mGridTriggerSecondaryA */ 0 -, /* mGridTriggerSecondaryB */ 1 -, /* mGridWaitForPrimaryEarlyExit */ 1 -, /* mGridWaitForPrimaryA */ 0 -, /* mGridWaitForPrimaryB */ 1 -, /* mHoistLoadTaskInit */ 1 -, /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 512 -, /* mKernelTraits */ {} -, /* mLayoutA */ gemm::MatrixLayout(0) -, /* mLayoutB */ gemm::MatrixLayout(0) -, /* mM */ 256 -, /* mMmaK */ 64 -, /* mMmaKind */ trtllm::gen::MmaKind(4) -, /* mMmaM */ 256 -, /* mMmaN */ 32 -, /* mMmaTileK */ 512 -, /* mMockAllReduce */ 0 -, /* mN */ 256 -, /* mNumEpilogueWarps */ 4 -, /* mNumRegsCastAWarps */ 0 -, /* mNumRegsCopySfLdsSttm */ 0 -, /* mNumRegsCopySparsityInfo */ 0 -, /* mNumRegsPerThreadEpilogueWarp */ 0 -, /* mNumRegsPerThreadNonEpilogueWarp */ 0 -, /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 5 -, /* mNumStagesB */ 5 -, /* mNumStagesMma */ 1 -, /* mNumStagesMmaWithinWorkTile */ 1 -, /* mNumStagesMmaAcrossWorkTile */ 1 -, /* mNumStagesSmemSfA */ 5 -, /* mNumStagesSmemSfB */ 5 -, /* mNumStagesTmemSfA */ 5 -, /* mNumStagesTmemSfB */ 5 -, /* mNumStagesWorkId */ 3 -, /* mOutputDebugTensors */ 0 -, /* mPatchF2fp */ 0 -, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) -, /* mSchedHostTask */ gemm::SchedHostTask(0) -, /* mSfBlockSizeA */ 16 -, /* mSfBlockSizeB */ 16 -, /* mSfBlockSizeC */ 16 -, /* mSfLayoutA */ trtllm::gen::SfLayout(3) -, /* mSfLayoutB */ trtllm::gen::SfLayout(0) -, /* mSfLayoutC */ trtllm::gen::SfLayout(1) -, /* mSfReshapeFactor */ 1 -, /* mSparsityA */ trtllm::gen::Sparsity(0) -, /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 512 -, /* mTileM */ 128 -, /* mTileN */ 32 -, /* mTileScheduler */ gemm::TileScheduler(0) -, /* mTransposeMmaOutput */ 1 -, /* mUseCustomizedMma3xNvFp4 */ 0 -, /* mUseCustomMmaSchedule */ 1 -, /* mUseDeepSeekFp8 */ 0 -, /* mUseFlexibleClusterDims */ 0 -, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 -, /* mUseMaxTmemOverlap */ 0 -, /* mUsePerTokenSfA */ 0 -, /* mUsePerTokenSfB */ 0 -, /* mUseShuffledMatrix */ 1 -, /* mUseTmaStore */ 1 -, /* mUseTwoTmaLoadWarps */ 1 -, /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 0 +, /* mUseUnrollLoop2xForMma */ 1 , /* mValidM */ 256 , /* mValidN */ 256 , /* mValidK */ 512 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(1) +, /* mActType */ gemmGatedAct::ActType(0) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} @@ -42270,12 +48153,12 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 , /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 216600, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 768, "4daf61574a0f60ef270e43398d75bfc0911b710f217668f6ff265a4d2a558cfc", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 151448, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 640, "e22503d430ea57e6fc4b0d2661461a78e2eed7c521b872368edb94e0dce6ec0c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -42291,13 +48174,13 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEltwiseActType */ gemm::EltwiseActType(2) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 , /* mEpilogueLdtmBits */ 256 , /* mEpilogueTileM */ 128 -, /* mEpilogueTileN */ 32 +, /* mEpilogueTileN */ 64 , /* mFallbackClusterDimX */ 1 , /* mFallbackClusterDimY */ 1 , /* mFallbackClusterDimZ */ 1 @@ -42319,8 +48202,8 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) , /* mMmaM */ 256 -, /* mMmaN */ 32 -, /* mMmaTileK */ 512 +, /* mMmaN */ 64 +, /* mMmaTileK */ 256 , /* mMockAllReduce */ 0 , /* mN */ 256 , /* mNumEpilogueWarps */ 4 @@ -42328,17 +48211,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySfLdsSttm */ 0 , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 -, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 48 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 5 -, /* mNumStagesB */ 5 -, /* mNumStagesMma */ 1 +, /* mNumStagesA */ 6 +, /* mNumStagesB */ 6 +, /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 -, /* mNumStagesMmaAcrossWorkTile */ 1 -, /* mNumStagesSmemSfA */ 5 -, /* mNumStagesSmemSfB */ 5 -, /* mNumStagesTmemSfA */ 5 -, /* mNumStagesTmemSfB */ 5 +, /* mNumStagesMmaAcrossWorkTile */ 2 +, /* mNumStagesSmemSfA */ 6 +, /* mNumStagesSmemSfB */ 6 +, /* mNumStagesTmemSfA */ 6 +, /* mNumStagesTmemSfB */ 6 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -42353,10 +48236,10 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) , /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 512 +, /* mTileK */ 256 , /* mTileM */ 128 -, /* mTileN */ 32 -, /* mTileScheduler */ gemm::TileScheduler(0) +, /* mTileN */ 64 +, /* mTileScheduler */ gemm::TileScheduler(1) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 , /* mUseCustomMmaSchedule */ 1 @@ -42370,18 +48253,18 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseTmaStore */ 1 , /* mUseTwoTmaLoadWarps */ 1 , /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 0 +, /* mUseUnrollLoop2xForMma */ 1 , /* mValidM */ 256 , /* mValidN */ 256 , /* mValidK */ 512 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(0) +, /* mActType */ gemmGatedAct::ActType(3) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 1 +, /* mFusedAct */ 0 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 @@ -42396,12 +48279,12 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 , /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 216600, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 768, "ab56ae46fb46980898000d681f492bbe88e237438ee945a13836774b37224c7e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 151448, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 640, "e5e39ede7414dc728cf6a07992dac91d4d1579780f776b73e625054b569efb16", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -42417,13 +48300,13 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(2) +, /* mEltwiseActType */ gemm::EltwiseActType(3) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 , /* mEpilogueLdtmBits */ 256 , /* mEpilogueTileM */ 128 -, /* mEpilogueTileN */ 32 +, /* mEpilogueTileN */ 64 , /* mFallbackClusterDimX */ 1 , /* mFallbackClusterDimY */ 1 , /* mFallbackClusterDimZ */ 1 @@ -42445,8 +48328,8 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) , /* mMmaM */ 256 -, /* mMmaN */ 32 -, /* mMmaTileK */ 512 +, /* mMmaN */ 64 +, /* mMmaTileK */ 256 , /* mMockAllReduce */ 0 , /* mN */ 256 , /* mNumEpilogueWarps */ 4 @@ -42454,17 +48337,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySfLdsSttm */ 0 , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 -, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 48 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 5 -, /* mNumStagesB */ 5 -, /* mNumStagesMma */ 1 +, /* mNumStagesA */ 6 +, /* mNumStagesB */ 6 +, /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 -, /* mNumStagesMmaAcrossWorkTile */ 1 -, /* mNumStagesSmemSfA */ 5 -, /* mNumStagesSmemSfB */ 5 -, /* mNumStagesTmemSfA */ 5 -, /* mNumStagesTmemSfB */ 5 +, /* mNumStagesMmaAcrossWorkTile */ 2 +, /* mNumStagesSmemSfA */ 6 +, /* mNumStagesSmemSfB */ 6 +, /* mNumStagesTmemSfA */ 6 +, /* mNumStagesTmemSfB */ 6 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -42479,10 +48362,10 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) , /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 512 +, /* mTileK */ 256 , /* mTileM */ 128 -, /* mTileN */ 32 -, /* mTileScheduler */ gemm::TileScheduler(0) +, /* mTileN */ 64 +, /* mTileScheduler */ gemm::TileScheduler(1) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 , /* mUseCustomMmaSchedule */ 1 @@ -42496,12 +48379,12 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseTmaStore */ 1 , /* mUseTwoTmaLoadWarps */ 1 , /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 0 +, /* mUseUnrollLoop2xForMma */ 1 , /* mValidM */ 256 , /* mValidN */ 256 , /* mValidK */ 512 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(0) +, /* mActType */ gemmGatedAct::ActType(3) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} @@ -42522,12 +48405,12 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 , /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 216600, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 768, "630f4be5c4efaa7813b64baf8992f19e6d6bc128c92be4f8e743b660a7380aa9", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 196248, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f", 896, "df270919a01b49ffcfdd9ba746dfbf811f2ce4f18e450c61b36ba2a7f9bd6eff", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -42543,7 +48426,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(3) +, /* mEltwiseActType */ gemm::EltwiseActType(0) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 @@ -42571,7 +48454,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) , /* mMmaM */ 256 -, /* mMmaN */ 32 +, /* mMmaN */ 64 , /* mMmaTileK */ 512 , /* mMockAllReduce */ 0 , /* mN */ 256 @@ -42580,17 +48463,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySfLdsSttm */ 0 , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 -, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 48 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 5 -, /* mNumStagesB */ 5 -, /* mNumStagesMma */ 1 +, /* mNumStagesA */ 4 +, /* mNumStagesB */ 4 +, /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 -, /* mNumStagesMmaAcrossWorkTile */ 1 -, /* mNumStagesSmemSfA */ 5 -, /* mNumStagesSmemSfB */ 5 -, /* mNumStagesTmemSfA */ 5 -, /* mNumStagesTmemSfB */ 5 +, /* mNumStagesMmaAcrossWorkTile */ 2 +, /* mNumStagesSmemSfA */ 4 +, /* mNumStagesSmemSfB */ 4 +, /* mNumStagesTmemSfA */ 4 +, /* mNumStagesTmemSfB */ 4 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -42607,8 +48490,8 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSplitK */ gemm::SplitK(0) , /* mTileK */ 512 , /* mTileM */ 128 -, /* mTileN */ 32 -, /* mTileScheduler */ gemm::TileScheduler(0) +, /* mTileN */ 64 +, /* mTileScheduler */ gemm::TileScheduler(1) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 , /* mUseCustomMmaSchedule */ 1 @@ -42627,13 +48510,13 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mValidN */ 256 , /* mValidK */ 512 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(3) +, /* mActType */ gemmGatedAct::ActType(1) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 0 +, /* mFusedAct */ 1 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 @@ -42653,7 +48536,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 216984, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f", 896, "4de88bd855155f2357e30c64a5ec6367dd1ef7fcac71a3b6755e0ff5eb698a30", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 196248, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f", 896, "229f597ff788b40c75f4205a163aa44b699a54ca72b236498f7514bfe5360c90", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -42689,7 +48572,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 1024 +, /* mK */ 512 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) @@ -42697,7 +48580,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) , /* mMmaM */ 256 -, /* mMmaN */ 32 +, /* mMmaN */ 64 , /* mMmaTileK */ 512 , /* mMockAllReduce */ 0 , /* mN */ 256 @@ -42706,17 +48589,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySfLdsSttm */ 0 , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 -, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 48 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 5 -, /* mNumStagesB */ 5 +, /* mNumStagesA */ 4 +, /* mNumStagesB */ 4 , /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 2 -, /* mNumStagesSmemSfA */ 5 -, /* mNumStagesSmemSfB */ 5 -, /* mNumStagesTmemSfA */ 5 -, /* mNumStagesTmemSfB */ 5 +, /* mNumStagesSmemSfA */ 4 +, /* mNumStagesSmemSfB */ 4 +, /* mNumStagesTmemSfA */ 4 +, /* mNumStagesTmemSfB */ 4 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -42733,7 +48616,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSplitK */ gemm::SplitK(0) , /* mTileK */ 512 , /* mTileM */ 128 -, /* mTileN */ 32 +, /* mTileN */ 64 , /* mTileScheduler */ gemm::TileScheduler(1) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 @@ -42748,12 +48631,12 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseTmaStore */ 1 , /* mUseTwoTmaLoadWarps */ 1 , /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 1 +, /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 1024 +, /* mValidK */ 512 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(1) +, /* mActType */ gemmGatedAct::ActType(2) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} @@ -42779,7 +48662,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 216984, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 896, "77a54feeb2e3e1e78e5d21e0f9a38620e8930ed96322e7c641098fd3969dc50c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 196248, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 896, "f430a2924034324777e5789c71167efde1ac53027ded578be3155e525fecbc7a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -42815,7 +48698,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 1024 +, /* mK */ 512 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) @@ -42823,7 +48706,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) , /* mMmaM */ 256 -, /* mMmaN */ 32 +, /* mMmaN */ 64 , /* mMmaTileK */ 512 , /* mMockAllReduce */ 0 , /* mN */ 256 @@ -42832,17 +48715,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySfLdsSttm */ 0 , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 -, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 48 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 5 -, /* mNumStagesB */ 5 +, /* mNumStagesA */ 4 +, /* mNumStagesB */ 4 , /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 2 -, /* mNumStagesSmemSfA */ 5 -, /* mNumStagesSmemSfB */ 5 -, /* mNumStagesTmemSfA */ 5 -, /* mNumStagesTmemSfB */ 5 +, /* mNumStagesSmemSfA */ 4 +, /* mNumStagesSmemSfB */ 4 +, /* mNumStagesTmemSfA */ 4 +, /* mNumStagesTmemSfB */ 4 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -42859,7 +48742,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSplitK */ gemm::SplitK(0) , /* mTileK */ 512 , /* mTileM */ 128 -, /* mTileN */ 32 +, /* mTileN */ 64 , /* mTileScheduler */ gemm::TileScheduler(1) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 @@ -42874,10 +48757,10 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseTmaStore */ 1 , /* mUseTwoTmaLoadWarps */ 1 , /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 1 +, /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 1024 +, /* mValidK */ 512 , /* mWorldSize */ 1 , /* mActType */ gemmGatedAct::ActType(0) , /* mClampBeforeAct */ 1 @@ -42905,7 +48788,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 216984, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 896, "6dd592dcae0c2e8fcad78ddca9ec04879ce787be1167debe2bdbb8657f1d7dc9", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 196248, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 896, "cd7955042082eedb04af18579bfd97124701eb36d7bc6cfdf210542a30d6031d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -42941,7 +48824,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 1024 +, /* mK */ 512 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) @@ -42949,7 +48832,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) , /* mMmaM */ 256 -, /* mMmaN */ 32 +, /* mMmaN */ 64 , /* mMmaTileK */ 512 , /* mMockAllReduce */ 0 , /* mN */ 256 @@ -42958,17 +48841,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySfLdsSttm */ 0 , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 -, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 48 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 5 -, /* mNumStagesB */ 5 +, /* mNumStagesA */ 4 +, /* mNumStagesB */ 4 , /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 2 -, /* mNumStagesSmemSfA */ 5 -, /* mNumStagesSmemSfB */ 5 -, /* mNumStagesTmemSfA */ 5 -, /* mNumStagesTmemSfB */ 5 +, /* mNumStagesSmemSfA */ 4 +, /* mNumStagesSmemSfB */ 4 +, /* mNumStagesTmemSfA */ 4 +, /* mNumStagesTmemSfB */ 4 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -42985,7 +48868,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSplitK */ gemm::SplitK(0) , /* mTileK */ 512 , /* mTileM */ 128 -, /* mTileN */ 32 +, /* mTileN */ 64 , /* mTileScheduler */ gemm::TileScheduler(1) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 @@ -43000,12 +48883,12 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseTmaStore */ 1 , /* mUseTwoTmaLoadWarps */ 1 , /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 1 +, /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 1024 +, /* mValidK */ 512 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(0) +, /* mActType */ gemmGatedAct::ActType(3) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} @@ -43031,7 +48914,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 216984, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 896, "3dda856f910bce58b174bfe891f8cfe00826ec9930427ac03c3a72ecb8b3ecbe", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 196248, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 896, "461059abc36c595f248c1408083f0f1171d87616d85ff5c6dba69adbbd0ded80", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -43067,7 +48950,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 1024 +, /* mK */ 512 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) @@ -43075,7 +48958,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) , /* mMmaM */ 256 -, /* mMmaN */ 32 +, /* mMmaN */ 64 , /* mMmaTileK */ 512 , /* mMockAllReduce */ 0 , /* mN */ 256 @@ -43084,17 +48967,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySfLdsSttm */ 0 , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 -, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 48 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 5 -, /* mNumStagesB */ 5 +, /* mNumStagesA */ 4 +, /* mNumStagesB */ 4 , /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 2 -, /* mNumStagesSmemSfA */ 5 -, /* mNumStagesSmemSfB */ 5 -, /* mNumStagesTmemSfA */ 5 -, /* mNumStagesTmemSfB */ 5 +, /* mNumStagesSmemSfA */ 4 +, /* mNumStagesSmemSfB */ 4 +, /* mNumStagesTmemSfA */ 4 +, /* mNumStagesTmemSfB */ 4 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -43111,7 +48994,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSplitK */ gemm::SplitK(0) , /* mTileK */ 512 , /* mTileM */ 128 -, /* mTileN */ 32 +, /* mTileN */ 64 , /* mTileScheduler */ gemm::TileScheduler(1) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 @@ -43126,10 +49009,10 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseTmaStore */ 1 , /* mUseTwoTmaLoadWarps */ 1 , /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 1 +, /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 1024 +, /* mValidK */ 512 , /* mWorldSize */ 1 , /* mActType */ gemmGatedAct::ActType(3) , /* mClampBeforeAct */ 1 @@ -43157,7 +49040,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 216600, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f", 768, "4fe840d8ed8840ea59b230b5cc72e019daef839cb13b02abcbc85ee47bcb9754", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 196248, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f", 896, "051614bbc3a9d4c1a118f1349abf1092b13037ec4d6f416ebb82577e9861ecf5", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -43201,7 +49084,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) , /* mMmaM */ 256 -, /* mMmaN */ 32 +, /* mMmaN */ 64 , /* mMmaTileK */ 512 , /* mMockAllReduce */ 0 , /* mN */ 256 @@ -43210,17 +49093,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySfLdsSttm */ 0 , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 -, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 48 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 5 -, /* mNumStagesB */ 5 -, /* mNumStagesMma */ 1 +, /* mNumStagesA */ 4 +, /* mNumStagesB */ 4 +, /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 -, /* mNumStagesMmaAcrossWorkTile */ 1 -, /* mNumStagesSmemSfA */ 5 -, /* mNumStagesSmemSfB */ 5 -, /* mNumStagesTmemSfA */ 5 -, /* mNumStagesTmemSfB */ 5 +, /* mNumStagesMmaAcrossWorkTile */ 2 +, /* mNumStagesSmemSfA */ 4 +, /* mNumStagesSmemSfB */ 4 +, /* mNumStagesTmemSfA */ 4 +, /* mNumStagesTmemSfB */ 4 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -43237,8 +49120,8 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSplitK */ gemm::SplitK(0) , /* mTileK */ 512 , /* mTileM */ 128 -, /* mTileN */ 32 -, /* mTileScheduler */ gemm::TileScheduler(0) +, /* mTileN */ 64 +, /* mTileScheduler */ gemm::TileScheduler(1) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 , /* mUseCustomMmaSchedule */ 1 @@ -43283,7 +49166,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 216600, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 768, "8c9cb5052e1810977bb09cbf8490a9bbd523f3c9d594c391e7cbd6b21668cd46", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 196248, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f", 896, "85e11dc78d9a3281805d5fa761aacdffa6d65c75eaa9f901e9620455a4b1464f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -43327,7 +49210,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) , /* mMmaM */ 256 -, /* mMmaN */ 32 +, /* mMmaN */ 64 , /* mMmaTileK */ 512 , /* mMockAllReduce */ 0 , /* mN */ 256 @@ -43336,17 +49219,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySfLdsSttm */ 0 , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 -, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 48 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 5 -, /* mNumStagesB */ 5 -, /* mNumStagesMma */ 1 +, /* mNumStagesA */ 4 +, /* mNumStagesB */ 4 +, /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 -, /* mNumStagesMmaAcrossWorkTile */ 1 -, /* mNumStagesSmemSfA */ 5 -, /* mNumStagesSmemSfB */ 5 -, /* mNumStagesTmemSfA */ 5 -, /* mNumStagesTmemSfB */ 5 +, /* mNumStagesMmaAcrossWorkTile */ 2 +, /* mNumStagesSmemSfA */ 4 +, /* mNumStagesSmemSfB */ 4 +, /* mNumStagesTmemSfA */ 4 +, /* mNumStagesTmemSfB */ 4 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -43363,8 +49246,8 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSplitK */ gemm::SplitK(0) , /* mTileK */ 512 , /* mTileM */ 128 -, /* mTileN */ 32 -, /* mTileScheduler */ gemm::TileScheduler(0) +, /* mTileN */ 64 +, /* mTileScheduler */ gemm::TileScheduler(1) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 , /* mUseCustomMmaSchedule */ 1 @@ -43383,7 +49266,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mValidN */ 256 , /* mValidK */ 1024 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(0) +, /* mActType */ gemmGatedAct::ActType(2) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} @@ -43409,7 +49292,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 216600, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 768, "8657274862f83ddf1c031bd3767342e05d357f15cbe9579097aae221369f4b7d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 196248, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 896, "62845a0e45537ef6d9a6812daaeb0738eb75a0285102aafe40a4f2392f26f8f1", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -43425,7 +49308,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(2) +, /* mEltwiseActType */ gemm::EltwiseActType(0) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 @@ -43453,7 +49336,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) , /* mMmaM */ 256 -, /* mMmaN */ 32 +, /* mMmaN */ 64 , /* mMmaTileK */ 512 , /* mMockAllReduce */ 0 , /* mN */ 256 @@ -43462,17 +49345,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySfLdsSttm */ 0 , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 -, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 48 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 5 -, /* mNumStagesB */ 5 -, /* mNumStagesMma */ 1 +, /* mNumStagesA */ 4 +, /* mNumStagesB */ 4 +, /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 -, /* mNumStagesMmaAcrossWorkTile */ 1 -, /* mNumStagesSmemSfA */ 5 -, /* mNumStagesSmemSfB */ 5 -, /* mNumStagesTmemSfA */ 5 -, /* mNumStagesTmemSfB */ 5 +, /* mNumStagesMmaAcrossWorkTile */ 2 +, /* mNumStagesSmemSfA */ 4 +, /* mNumStagesSmemSfB */ 4 +, /* mNumStagesTmemSfA */ 4 +, /* mNumStagesTmemSfB */ 4 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -43489,8 +49372,8 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSplitK */ gemm::SplitK(0) , /* mTileK */ 512 , /* mTileM */ 128 -, /* mTileN */ 32 -, /* mTileScheduler */ gemm::TileScheduler(0) +, /* mTileN */ 64 +, /* mTileScheduler */ gemm::TileScheduler(1) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 , /* mUseCustomMmaSchedule */ 1 @@ -43515,7 +49398,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 0 +, /* mFusedAct */ 1 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 @@ -43535,7 +49418,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 216600, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 768, "0cac5280ee597bfccc719993c7f3a04aa1c3306db481922d22f141a2fda81791", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 196248, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 896, "f64ca1f42100912114ef1783eebcbeae022011532de3cb62db5f036d17391b7b", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -43551,7 +49434,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(3) +, /* mEltwiseActType */ gemm::EltwiseActType(2) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 @@ -43579,7 +49462,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) , /* mMmaM */ 256 -, /* mMmaN */ 32 +, /* mMmaN */ 64 , /* mMmaTileK */ 512 , /* mMockAllReduce */ 0 , /* mN */ 256 @@ -43588,17 +49471,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySfLdsSttm */ 0 , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 -, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 48 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 5 -, /* mNumStagesB */ 5 -, /* mNumStagesMma */ 1 +, /* mNumStagesA */ 4 +, /* mNumStagesB */ 4 +, /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 -, /* mNumStagesMmaAcrossWorkTile */ 1 -, /* mNumStagesSmemSfA */ 5 -, /* mNumStagesSmemSfB */ 5 -, /* mNumStagesTmemSfA */ 5 -, /* mNumStagesTmemSfB */ 5 +, /* mNumStagesMmaAcrossWorkTile */ 2 +, /* mNumStagesSmemSfA */ 4 +, /* mNumStagesSmemSfB */ 4 +, /* mNumStagesTmemSfA */ 4 +, /* mNumStagesTmemSfB */ 4 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -43615,8 +49498,8 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSplitK */ gemm::SplitK(0) , /* mTileK */ 512 , /* mTileM */ 128 -, /* mTileN */ 32 -, /* mTileScheduler */ gemm::TileScheduler(0) +, /* mTileN */ 64 +, /* mTileScheduler */ gemm::TileScheduler(1) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 , /* mUseCustomMmaSchedule */ 1 @@ -43661,7 +49544,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 151448, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f", 640, "49ed9329080255d7d7dfd8926e40ce55102b05eff1ec4a302cba798261c1e52a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 196248, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 896, "e0682b36440a3beb6a6abd2d10ffda118160bdfa178bd1a616f7545278c591d0", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -43677,13 +49560,13 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEltwiseActType */ gemm::EltwiseActType(3) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 , /* mEpilogueLdtmBits */ 256 , /* mEpilogueTileM */ 128 -, /* mEpilogueTileN */ 64 +, /* mEpilogueTileN */ 32 , /* mFallbackClusterDimX */ 1 , /* mFallbackClusterDimY */ 1 , /* mFallbackClusterDimZ */ 1 @@ -43697,7 +49580,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 256 +, /* mK */ 1024 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) @@ -43706,7 +49589,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mMmaKind */ trtllm::gen::MmaKind(4) , /* mMmaM */ 256 , /* mMmaN */ 64 -, /* mMmaTileK */ 256 +, /* mMmaTileK */ 512 , /* mMockAllReduce */ 0 , /* mN */ 256 , /* mNumEpilogueWarps */ 4 @@ -43716,15 +49599,15 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 48 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 6 -, /* mNumStagesB */ 6 +, /* mNumStagesA */ 4 +, /* mNumStagesB */ 4 , /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 2 -, /* mNumStagesSmemSfA */ 6 -, /* mNumStagesSmemSfB */ 6 -, /* mNumStagesTmemSfA */ 6 -, /* mNumStagesTmemSfB */ 6 +, /* mNumStagesSmemSfA */ 4 +, /* mNumStagesSmemSfB */ 4 +, /* mNumStagesTmemSfA */ 4 +, /* mNumStagesTmemSfB */ 4 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -43739,7 +49622,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) , /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 256 +, /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 64 , /* mTileScheduler */ gemm::TileScheduler(1) @@ -43756,18 +49639,18 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseTmaStore */ 1 , /* mUseTwoTmaLoadWarps */ 1 , /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 0 +, /* mUseUnrollLoop2xForMma */ 1 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 256 +, /* mValidK */ 1024 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(1) +, /* mActType */ gemmGatedAct::ActType(3) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 1 +, /* mFusedAct */ 0 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 @@ -43782,17 +49665,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 , /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 151448, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "eca147c07d5ce270e57cf83db557d6bc4e384876a776dd025d9af52caaa1c17e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 183440, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f", 512, "95ac047a99bc456d914ce52e9549ed4b7ec58646ff3073174178ab2559aedbda", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 -, /* mClusterDimX */ 2 +, /* mClusterDimX */ 1 , /* mClusterDimY */ 1 , /* mClusterDimZ */ 1 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) @@ -43809,7 +49692,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mEpilogueLdtmDps */ 16 , /* mEpilogueLdtmBits */ 256 , /* mEpilogueTileM */ 128 -, /* mEpilogueTileN */ 64 +, /* mEpilogueTileN */ 8 , /* mFallbackClusterDimX */ 1 , /* mFallbackClusterDimY */ 1 , /* mFallbackClusterDimZ */ 1 @@ -43830,8 +49713,8 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mM */ 256 , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) -, /* mMmaM */ 256 -, /* mMmaN */ 64 +, /* mMmaM */ 128 +, /* mMmaN */ 8 , /* mMmaTileK */ 256 , /* mMockAllReduce */ 0 , /* mN */ 256 @@ -43840,17 +49723,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySfLdsSttm */ 0 , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 -, /* mNumRegsPerThreadNonEpilogueWarp */ 48 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 6 -, /* mNumStagesB */ 6 +, /* mNumStagesA */ 9 +, /* mNumStagesB */ 9 , /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 2 -, /* mNumStagesSmemSfA */ 6 -, /* mNumStagesSmemSfB */ 6 -, /* mNumStagesTmemSfA */ 6 -, /* mNumStagesTmemSfB */ 6 +, /* mNumStagesSmemSfA */ 9 +, /* mNumStagesSmemSfB */ 9 +, /* mNumStagesTmemSfA */ 9 +, /* mNumStagesTmemSfB */ 9 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -43867,7 +49750,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSplitK */ gemm::SplitK(0) , /* mTileK */ 256 , /* mTileM */ 128 -, /* mTileN */ 64 +, /* mTileN */ 8 , /* mTileScheduler */ gemm::TileScheduler(1) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 @@ -43887,7 +49770,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mValidN */ 256 , /* mValidK */ 256 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(0) +, /* mActType */ gemmGatedAct::ActType(1) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} @@ -43913,12 +49796,12 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 151448, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 640, "a811661feae910a1198dc330f1ecf549bc38f89d74017163776cfd21ca2fbede", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 183440, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "aee4ee176a04807b82ca840470e804d2782cce0db1ff3326bcfcdbb66767b695", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 -, /* mClusterDimX */ 2 +, /* mClusterDimX */ 1 , /* mClusterDimY */ 1 , /* mClusterDimZ */ 1 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) @@ -43929,13 +49812,13 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(2) +, /* mEltwiseActType */ gemm::EltwiseActType(0) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 , /* mEpilogueLdtmBits */ 256 , /* mEpilogueTileM */ 128 -, /* mEpilogueTileN */ 64 +, /* mEpilogueTileN */ 8 , /* mFallbackClusterDimX */ 1 , /* mFallbackClusterDimY */ 1 , /* mFallbackClusterDimZ */ 1 @@ -43956,8 +49839,8 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mM */ 256 , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) -, /* mMmaM */ 256 -, /* mMmaN */ 64 +, /* mMmaM */ 128 +, /* mMmaN */ 8 , /* mMmaTileK */ 256 , /* mMockAllReduce */ 0 , /* mN */ 256 @@ -43966,17 +49849,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySfLdsSttm */ 0 , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 -, /* mNumRegsPerThreadNonEpilogueWarp */ 48 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 6 -, /* mNumStagesB */ 6 +, /* mNumStagesA */ 9 +, /* mNumStagesB */ 9 , /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 2 -, /* mNumStagesSmemSfA */ 6 -, /* mNumStagesSmemSfB */ 6 -, /* mNumStagesTmemSfA */ 6 -, /* mNumStagesTmemSfB */ 6 +, /* mNumStagesSmemSfA */ 9 +, /* mNumStagesSmemSfB */ 9 +, /* mNumStagesTmemSfA */ 9 +, /* mNumStagesTmemSfB */ 9 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -43993,7 +49876,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSplitK */ gemm::SplitK(0) , /* mTileK */ 256 , /* mTileM */ 128 -, /* mTileN */ 64 +, /* mTileN */ 8 , /* mTileScheduler */ gemm::TileScheduler(1) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 @@ -44013,13 +49896,13 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mValidN */ 256 , /* mValidK */ 256 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(3) +, /* mActType */ gemmGatedAct::ActType(2) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 0 +, /* mFusedAct */ 1 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 @@ -44039,12 +49922,12 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 151448, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 640, "afa1068f89495f40462f9cd960a99ad7acda322dd66632c02a1c5a1fb0453333", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 183440, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "2c8c4e6be5720166607c450616cdbfae6c0ab87c119d7da59917997126f9e075", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 -, /* mClusterDimX */ 2 +, /* mClusterDimX */ 1 , /* mClusterDimY */ 1 , /* mClusterDimZ */ 1 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) @@ -44055,13 +49938,13 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(3) +, /* mEltwiseActType */ gemm::EltwiseActType(0) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 , /* mEpilogueLdtmBits */ 256 , /* mEpilogueTileM */ 128 -, /* mEpilogueTileN */ 64 +, /* mEpilogueTileN */ 8 , /* mFallbackClusterDimX */ 1 , /* mFallbackClusterDimY */ 1 , /* mFallbackClusterDimZ */ 1 @@ -44082,8 +49965,8 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mM */ 256 , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) -, /* mMmaM */ 256 -, /* mMmaN */ 64 +, /* mMmaM */ 128 +, /* mMmaN */ 8 , /* mMmaTileK */ 256 , /* mMockAllReduce */ 0 , /* mN */ 256 @@ -44092,17 +49975,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySfLdsSttm */ 0 , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 -, /* mNumRegsPerThreadNonEpilogueWarp */ 48 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 6 -, /* mNumStagesB */ 6 +, /* mNumStagesA */ 9 +, /* mNumStagesB */ 9 , /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 2 -, /* mNumStagesSmemSfA */ 6 -, /* mNumStagesSmemSfB */ 6 -, /* mNumStagesTmemSfA */ 6 -, /* mNumStagesTmemSfB */ 6 +, /* mNumStagesSmemSfA */ 9 +, /* mNumStagesSmemSfB */ 9 +, /* mNumStagesTmemSfA */ 9 +, /* mNumStagesTmemSfB */ 9 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -44119,7 +50002,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSplitK */ gemm::SplitK(0) , /* mTileK */ 256 , /* mTileM */ 128 -, /* mTileN */ 64 +, /* mTileN */ 8 , /* mTileScheduler */ gemm::TileScheduler(1) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 @@ -44139,133 +50022,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mValidN */ 256 , /* mValidK */ 256 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(3) -, /* mClampBeforeAct */ 1 -, /* mBatchedM */ {} -, /* mBatchedN */ {} -, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) -, /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 0 -, /* mGridWaitForPrimaryRouting */ 1 -, /* mIsStaticBatch */ 0 -, /* mIsUniformNumTokensPerBatch */ 0 -, /* mNumBatches */ 128 -, /* mNumRegsPerThreadLoadA */ 0 -, /* mNumRegsPerThreadLoadB */ 0 -, /* mNumRegsPerThreadLoadSfA */ 0 -, /* mNumRegsPerThreadLoadSfB */ 0 -, /* mNumTokens */ 2 -, /* mNumWarpsLoadA */ 0 -, /* mNumWarpsLoadB */ 0 -, /* mNumWarpsLoadSfA */ 0 -, /* mNumWarpsLoadSfB */ 0 -, /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} -, /* mPrefetchB */ 0 -, /* mUseCMultiCast */ 0 -, /* mUseTmaOobOpt */ 1 - }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 151448, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f", 640, "3fe9134ec5299e8516ab1544ade686faf7887f62a23a8ffe1bde57eaa6cd9855", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) -, /* mBiasDtype */ trtllm::gen::Dtype(1056776) -, /* mBiasType */ gemm::BiasType(1) -, /* mBlockK */ -1 -, /* mClcFastDrain */ 1 -, /* mClusterDimX */ 2 -, /* mClusterDimY */ 1 -, /* mClusterDimZ */ 1 -, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) -, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) -, /* mDtypeA */ trtllm::gen::Dtype(17826818) -, /* mDtypeB */ trtllm::gen::Dtype(17826818) -, /* mDtypeC */ trtllm::gen::Dtype(17826818) -, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) -, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) -, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(0) -, /* mEnablesEarlyExit */ 1 -, /* mEnablesDelayedEarlyExit */ 0 -, /* mEpilogueLdtmDps */ 16 -, /* mEpilogueLdtmBits */ 256 -, /* mEpilogueTileM */ 128 -, /* mEpilogueTileN */ 64 -, /* mFallbackClusterDimX */ 1 -, /* mFallbackClusterDimY */ 1 -, /* mFallbackClusterDimZ */ 1 -, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) -, /* mFuseLoadSfTask */ 0 -, /* mFuseUtccpWithUtcmma */ 0 -, /* mGridTriggerSecondaryA */ 0 -, /* mGridTriggerSecondaryB */ 1 -, /* mGridWaitForPrimaryEarlyExit */ 1 -, /* mGridWaitForPrimaryA */ 0 -, /* mGridWaitForPrimaryB */ 1 -, /* mHoistLoadTaskInit */ 1 -, /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 512 -, /* mKernelTraits */ {} -, /* mLayoutA */ gemm::MatrixLayout(0) -, /* mLayoutB */ gemm::MatrixLayout(0) -, /* mM */ 256 -, /* mMmaK */ 64 -, /* mMmaKind */ trtllm::gen::MmaKind(4) -, /* mMmaM */ 256 -, /* mMmaN */ 64 -, /* mMmaTileK */ 256 -, /* mMockAllReduce */ 0 -, /* mN */ 256 -, /* mNumEpilogueWarps */ 4 -, /* mNumRegsCastAWarps */ 0 -, /* mNumRegsCopySfLdsSttm */ 0 -, /* mNumRegsCopySparsityInfo */ 0 -, /* mNumRegsPerThreadEpilogueWarp */ 0 -, /* mNumRegsPerThreadNonEpilogueWarp */ 48 -, /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 6 -, /* mNumStagesB */ 6 -, /* mNumStagesMma */ 2 -, /* mNumStagesMmaWithinWorkTile */ 1 -, /* mNumStagesMmaAcrossWorkTile */ 2 -, /* mNumStagesSmemSfA */ 6 -, /* mNumStagesSmemSfB */ 6 -, /* mNumStagesTmemSfA */ 6 -, /* mNumStagesTmemSfB */ 6 -, /* mNumStagesWorkId */ 3 -, /* mOutputDebugTensors */ 0 -, /* mPatchF2fp */ 0 -, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) -, /* mSchedHostTask */ gemm::SchedHostTask(0) -, /* mSfBlockSizeA */ 16 -, /* mSfBlockSizeB */ 16 -, /* mSfBlockSizeC */ 16 -, /* mSfLayoutA */ trtllm::gen::SfLayout(3) -, /* mSfLayoutB */ trtllm::gen::SfLayout(0) -, /* mSfLayoutC */ trtllm::gen::SfLayout(1) -, /* mSfReshapeFactor */ 1 -, /* mSparsityA */ trtllm::gen::Sparsity(0) -, /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 256 -, /* mTileM */ 128 -, /* mTileN */ 64 -, /* mTileScheduler */ gemm::TileScheduler(1) -, /* mTransposeMmaOutput */ 1 -, /* mUseCustomizedMma3xNvFp4 */ 0 -, /* mUseCustomMmaSchedule */ 1 -, /* mUseDeepSeekFp8 */ 0 -, /* mUseFlexibleClusterDims */ 0 -, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 -, /* mUseMaxTmemOverlap */ 0 -, /* mUsePerTokenSfA */ 0 -, /* mUsePerTokenSfB */ 0 -, /* mUseShuffledMatrix */ 1 -, /* mUseTmaStore */ 1 -, /* mUseTwoTmaLoadWarps */ 1 -, /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 1 -, /* mValidM */ 256 -, /* mValidN */ 256 -, /* mValidK */ 512 -, /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(1) +, /* mActType */ gemmGatedAct::ActType(0) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} @@ -44291,12 +50048,12 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 151448, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "15b02b5ff92bbdf7f5668e58929cb75676e68dbbca63b30f646bdfd57c786cf9", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 183440, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 512, "da74c5b5ca59015731f123e9c9ec7a2995c8bcfadf178f405201ca6512cf8852", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 -, /* mClusterDimX */ 2 +, /* mClusterDimX */ 1 , /* mClusterDimY */ 1 , /* mClusterDimZ */ 1 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) @@ -44307,13 +50064,13 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEltwiseActType */ gemm::EltwiseActType(2) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 , /* mEpilogueLdtmBits */ 256 , /* mEpilogueTileM */ 128 -, /* mEpilogueTileN */ 64 +, /* mEpilogueTileN */ 8 , /* mFallbackClusterDimX */ 1 , /* mFallbackClusterDimY */ 1 , /* mFallbackClusterDimZ */ 1 @@ -44327,15 +50084,15 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 512 +, /* mK */ 256 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) , /* mM */ 256 , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) -, /* mMmaM */ 256 -, /* mMmaN */ 64 +, /* mMmaM */ 128 +, /* mMmaN */ 8 , /* mMmaTileK */ 256 , /* mMockAllReduce */ 0 , /* mN */ 256 @@ -44344,17 +50101,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySfLdsSttm */ 0 , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 -, /* mNumRegsPerThreadNonEpilogueWarp */ 48 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 6 -, /* mNumStagesB */ 6 +, /* mNumStagesA */ 9 +, /* mNumStagesB */ 9 , /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 2 -, /* mNumStagesSmemSfA */ 6 -, /* mNumStagesSmemSfB */ 6 -, /* mNumStagesTmemSfA */ 6 -, /* mNumStagesTmemSfB */ 6 +, /* mNumStagesSmemSfA */ 9 +, /* mNumStagesSmemSfB */ 9 +, /* mNumStagesTmemSfA */ 9 +, /* mNumStagesTmemSfB */ 9 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -44371,7 +50128,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSplitK */ gemm::SplitK(0) , /* mTileK */ 256 , /* mTileM */ 128 -, /* mTileN */ 64 +, /* mTileN */ 8 , /* mTileScheduler */ gemm::TileScheduler(1) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 @@ -44386,10 +50143,10 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseTmaStore */ 1 , /* mUseTwoTmaLoadWarps */ 1 , /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 1 +, /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 512 +, /* mValidK */ 256 , /* mWorldSize */ 1 , /* mActType */ gemmGatedAct::ActType(0) , /* mClampBeforeAct */ 1 @@ -44397,7 +50154,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 1 +, /* mFusedAct */ 0 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 @@ -44417,12 +50174,12 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 151448, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 640, "18dceee0abf90ec7b76dfe61f7e0bc7dad09ccd398853119a3f2553754ed4dbc", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 183440, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 512, "0ee6d314493dc77763ab87dd1bb9d8bbb62680d7454507dfd3cdea4186ea3bf4", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 -, /* mClusterDimX */ 2 +, /* mClusterDimX */ 1 , /* mClusterDimY */ 1 , /* mClusterDimZ */ 1 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) @@ -44433,13 +50190,13 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(2) +, /* mEltwiseActType */ gemm::EltwiseActType(3) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 , /* mEpilogueLdtmBits */ 256 , /* mEpilogueTileM */ 128 -, /* mEpilogueTileN */ 64 +, /* mEpilogueTileN */ 8 , /* mFallbackClusterDimX */ 1 , /* mFallbackClusterDimY */ 1 , /* mFallbackClusterDimZ */ 1 @@ -44453,15 +50210,15 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 512 +, /* mK */ 256 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) , /* mM */ 256 , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) -, /* mMmaM */ 256 -, /* mMmaN */ 64 +, /* mMmaM */ 128 +, /* mMmaN */ 8 , /* mMmaTileK */ 256 , /* mMockAllReduce */ 0 , /* mN */ 256 @@ -44470,17 +50227,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySfLdsSttm */ 0 , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 -, /* mNumRegsPerThreadNonEpilogueWarp */ 48 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 6 -, /* mNumStagesB */ 6 +, /* mNumStagesA */ 9 +, /* mNumStagesB */ 9 , /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 2 -, /* mNumStagesSmemSfA */ 6 -, /* mNumStagesSmemSfB */ 6 -, /* mNumStagesTmemSfA */ 6 -, /* mNumStagesTmemSfB */ 6 +, /* mNumStagesSmemSfA */ 9 +, /* mNumStagesSmemSfB */ 9 +, /* mNumStagesTmemSfA */ 9 +, /* mNumStagesTmemSfB */ 9 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -44497,7 +50254,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSplitK */ gemm::SplitK(0) , /* mTileK */ 256 , /* mTileM */ 128 -, /* mTileN */ 64 +, /* mTileN */ 8 , /* mTileScheduler */ gemm::TileScheduler(1) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 @@ -44512,10 +50269,10 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseTmaStore */ 1 , /* mUseTwoTmaLoadWarps */ 1 , /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 1 +, /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 512 +, /* mValidK */ 256 , /* mWorldSize */ 1 , /* mActType */ gemmGatedAct::ActType(3) , /* mClampBeforeAct */ 1 @@ -44543,12 +50300,12 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 151448, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 640, "1fffb302b9c73c6ef33ad9eede6b63830ff3384de515b55a29c1dc232d9916a7", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 183312, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f", 512, "96cb72309a1d3916a96c168f646d3c1c5d843b699703f88844c6118c06a9df49", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 -, /* mClusterDimX */ 2 +, /* mClusterDimX */ 1 , /* mClusterDimY */ 1 , /* mClusterDimZ */ 1 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) @@ -44559,13 +50316,13 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(3) +, /* mEltwiseActType */ gemm::EltwiseActType(0) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 , /* mEpilogueLdtmBits */ 256 , /* mEpilogueTileM */ 128 -, /* mEpilogueTileN */ 64 +, /* mEpilogueTileN */ 8 , /* mFallbackClusterDimX */ 1 , /* mFallbackClusterDimY */ 1 , /* mFallbackClusterDimZ */ 1 @@ -44579,15 +50336,15 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 512 +, /* mK */ 256 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) , /* mM */ 256 , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) -, /* mMmaM */ 256 -, /* mMmaN */ 64 +, /* mMmaM */ 128 +, /* mMmaN */ 8 , /* mMmaTileK */ 256 , /* mMockAllReduce */ 0 , /* mN */ 256 @@ -44596,17 +50353,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySfLdsSttm */ 0 , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 -, /* mNumRegsPerThreadNonEpilogueWarp */ 48 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 6 -, /* mNumStagesB */ 6 -, /* mNumStagesMma */ 2 +, /* mNumStagesA */ 9 +, /* mNumStagesB */ 9 +, /* mNumStagesMma */ 1 , /* mNumStagesMmaWithinWorkTile */ 1 -, /* mNumStagesMmaAcrossWorkTile */ 2 -, /* mNumStagesSmemSfA */ 6 -, /* mNumStagesSmemSfB */ 6 -, /* mNumStagesTmemSfA */ 6 -, /* mNumStagesTmemSfB */ 6 +, /* mNumStagesMmaAcrossWorkTile */ 1 +, /* mNumStagesSmemSfA */ 9 +, /* mNumStagesSmemSfB */ 9 +, /* mNumStagesTmemSfA */ 9 +, /* mNumStagesTmemSfB */ 9 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -44623,8 +50380,8 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSplitK */ gemm::SplitK(0) , /* mTileK */ 256 , /* mTileM */ 128 -, /* mTileN */ 64 -, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTileN */ 8 +, /* mTileScheduler */ gemm::TileScheduler(0) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 , /* mUseCustomMmaSchedule */ 1 @@ -44638,18 +50395,18 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseTmaStore */ 1 , /* mUseTwoTmaLoadWarps */ 1 , /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 1 +, /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 512 +, /* mValidK */ 256 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(3) +, /* mActType */ gemmGatedAct::ActType(1) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 0 +, /* mFusedAct */ 1 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 @@ -44669,12 +50426,12 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 196248, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f", 896, "a5ff5c5231d93ce9715272178b4699ab1085f82a2b7621d872343e0f3f626c98", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 183312, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "16aba6c094a121b132e94331bba2fa4474c8a4cdda5a8f931b139f1111cf2b61", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 -, /* mClusterDimX */ 2 +, /* mClusterDimX */ 1 , /* mClusterDimY */ 1 , /* mClusterDimZ */ 1 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) @@ -44691,7 +50448,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mEpilogueLdtmDps */ 16 , /* mEpilogueLdtmBits */ 256 , /* mEpilogueTileM */ 128 -, /* mEpilogueTileN */ 32 +, /* mEpilogueTileN */ 8 , /* mFallbackClusterDimX */ 1 , /* mFallbackClusterDimY */ 1 , /* mFallbackClusterDimZ */ 1 @@ -44705,16 +50462,16 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 512 +, /* mK */ 256 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) , /* mM */ 256 , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) -, /* mMmaM */ 256 -, /* mMmaN */ 64 -, /* mMmaTileK */ 512 +, /* mMmaM */ 128 +, /* mMmaN */ 8 +, /* mMmaTileK */ 256 , /* mMockAllReduce */ 0 , /* mN */ 256 , /* mNumEpilogueWarps */ 4 @@ -44722,17 +50479,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySfLdsSttm */ 0 , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 -, /* mNumRegsPerThreadNonEpilogueWarp */ 48 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 4 -, /* mNumStagesB */ 4 -, /* mNumStagesMma */ 2 +, /* mNumStagesA */ 9 +, /* mNumStagesB */ 9 +, /* mNumStagesMma */ 1 , /* mNumStagesMmaWithinWorkTile */ 1 -, /* mNumStagesMmaAcrossWorkTile */ 2 -, /* mNumStagesSmemSfA */ 4 -, /* mNumStagesSmemSfB */ 4 -, /* mNumStagesTmemSfA */ 4 -, /* mNumStagesTmemSfB */ 4 +, /* mNumStagesMmaAcrossWorkTile */ 1 +, /* mNumStagesSmemSfA */ 9 +, /* mNumStagesSmemSfB */ 9 +, /* mNumStagesTmemSfA */ 9 +, /* mNumStagesTmemSfB */ 9 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -44747,10 +50504,10 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) , /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 512 +, /* mTileK */ 256 , /* mTileM */ 128 -, /* mTileN */ 64 -, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTileN */ 8 +, /* mTileScheduler */ gemm::TileScheduler(0) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 , /* mUseCustomMmaSchedule */ 1 @@ -44767,9 +50524,9 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 512 +, /* mValidK */ 256 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(1) +, /* mActType */ gemmGatedAct::ActType(2) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} @@ -44790,17 +50547,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 , /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 196248, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 896, "a2895d97cd26211bb3714a1ad919a2fb87b4bebbb9eafce8edc3e6bb1621779a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 183312, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "7e4b0c1b68a582f7426b9be600a6c4f620e83007ad25e5cff72b30b427359cd8", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 -, /* mClusterDimX */ 2 +, /* mClusterDimX */ 1 , /* mClusterDimY */ 1 , /* mClusterDimZ */ 1 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) @@ -44817,7 +50574,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mEpilogueLdtmDps */ 16 , /* mEpilogueLdtmBits */ 256 , /* mEpilogueTileM */ 128 -, /* mEpilogueTileN */ 32 +, /* mEpilogueTileN */ 8 , /* mFallbackClusterDimX */ 1 , /* mFallbackClusterDimY */ 1 , /* mFallbackClusterDimZ */ 1 @@ -44831,16 +50588,16 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 512 +, /* mK */ 256 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) , /* mM */ 256 , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) -, /* mMmaM */ 256 -, /* mMmaN */ 64 -, /* mMmaTileK */ 512 +, /* mMmaM */ 128 +, /* mMmaN */ 8 +, /* mMmaTileK */ 256 , /* mMockAllReduce */ 0 , /* mN */ 256 , /* mNumEpilogueWarps */ 4 @@ -44848,17 +50605,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySfLdsSttm */ 0 , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 -, /* mNumRegsPerThreadNonEpilogueWarp */ 48 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 4 -, /* mNumStagesB */ 4 -, /* mNumStagesMma */ 2 +, /* mNumStagesA */ 9 +, /* mNumStagesB */ 9 +, /* mNumStagesMma */ 1 , /* mNumStagesMmaWithinWorkTile */ 1 -, /* mNumStagesMmaAcrossWorkTile */ 2 -, /* mNumStagesSmemSfA */ 4 -, /* mNumStagesSmemSfB */ 4 -, /* mNumStagesTmemSfA */ 4 -, /* mNumStagesTmemSfB */ 4 +, /* mNumStagesMmaAcrossWorkTile */ 1 +, /* mNumStagesSmemSfA */ 9 +, /* mNumStagesSmemSfB */ 9 +, /* mNumStagesTmemSfA */ 9 +, /* mNumStagesTmemSfB */ 9 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -44873,10 +50630,10 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) , /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 512 +, /* mTileK */ 256 , /* mTileM */ 128 -, /* mTileN */ 64 -, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTileN */ 8 +, /* mTileScheduler */ gemm::TileScheduler(0) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 , /* mUseCustomMmaSchedule */ 1 @@ -44893,7 +50650,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 512 +, /* mValidK */ 256 , /* mWorldSize */ 1 , /* mActType */ gemmGatedAct::ActType(0) , /* mClampBeforeAct */ 1 @@ -44916,17 +50673,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 , /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 196248, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 896, "2ca55f92d11ed8b10f429929dd1246e91e25589eae6a2fbc9d4c5e37b33ab4a4", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 183312, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 512, "091582053242d9f42805fa2d016e416d6663163c32623a7f987ad33bf8e189da", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 -, /* mClusterDimX */ 2 +, /* mClusterDimX */ 1 , /* mClusterDimY */ 1 , /* mClusterDimZ */ 1 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) @@ -44943,7 +50700,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mEpilogueLdtmDps */ 16 , /* mEpilogueLdtmBits */ 256 , /* mEpilogueTileM */ 128 -, /* mEpilogueTileN */ 32 +, /* mEpilogueTileN */ 8 , /* mFallbackClusterDimX */ 1 , /* mFallbackClusterDimY */ 1 , /* mFallbackClusterDimZ */ 1 @@ -44957,16 +50714,16 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 512 +, /* mK */ 256 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) , /* mM */ 256 , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) -, /* mMmaM */ 256 -, /* mMmaN */ 64 -, /* mMmaTileK */ 512 +, /* mMmaM */ 128 +, /* mMmaN */ 8 +, /* mMmaTileK */ 256 , /* mMockAllReduce */ 0 , /* mN */ 256 , /* mNumEpilogueWarps */ 4 @@ -44974,17 +50731,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySfLdsSttm */ 0 , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 -, /* mNumRegsPerThreadNonEpilogueWarp */ 48 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 4 -, /* mNumStagesB */ 4 -, /* mNumStagesMma */ 2 +, /* mNumStagesA */ 9 +, /* mNumStagesB */ 9 +, /* mNumStagesMma */ 1 , /* mNumStagesMmaWithinWorkTile */ 1 -, /* mNumStagesMmaAcrossWorkTile */ 2 -, /* mNumStagesSmemSfA */ 4 -, /* mNumStagesSmemSfB */ 4 -, /* mNumStagesTmemSfA */ 4 -, /* mNumStagesTmemSfB */ 4 +, /* mNumStagesMmaAcrossWorkTile */ 1 +, /* mNumStagesSmemSfA */ 9 +, /* mNumStagesSmemSfB */ 9 +, /* mNumStagesTmemSfA */ 9 +, /* mNumStagesTmemSfB */ 9 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -44999,10 +50756,10 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) , /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 512 +, /* mTileK */ 256 , /* mTileM */ 128 -, /* mTileN */ 64 -, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTileN */ 8 +, /* mTileScheduler */ gemm::TileScheduler(0) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 , /* mUseCustomMmaSchedule */ 1 @@ -45019,9 +50776,9 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 512 +, /* mValidK */ 256 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(3) +, /* mActType */ gemmGatedAct::ActType(0) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} @@ -45042,17 +50799,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 , /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 196248, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 896, "1870628d0c59ffe92c2dd22f42c52c87f14f4ad614fcc64d8be28c9b03268cc6", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 183312, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 512, "de185c2401fecc591797ea72ceb744d00f5370fb66d6ad87c0ec33bc2c53ff00", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 -, /* mClusterDimX */ 2 +, /* mClusterDimX */ 1 , /* mClusterDimY */ 1 , /* mClusterDimZ */ 1 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) @@ -45069,7 +50826,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mEpilogueLdtmDps */ 16 , /* mEpilogueLdtmBits */ 256 , /* mEpilogueTileM */ 128 -, /* mEpilogueTileN */ 32 +, /* mEpilogueTileN */ 8 , /* mFallbackClusterDimX */ 1 , /* mFallbackClusterDimY */ 1 , /* mFallbackClusterDimZ */ 1 @@ -45083,16 +50840,16 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 512 +, /* mK */ 256 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) , /* mM */ 256 , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) -, /* mMmaM */ 256 -, /* mMmaN */ 64 -, /* mMmaTileK */ 512 +, /* mMmaM */ 128 +, /* mMmaN */ 8 +, /* mMmaTileK */ 256 , /* mMockAllReduce */ 0 , /* mN */ 256 , /* mNumEpilogueWarps */ 4 @@ -45100,17 +50857,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySfLdsSttm */ 0 , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 -, /* mNumRegsPerThreadNonEpilogueWarp */ 48 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 4 -, /* mNumStagesB */ 4 -, /* mNumStagesMma */ 2 +, /* mNumStagesA */ 9 +, /* mNumStagesB */ 9 +, /* mNumStagesMma */ 1 , /* mNumStagesMmaWithinWorkTile */ 1 -, /* mNumStagesMmaAcrossWorkTile */ 2 -, /* mNumStagesSmemSfA */ 4 -, /* mNumStagesSmemSfB */ 4 -, /* mNumStagesTmemSfA */ 4 -, /* mNumStagesTmemSfB */ 4 +, /* mNumStagesMmaAcrossWorkTile */ 1 +, /* mNumStagesSmemSfA */ 9 +, /* mNumStagesSmemSfB */ 9 +, /* mNumStagesTmemSfA */ 9 +, /* mNumStagesTmemSfB */ 9 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -45125,10 +50882,10 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) , /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 512 +, /* mTileK */ 256 , /* mTileM */ 128 -, /* mTileN */ 64 -, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTileN */ 8 +, /* mTileScheduler */ gemm::TileScheduler(0) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 , /* mUseCustomMmaSchedule */ 1 @@ -45145,7 +50902,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 512 +, /* mValidK */ 256 , /* mWorldSize */ 1 , /* mActType */ gemmGatedAct::ActType(3) , /* mClampBeforeAct */ 1 @@ -45168,17 +50925,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 , /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 196248, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f", 896, "55b8297bccfdb00c698b21aad6e52ac1100e0e92d1dbfee1bc04fc43b9cf727d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 183440, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f", 512, "6dfa8b29e5341c45ea317a6a3a2c1284fd36478665f6a4fdc6a3f3a6c7ebbb0b", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 -, /* mClusterDimX */ 2 +, /* mClusterDimX */ 1 , /* mClusterDimY */ 1 , /* mClusterDimZ */ 1 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) @@ -45195,7 +50952,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mEpilogueLdtmDps */ 16 , /* mEpilogueLdtmBits */ 256 , /* mEpilogueTileM */ 128 -, /* mEpilogueTileN */ 32 +, /* mEpilogueTileN */ 8 , /* mFallbackClusterDimX */ 1 , /* mFallbackClusterDimY */ 1 , /* mFallbackClusterDimZ */ 1 @@ -45209,16 +50966,16 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 1024 +, /* mK */ 512 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) , /* mM */ 256 , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) -, /* mMmaM */ 256 -, /* mMmaN */ 64 -, /* mMmaTileK */ 512 +, /* mMmaM */ 128 +, /* mMmaN */ 8 +, /* mMmaTileK */ 256 , /* mMockAllReduce */ 0 , /* mN */ 256 , /* mNumEpilogueWarps */ 4 @@ -45226,17 +50983,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySfLdsSttm */ 0 , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 -, /* mNumRegsPerThreadNonEpilogueWarp */ 48 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 4 -, /* mNumStagesB */ 4 +, /* mNumStagesA */ 9 +, /* mNumStagesB */ 9 , /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 2 -, /* mNumStagesSmemSfA */ 4 -, /* mNumStagesSmemSfB */ 4 -, /* mNumStagesTmemSfA */ 4 -, /* mNumStagesTmemSfB */ 4 +, /* mNumStagesSmemSfA */ 9 +, /* mNumStagesSmemSfB */ 9 +, /* mNumStagesTmemSfA */ 9 +, /* mNumStagesTmemSfB */ 9 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -45251,9 +51008,9 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) , /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 512 +, /* mTileK */ 256 , /* mTileM */ 128 -, /* mTileN */ 64 +, /* mTileN */ 8 , /* mTileScheduler */ gemm::TileScheduler(1) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 @@ -45271,7 +51028,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseUnrollLoop2xForMma */ 1 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 1024 +, /* mValidK */ 512 , /* mWorldSize */ 1 , /* mActType */ gemmGatedAct::ActType(1) , /* mClampBeforeAct */ 1 @@ -45294,17 +51051,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 , /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 196248, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 896, "3926328a24d9c0345c271e1c1abe8788f115f28ffda89d63f959f1eda0232327", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 183440, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "9757da7a05a6bf705f0cea78e0f5878d77e1016960064573b035c515d43857cf", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 -, /* mClusterDimX */ 2 +, /* mClusterDimX */ 1 , /* mClusterDimY */ 1 , /* mClusterDimZ */ 1 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) @@ -45321,7 +51078,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mEpilogueLdtmDps */ 16 , /* mEpilogueLdtmBits */ 256 , /* mEpilogueTileM */ 128 -, /* mEpilogueTileN */ 32 +, /* mEpilogueTileN */ 8 , /* mFallbackClusterDimX */ 1 , /* mFallbackClusterDimY */ 1 , /* mFallbackClusterDimZ */ 1 @@ -45335,16 +51092,16 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 1024 +, /* mK */ 512 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) , /* mM */ 256 , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) -, /* mMmaM */ 256 -, /* mMmaN */ 64 -, /* mMmaTileK */ 512 +, /* mMmaM */ 128 +, /* mMmaN */ 8 +, /* mMmaTileK */ 256 , /* mMockAllReduce */ 0 , /* mN */ 256 , /* mNumEpilogueWarps */ 4 @@ -45352,17 +51109,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySfLdsSttm */ 0 , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 -, /* mNumRegsPerThreadNonEpilogueWarp */ 48 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 4 -, /* mNumStagesB */ 4 +, /* mNumStagesA */ 9 +, /* mNumStagesB */ 9 , /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 2 -, /* mNumStagesSmemSfA */ 4 -, /* mNumStagesSmemSfB */ 4 -, /* mNumStagesTmemSfA */ 4 -, /* mNumStagesTmemSfB */ 4 +, /* mNumStagesSmemSfA */ 9 +, /* mNumStagesSmemSfB */ 9 +, /* mNumStagesTmemSfA */ 9 +, /* mNumStagesTmemSfB */ 9 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -45377,9 +51134,9 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) , /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 512 +, /* mTileK */ 256 , /* mTileM */ 128 -, /* mTileN */ 64 +, /* mTileN */ 8 , /* mTileScheduler */ gemm::TileScheduler(1) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 @@ -45397,9 +51154,9 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseUnrollLoop2xForMma */ 1 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 1024 +, /* mValidK */ 512 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(0) +, /* mActType */ gemmGatedAct::ActType(2) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} @@ -45420,143 +51177,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 , /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} -, /* mPrefetchB */ 0 -, /* mUseCMultiCast */ 0 -, /* mUseTmaOobOpt */ 1 - }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 196248, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 896, "d68a779a16ef7e5112e6cf9178ba4c90a9cafd33f3ac3cc473bded7b4f12b38c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) -, /* mBiasDtype */ trtllm::gen::Dtype(1056776) -, /* mBiasType */ gemm::BiasType(1) -, /* mBlockK */ -1 -, /* mClcFastDrain */ 1 -, /* mClusterDimX */ 2 -, /* mClusterDimY */ 1 -, /* mClusterDimZ */ 1 -, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) -, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) -, /* mDtypeA */ trtllm::gen::Dtype(17826818) -, /* mDtypeB */ trtllm::gen::Dtype(17826818) -, /* mDtypeC */ trtllm::gen::Dtype(17826818) -, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) -, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) -, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(2) -, /* mEnablesEarlyExit */ 1 -, /* mEnablesDelayedEarlyExit */ 0 -, /* mEpilogueLdtmDps */ 16 -, /* mEpilogueLdtmBits */ 256 -, /* mEpilogueTileM */ 128 -, /* mEpilogueTileN */ 32 -, /* mFallbackClusterDimX */ 1 -, /* mFallbackClusterDimY */ 1 -, /* mFallbackClusterDimZ */ 1 -, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) -, /* mFuseLoadSfTask */ 0 -, /* mFuseUtccpWithUtcmma */ 0 -, /* mGridTriggerSecondaryA */ 0 -, /* mGridTriggerSecondaryB */ 1 -, /* mGridWaitForPrimaryEarlyExit */ 1 -, /* mGridWaitForPrimaryA */ 0 -, /* mGridWaitForPrimaryB */ 1 -, /* mHoistLoadTaskInit */ 1 -, /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 1024 -, /* mKernelTraits */ {} -, /* mLayoutA */ gemm::MatrixLayout(0) -, /* mLayoutB */ gemm::MatrixLayout(0) -, /* mM */ 256 -, /* mMmaK */ 64 -, /* mMmaKind */ trtllm::gen::MmaKind(4) -, /* mMmaM */ 256 -, /* mMmaN */ 64 -, /* mMmaTileK */ 512 -, /* mMockAllReduce */ 0 -, /* mN */ 256 -, /* mNumEpilogueWarps */ 4 -, /* mNumRegsCastAWarps */ 0 -, /* mNumRegsCopySfLdsSttm */ 0 -, /* mNumRegsCopySparsityInfo */ 0 -, /* mNumRegsPerThreadEpilogueWarp */ 0 -, /* mNumRegsPerThreadNonEpilogueWarp */ 48 -, /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 4 -, /* mNumStagesB */ 4 -, /* mNumStagesMma */ 2 -, /* mNumStagesMmaWithinWorkTile */ 1 -, /* mNumStagesMmaAcrossWorkTile */ 2 -, /* mNumStagesSmemSfA */ 4 -, /* mNumStagesSmemSfB */ 4 -, /* mNumStagesTmemSfA */ 4 -, /* mNumStagesTmemSfB */ 4 -, /* mNumStagesWorkId */ 3 -, /* mOutputDebugTensors */ 0 -, /* mPatchF2fp */ 0 -, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) -, /* mSchedHostTask */ gemm::SchedHostTask(0) -, /* mSfBlockSizeA */ 16 -, /* mSfBlockSizeB */ 16 -, /* mSfBlockSizeC */ 16 -, /* mSfLayoutA */ trtllm::gen::SfLayout(3) -, /* mSfLayoutB */ trtllm::gen::SfLayout(0) -, /* mSfLayoutC */ trtllm::gen::SfLayout(1) -, /* mSfReshapeFactor */ 1 -, /* mSparsityA */ trtllm::gen::Sparsity(0) -, /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 512 -, /* mTileM */ 128 -, /* mTileN */ 64 -, /* mTileScheduler */ gemm::TileScheduler(1) -, /* mTransposeMmaOutput */ 1 -, /* mUseCustomizedMma3xNvFp4 */ 0 -, /* mUseCustomMmaSchedule */ 1 -, /* mUseDeepSeekFp8 */ 0 -, /* mUseFlexibleClusterDims */ 0 -, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 -, /* mUseMaxTmemOverlap */ 0 -, /* mUsePerTokenSfA */ 0 -, /* mUsePerTokenSfB */ 0 -, /* mUseShuffledMatrix */ 1 -, /* mUseTmaStore */ 1 -, /* mUseTwoTmaLoadWarps */ 1 -, /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 1 -, /* mValidM */ 256 -, /* mValidN */ 256 -, /* mValidK */ 1024 -, /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(3) -, /* mClampBeforeAct */ 1 -, /* mBatchedM */ {} -, /* mBatchedN */ {} -, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) -, /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 0 -, /* mGridWaitForPrimaryRouting */ 1 -, /* mIsStaticBatch */ 0 -, /* mIsUniformNumTokensPerBatch */ 0 -, /* mNumBatches */ 128 -, /* mNumRegsPerThreadLoadA */ 0 -, /* mNumRegsPerThreadLoadB */ 0 -, /* mNumRegsPerThreadLoadSfA */ 0 -, /* mNumRegsPerThreadLoadSfB */ 0 -, /* mNumTokens */ 2 -, /* mNumWarpsLoadA */ 0 -, /* mNumWarpsLoadB */ 0 -, /* mNumWarpsLoadSfA */ 0 -, /* mNumWarpsLoadSfB */ 0 -, /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 196248, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 896, "2de609606d699ad746d292251fe04f2bcfd725517405ca8584e54d4183e3c72f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 183440, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "22565f6ad545cc3ceae6c05acbea5ebb8b972a33962993569f77eff56dac12f9", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 -, /* mClusterDimX */ 2 +, /* mClusterDimX */ 1 , /* mClusterDimY */ 1 , /* mClusterDimZ */ 1 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) @@ -45567,13 +51198,13 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(3) +, /* mEltwiseActType */ gemm::EltwiseActType(0) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 , /* mEpilogueLdtmBits */ 256 , /* mEpilogueTileM */ 128 -, /* mEpilogueTileN */ 32 +, /* mEpilogueTileN */ 8 , /* mFallbackClusterDimX */ 1 , /* mFallbackClusterDimY */ 1 , /* mFallbackClusterDimZ */ 1 @@ -45587,16 +51218,16 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 1024 +, /* mK */ 512 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) , /* mM */ 256 , /* mMmaK */ 64 , /* mMmaKind */ trtllm::gen::MmaKind(4) -, /* mMmaM */ 256 -, /* mMmaN */ 64 -, /* mMmaTileK */ 512 +, /* mMmaM */ 128 +, /* mMmaN */ 8 +, /* mMmaTileK */ 256 , /* mMockAllReduce */ 0 , /* mN */ 256 , /* mNumEpilogueWarps */ 4 @@ -45604,17 +51235,17 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySfLdsSttm */ 0 , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 -, /* mNumRegsPerThreadNonEpilogueWarp */ 48 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 4 -, /* mNumStagesB */ 4 +, /* mNumStagesA */ 9 +, /* mNumStagesB */ 9 , /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 2 -, /* mNumStagesSmemSfA */ 4 -, /* mNumStagesSmemSfB */ 4 -, /* mNumStagesTmemSfA */ 4 -, /* mNumStagesTmemSfB */ 4 +, /* mNumStagesSmemSfA */ 9 +, /* mNumStagesSmemSfB */ 9 +, /* mNumStagesTmemSfA */ 9 +, /* mNumStagesTmemSfB */ 9 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -45629,9 +51260,9 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) , /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 512 +, /* mTileK */ 256 , /* mTileM */ 128 -, /* mTileN */ 64 +, /* mTileN */ 8 , /* mTileScheduler */ gemm::TileScheduler(1) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 @@ -45649,15 +51280,15 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseUnrollLoop2xForMma */ 1 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 1024 +, /* mValidK */ 512 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(3) +, /* mActType */ gemmGatedAct::ActType(0) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 0 +, /* mFusedAct */ 1 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 @@ -45672,12 +51303,12 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 , /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 183440, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f", 512, "8e29c1dce23de3e3a78bb3a3de3624feb581f37141b2116ead5abb6dc8b8d756", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 183440, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 512, "51b4f9ebf670997dca520fbad7a9019aad3ddb761851179ac11923323273efb2", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -45693,7 +51324,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEltwiseActType */ gemm::EltwiseActType(2) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 @@ -45713,7 +51344,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 256 +, /* mK */ 512 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) @@ -45772,18 +51403,18 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseTmaStore */ 1 , /* mUseTwoTmaLoadWarps */ 1 , /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 0 +, /* mUseUnrollLoop2xForMma */ 1 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 256 +, /* mValidK */ 512 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(1) +, /* mActType */ gemmGatedAct::ActType(0) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 1 +, /* mFusedAct */ 0 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 @@ -45803,7 +51434,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 183440, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "71fea993e95fdcad9b243119b403ade3224f3bb3f592b301690b63ce070bb3e7", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 183440, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 512, "479047859e0a155515477c5662b342cb12d2f185a5f320591af38fa781e88c7f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -45819,7 +51450,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEltwiseActType */ gemm::EltwiseActType(3) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 @@ -45839,7 +51470,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 256 +, /* mK */ 512 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) @@ -45898,18 +51529,18 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseTmaStore */ 1 , /* mUseTwoTmaLoadWarps */ 1 , /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 0 +, /* mUseUnrollLoop2xForMma */ 1 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 256 +, /* mValidK */ 512 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(0) +, /* mActType */ gemmGatedAct::ActType(3) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 1 +, /* mFusedAct */ 0 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 @@ -45929,7 +51560,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 183440, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 512, "1d9ec1d73577ce1239a44fb4c42f38370e7b213751fac7026e285710065feab2", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 183312, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f", 512, "99f84a0039d6b734ae26f9942610bb900bc1212b10e89dffbafb17680c4eb625", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -45945,7 +51576,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(2) +, /* mEltwiseActType */ gemm::EltwiseActType(0) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 @@ -45965,7 +51596,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 256 +, /* mK */ 512 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) @@ -45986,9 +51617,9 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumSlicesForSplitK */ 1 , /* mNumStagesA */ 9 , /* mNumStagesB */ 9 -, /* mNumStagesMma */ 2 +, /* mNumStagesMma */ 1 , /* mNumStagesMmaWithinWorkTile */ 1 -, /* mNumStagesMmaAcrossWorkTile */ 2 +, /* mNumStagesMmaAcrossWorkTile */ 1 , /* mNumStagesSmemSfA */ 9 , /* mNumStagesSmemSfB */ 9 , /* mNumStagesTmemSfA */ 9 @@ -46010,7 +51641,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mTileK */ 256 , /* mTileM */ 128 , /* mTileN */ 8 -, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTileScheduler */ gemm::TileScheduler(0) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 , /* mUseCustomMmaSchedule */ 1 @@ -46024,18 +51655,18 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseTmaStore */ 1 , /* mUseTwoTmaLoadWarps */ 1 , /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 0 +, /* mUseUnrollLoop2xForMma */ 1 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 256 +, /* mValidK */ 512 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(0) +, /* mActType */ gemmGatedAct::ActType(1) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 0 +, /* mFusedAct */ 1 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 @@ -46055,7 +51686,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 183440, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 512, "6d45a517f003dbda07c4561c3b2074e0534109f9e5ffacdd48c0c418b54adc62", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 183312, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "a70aa89880e5277d6329b2bb5073d2cb63e1a5a8af73295bcad61d8ce55ca3fd", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -46071,7 +51702,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(3) +, /* mEltwiseActType */ gemm::EltwiseActType(0) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 @@ -46091,7 +51722,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 256 +, /* mK */ 512 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) @@ -46112,9 +51743,9 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumSlicesForSplitK */ 1 , /* mNumStagesA */ 9 , /* mNumStagesB */ 9 -, /* mNumStagesMma */ 2 +, /* mNumStagesMma */ 1 , /* mNumStagesMmaWithinWorkTile */ 1 -, /* mNumStagesMmaAcrossWorkTile */ 2 +, /* mNumStagesMmaAcrossWorkTile */ 1 , /* mNumStagesSmemSfA */ 9 , /* mNumStagesSmemSfB */ 9 , /* mNumStagesTmemSfA */ 9 @@ -46136,7 +51767,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mTileK */ 256 , /* mTileM */ 128 , /* mTileN */ 8 -, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTileScheduler */ gemm::TileScheduler(0) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 , /* mUseCustomMmaSchedule */ 1 @@ -46150,18 +51781,18 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseTmaStore */ 1 , /* mUseTwoTmaLoadWarps */ 1 , /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 0 +, /* mUseUnrollLoop2xForMma */ 1 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 256 +, /* mValidK */ 512 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(3) +, /* mActType */ gemmGatedAct::ActType(2) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 0 +, /* mFusedAct */ 1 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 @@ -46181,7 +51812,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 183312, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f", 512, "6bf7414821ee2bd06588ffd1105454750ea587b4d818367d51e9c0a0e1b6288f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 183312, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "37528a2f425a1664d0969810281d6f7e8b04361954adf72ea71098a94a2a83ce", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -46217,7 +51848,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 256 +, /* mK */ 512 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) @@ -46276,12 +51907,12 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseTmaStore */ 1 , /* mUseTwoTmaLoadWarps */ 1 , /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 0 +, /* mUseUnrollLoop2xForMma */ 1 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 256 +, /* mValidK */ 512 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(1) +, /* mActType */ gemmGatedAct::ActType(0) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} @@ -46307,7 +51938,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 183312, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "e03d90cd77399704927d22f3cb3078e2e6b83800aeb5868345abcd5ae999d21c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 183312, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 512, "144646aaefee3b59f0c0943e6252c1067851508e10ee6c62e2876b1e878a133d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -46323,7 +51954,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEltwiseActType */ gemm::EltwiseActType(2) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 @@ -46343,7 +51974,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 256 +, /* mK */ 512 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) @@ -46402,10 +52033,10 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseTmaStore */ 1 , /* mUseTwoTmaLoadWarps */ 1 , /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 0 +, /* mUseUnrollLoop2xForMma */ 1 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 256 +, /* mValidK */ 512 , /* mWorldSize */ 1 , /* mActType */ gemmGatedAct::ActType(0) , /* mClampBeforeAct */ 1 @@ -46413,7 +52044,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 1 +, /* mFusedAct */ 0 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 @@ -46433,7 +52064,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 183312, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 512, "f2ecc35a2129d90cb3a1bf9277385d3a25aead39553b7e7f0542a256da497e5e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 183312, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 512, "2ac47e7b2f8c10ff2a2f749029efff7269f9e3f9fa501d6da8c00b30fa86e5e4", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -46449,7 +52080,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(2) +, /* mEltwiseActType */ gemm::EltwiseActType(3) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 @@ -46469,7 +52100,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 256 +, /* mK */ 512 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) @@ -46528,12 +52159,12 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseTmaStore */ 1 , /* mUseTwoTmaLoadWarps */ 1 , /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 0 +, /* mUseUnrollLoop2xForMma */ 1 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 256 +, /* mValidK */ 512 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(0) +, /* mActType */ gemmGatedAct::ActType(3) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} @@ -46559,7 +52190,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 183312, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 512, "4a973cfde8c9634215edbbede3bdbdca99049e07aedbe40e4f0d70bb7d107fc7", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f_cubin_len, 162320, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f", 512, "166e385f7cab611c883b79ae0d95077489dd9f6bf689bad9496282eb16a08b30", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -46575,8 +52206,8 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(3) -, /* mEnablesEarlyExit */ 1 +, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEnablesEarlyExit */ 0 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 , /* mEpilogueLdtmBits */ 256 @@ -46595,7 +52226,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 256 +, /* mK */ 512 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) @@ -46604,7 +52235,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mMmaKind */ trtllm::gen::MmaKind(4) , /* mMmaM */ 128 , /* mMmaN */ 8 -, /* mMmaTileK */ 256 +, /* mMmaTileK */ 512 , /* mMockAllReduce */ 0 , /* mN */ 256 , /* mNumEpilogueWarps */ 4 @@ -46614,15 +52245,15 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 9 -, /* mNumStagesB */ 9 +, /* mNumStagesA */ 4 +, /* mNumStagesB */ 4 , /* mNumStagesMma */ 1 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 1 -, /* mNumStagesSmemSfA */ 9 -, /* mNumStagesSmemSfB */ 9 -, /* mNumStagesTmemSfA */ 9 -, /* mNumStagesTmemSfB */ 9 +, /* mNumStagesSmemSfA */ 4 +, /* mNumStagesSmemSfB */ 4 +, /* mNumStagesTmemSfA */ 4 +, /* mNumStagesTmemSfB */ 4 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -46632,12 +52263,12 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfBlockSizeB */ 16 , /* mSfBlockSizeC */ 16 , /* mSfLayoutA */ trtllm::gen::SfLayout(3) -, /* mSfLayoutB */ trtllm::gen::SfLayout(0) +, /* mSfLayoutB */ trtllm::gen::SfLayout(1) , /* mSfLayoutC */ trtllm::gen::SfLayout(1) , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) , /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 256 +, /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 8 , /* mTileScheduler */ gemm::TileScheduler(0) @@ -46657,9 +52288,9 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 256 +, /* mValidK */ 512 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(3) +, /* mActType */ gemmGatedAct::ActType(0) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} @@ -46667,6 +52298,132 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mBatchStrideInTokens */ -1 , /* mFusedAct */ 0 , /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 1 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 2 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 0 +, /* mNumRegsPerThreadLoadSfA */ 0 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 0 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 0 +, /* mNumWarpsLoadSfA */ 0 +, /* mNumWarpsLoadSfB */ 0 +, /* mRouteImpl */ batchedGemm::RouteImpl(0) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(0)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 202640, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f", 640, "8df73b4616585f450d43702bb77e24f15929b9b42efb6afc36e36c80c49ac480", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(1) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 1 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 1 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEnablesEarlyExit */ 1 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 16 +, /* mEpilogueLdtmBits */ 256 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 8 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 0 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 512 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 128 +, /* mMmaN */ 8 +, /* mMmaTileK */ 512 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 5 +, /* mNumStagesB */ 5 +, /* mNumStagesMma */ 2 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 2 +, /* mNumStagesSmemSfA */ 5 +, /* mNumStagesSmemSfB */ 5 +, /* mNumStagesTmemSfA */ 5 +, /* mNumStagesTmemSfB */ 5 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(3) +, /* mSfLayoutB */ trtllm::gen::SfLayout(0) +, /* mSfLayoutC */ trtllm::gen::SfLayout(1) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 512 +, /* mTileM */ 128 +, /* mTileN */ 8 +, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTransposeMmaOutput */ 1 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 0 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 0 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 512 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(1) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 1 +, /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 , /* mNumBatches */ 128 @@ -46680,12 +52437,12 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 , /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 183440, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f", 512, "65be3280bd48adfaa047f33bcd3da29f208aca32eba52d2155d00385e0a99db1", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 202640, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f", 640, "c13ac075b89b8f69d108f48a37fb769e690e1b9bf8e5994b27de9e7bae819f5c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -46730,7 +52487,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mMmaKind */ trtllm::gen::MmaKind(4) , /* mMmaM */ 128 , /* mMmaN */ 8 -, /* mMmaTileK */ 256 +, /* mMmaTileK */ 512 , /* mMockAllReduce */ 0 , /* mN */ 256 , /* mNumEpilogueWarps */ 4 @@ -46740,15 +52497,15 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 9 -, /* mNumStagesB */ 9 +, /* mNumStagesA */ 5 +, /* mNumStagesB */ 5 , /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 2 -, /* mNumStagesSmemSfA */ 9 -, /* mNumStagesSmemSfB */ 9 -, /* mNumStagesTmemSfA */ 9 -, /* mNumStagesTmemSfB */ 9 +, /* mNumStagesSmemSfA */ 5 +, /* mNumStagesSmemSfB */ 5 +, /* mNumStagesTmemSfA */ 5 +, /* mNumStagesTmemSfB */ 5 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -46763,7 +52520,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) , /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 256 +, /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 8 , /* mTileScheduler */ gemm::TileScheduler(1) @@ -46780,12 +52537,12 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseTmaStore */ 1 , /* mUseTwoTmaLoadWarps */ 1 , /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 1 +, /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 , /* mValidK */ 512 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(1) +, /* mActType */ gemmGatedAct::ActType(2) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} @@ -46806,12 +52563,12 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 , /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 183440, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "6d4a57d9cbfe434b5c3fa5ec7105366557b0409eeead03654552ee451e8c3345", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 202640, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "a4501a19831bc40be77b8fd0f28e120c70bcb9991cbc3859c7b1e9e3feac8c15", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -46856,7 +52613,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mMmaKind */ trtllm::gen::MmaKind(4) , /* mMmaM */ 128 , /* mMmaN */ 8 -, /* mMmaTileK */ 256 +, /* mMmaTileK */ 512 , /* mMockAllReduce */ 0 , /* mN */ 256 , /* mNumEpilogueWarps */ 4 @@ -46866,15 +52623,15 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 9 -, /* mNumStagesB */ 9 +, /* mNumStagesA */ 5 +, /* mNumStagesB */ 5 , /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 2 -, /* mNumStagesSmemSfA */ 9 -, /* mNumStagesSmemSfB */ 9 -, /* mNumStagesTmemSfA */ 9 -, /* mNumStagesTmemSfB */ 9 +, /* mNumStagesSmemSfA */ 5 +, /* mNumStagesSmemSfB */ 5 +, /* mNumStagesTmemSfA */ 5 +, /* mNumStagesTmemSfB */ 5 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -46889,7 +52646,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) , /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 256 +, /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 8 , /* mTileScheduler */ gemm::TileScheduler(1) @@ -46906,7 +52663,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseTmaStore */ 1 , /* mUseTwoTmaLoadWarps */ 1 , /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 1 +, /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 , /* mValidK */ 512 @@ -46932,12 +52689,12 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 , /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 183440, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 512, "f4f6d78fbf0b9370f07da3d5054aad54596f132f6d7239e90f3255d362b1ba3e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 202640, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 640, "d7f9d16d930c4cbc527eba9c5f8b27e23bb06ab3d06afc19e79a851247fa6bf0", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -46982,7 +52739,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mMmaKind */ trtllm::gen::MmaKind(4) , /* mMmaM */ 128 , /* mMmaN */ 8 -, /* mMmaTileK */ 256 +, /* mMmaTileK */ 512 , /* mMockAllReduce */ 0 , /* mN */ 256 , /* mNumEpilogueWarps */ 4 @@ -46992,15 +52749,15 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 9 -, /* mNumStagesB */ 9 +, /* mNumStagesA */ 5 +, /* mNumStagesB */ 5 , /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 2 -, /* mNumStagesSmemSfA */ 9 -, /* mNumStagesSmemSfB */ 9 -, /* mNumStagesTmemSfA */ 9 -, /* mNumStagesTmemSfB */ 9 +, /* mNumStagesSmemSfA */ 5 +, /* mNumStagesSmemSfB */ 5 +, /* mNumStagesTmemSfA */ 5 +, /* mNumStagesTmemSfB */ 5 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -47015,7 +52772,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) , /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 256 +, /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 8 , /* mTileScheduler */ gemm::TileScheduler(1) @@ -47032,7 +52789,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseTmaStore */ 1 , /* mUseTwoTmaLoadWarps */ 1 , /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 1 +, /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 , /* mValidK */ 512 @@ -47058,12 +52815,12 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 , /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 183440, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 512, "1d135437b451de27407e1e3e2f2c94cae5319acc37d78e16467b91d5593dfef7", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 202640, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 640, "95e81feaef1da7e3bf7d74e6a6b8b60d0299091be68b8b6682cf9ba52712d7f2", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -47108,7 +52865,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mMmaKind */ trtllm::gen::MmaKind(4) , /* mMmaM */ 128 , /* mMmaN */ 8 -, /* mMmaTileK */ 256 +, /* mMmaTileK */ 512 , /* mMockAllReduce */ 0 , /* mN */ 256 , /* mNumEpilogueWarps */ 4 @@ -47118,15 +52875,15 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 9 -, /* mNumStagesB */ 9 +, /* mNumStagesA */ 5 +, /* mNumStagesB */ 5 , /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 2 -, /* mNumStagesSmemSfA */ 9 -, /* mNumStagesSmemSfB */ 9 -, /* mNumStagesTmemSfA */ 9 -, /* mNumStagesTmemSfB */ 9 +, /* mNumStagesSmemSfA */ 5 +, /* mNumStagesSmemSfB */ 5 +, /* mNumStagesTmemSfA */ 5 +, /* mNumStagesTmemSfB */ 5 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -47141,7 +52898,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) , /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 256 +, /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 8 , /* mTileScheduler */ gemm::TileScheduler(1) @@ -47158,7 +52915,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseTmaStore */ 1 , /* mUseTwoTmaLoadWarps */ 1 , /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 1 +, /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 , /* mValidK */ 512 @@ -47184,12 +52941,12 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 , /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 183312, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f", 512, "323d7117879c4856e8303c6436e4260b4cf67488231b908df2e1b6fdd6f376e7", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 202256, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f", 512, "48916ea8cf771ebfba5382d5a73665a7210028aba1c8e2bab0cdc92bf9a161af", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -47234,7 +52991,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mMmaKind */ trtllm::gen::MmaKind(4) , /* mMmaM */ 128 , /* mMmaN */ 8 -, /* mMmaTileK */ 256 +, /* mMmaTileK */ 512 , /* mMockAllReduce */ 0 , /* mN */ 256 , /* mNumEpilogueWarps */ 4 @@ -47244,15 +53001,15 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 9 -, /* mNumStagesB */ 9 +, /* mNumStagesA */ 5 +, /* mNumStagesB */ 5 , /* mNumStagesMma */ 1 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 1 -, /* mNumStagesSmemSfA */ 9 -, /* mNumStagesSmemSfB */ 9 -, /* mNumStagesTmemSfA */ 9 -, /* mNumStagesTmemSfB */ 9 +, /* mNumStagesSmemSfA */ 5 +, /* mNumStagesSmemSfB */ 5 +, /* mNumStagesTmemSfA */ 5 +, /* mNumStagesTmemSfB */ 5 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -47267,7 +53024,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) , /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 256 +, /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 8 , /* mTileScheduler */ gemm::TileScheduler(0) @@ -47284,7 +53041,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseTmaStore */ 1 , /* mUseTwoTmaLoadWarps */ 1 , /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 1 +, /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 , /* mValidK */ 512 @@ -47310,12 +53067,12 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 , /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 183312, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "9a3a54df273167d5f485cdd4e94b109885d31185aaf8c38cd6f1306fbebcfbe9", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 202256, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "994b7b9758ca1c250fc0b8c37ac85cc15f3325ec49417438b74177bf4d91dd79", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -47360,7 +53117,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mMmaKind */ trtllm::gen::MmaKind(4) , /* mMmaM */ 128 , /* mMmaN */ 8 -, /* mMmaTileK */ 256 +, /* mMmaTileK */ 512 , /* mMockAllReduce */ 0 , /* mN */ 256 , /* mNumEpilogueWarps */ 4 @@ -47370,15 +53127,15 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 9 -, /* mNumStagesB */ 9 +, /* mNumStagesA */ 5 +, /* mNumStagesB */ 5 , /* mNumStagesMma */ 1 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 1 -, /* mNumStagesSmemSfA */ 9 -, /* mNumStagesSmemSfB */ 9 -, /* mNumStagesTmemSfA */ 9 -, /* mNumStagesTmemSfB */ 9 +, /* mNumStagesSmemSfA */ 5 +, /* mNumStagesSmemSfB */ 5 +, /* mNumStagesTmemSfA */ 5 +, /* mNumStagesTmemSfB */ 5 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -47393,7 +53150,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) , /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 256 +, /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 8 , /* mTileScheduler */ gemm::TileScheduler(0) @@ -47410,12 +53167,12 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseTmaStore */ 1 , /* mUseTwoTmaLoadWarps */ 1 , /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 1 +, /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 , /* mValidK */ 512 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(0) +, /* mActType */ gemmGatedAct::ActType(2) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} @@ -47436,12 +53193,12 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 , /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 183312, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 512, "ab992021edc350e300aaafd13a2647b99d717033df48915de033cfe1129db856", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 202256, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "a881ee0cb80cbceb742f2e4133995b739a5110ebe06753c18a7dff4de6d3eb96", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -47457,7 +53214,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(2) +, /* mEltwiseActType */ gemm::EltwiseActType(0) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 @@ -47486,7 +53243,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mMmaKind */ trtllm::gen::MmaKind(4) , /* mMmaM */ 128 , /* mMmaN */ 8 -, /* mMmaTileK */ 256 +, /* mMmaTileK */ 512 , /* mMockAllReduce */ 0 , /* mN */ 256 , /* mNumEpilogueWarps */ 4 @@ -47496,15 +53253,15 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 9 -, /* mNumStagesB */ 9 +, /* mNumStagesA */ 5 +, /* mNumStagesB */ 5 , /* mNumStagesMma */ 1 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 1 -, /* mNumStagesSmemSfA */ 9 -, /* mNumStagesSmemSfB */ 9 -, /* mNumStagesTmemSfA */ 9 -, /* mNumStagesTmemSfB */ 9 +, /* mNumStagesSmemSfA */ 5 +, /* mNumStagesSmemSfB */ 5 +, /* mNumStagesTmemSfA */ 5 +, /* mNumStagesTmemSfB */ 5 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -47519,7 +53276,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) , /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 256 +, /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 8 , /* mTileScheduler */ gemm::TileScheduler(0) @@ -47536,7 +53293,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseTmaStore */ 1 , /* mUseTwoTmaLoadWarps */ 1 , /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 1 +, /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 , /* mValidK */ 512 @@ -47547,7 +53304,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 0 +, /* mFusedAct */ 1 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 @@ -47562,12 +53319,12 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 , /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f_cubin_len, 183312, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f", 512, "62d288482b3da6e09c46764ca604e47bc373e61b1ef64f70d51f985c7bfb94e9", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 202256, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 512, "4dc62475c7109d223e32a2def85efc72e96120afb129822fe15bc9f529ad09d3", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -47583,7 +53340,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(3) +, /* mEltwiseActType */ gemm::EltwiseActType(2) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 @@ -47612,7 +53369,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mMmaKind */ trtllm::gen::MmaKind(4) , /* mMmaM */ 128 , /* mMmaN */ 8 -, /* mMmaTileK */ 256 +, /* mMmaTileK */ 512 , /* mMockAllReduce */ 0 , /* mN */ 256 , /* mNumEpilogueWarps */ 4 @@ -47622,15 +53379,15 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 9 -, /* mNumStagesB */ 9 +, /* mNumStagesA */ 5 +, /* mNumStagesB */ 5 , /* mNumStagesMma */ 1 , /* mNumStagesMmaWithinWorkTile */ 1 , /* mNumStagesMmaAcrossWorkTile */ 1 -, /* mNumStagesSmemSfA */ 9 -, /* mNumStagesSmemSfB */ 9 -, /* mNumStagesTmemSfA */ 9 -, /* mNumStagesTmemSfB */ 9 +, /* mNumStagesSmemSfA */ 5 +, /* mNumStagesSmemSfB */ 5 +, /* mNumStagesTmemSfA */ 5 +, /* mNumStagesTmemSfB */ 5 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -47645,7 +53402,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) , /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 256 +, /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 8 , /* mTileScheduler */ gemm::TileScheduler(0) @@ -47662,12 +53419,12 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseTmaStore */ 1 , /* mUseTwoTmaLoadWarps */ 1 , /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 1 +, /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 , /* mValidK */ 512 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(3) +, /* mActType */ gemmGatedAct::ActType(0) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} @@ -47688,138 +53445,12 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 , /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(1)} -, /* mPrefetchB */ 0 -, /* mUseCMultiCast */ 0 -, /* mUseTmaOobOpt */ 1 - }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f_cubin_len, 162320, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f", 512, "86a4157a091c5f8bfd084c5d21ff497ae15b4d456e8f53630fe8ef98195e336d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) -, /* mBiasDtype */ trtllm::gen::Dtype(1056776) -, /* mBiasType */ gemm::BiasType(1) -, /* mBlockK */ -1 -, /* mClcFastDrain */ 1 -, /* mClusterDimX */ 1 -, /* mClusterDimY */ 1 -, /* mClusterDimZ */ 1 -, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) -, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) -, /* mDtypeA */ trtllm::gen::Dtype(17826818) -, /* mDtypeB */ trtllm::gen::Dtype(17826818) -, /* mDtypeC */ trtllm::gen::Dtype(17826818) -, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) -, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) -, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(0) -, /* mEnablesEarlyExit */ 0 -, /* mEnablesDelayedEarlyExit */ 0 -, /* mEpilogueLdtmDps */ 16 -, /* mEpilogueLdtmBits */ 256 -, /* mEpilogueTileM */ 128 -, /* mEpilogueTileN */ 8 -, /* mFallbackClusterDimX */ 1 -, /* mFallbackClusterDimY */ 1 -, /* mFallbackClusterDimZ */ 1 -, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) -, /* mFuseLoadSfTask */ 0 -, /* mFuseUtccpWithUtcmma */ 0 -, /* mGridTriggerSecondaryA */ 0 -, /* mGridTriggerSecondaryB */ 1 -, /* mGridWaitForPrimaryEarlyExit */ 1 -, /* mGridWaitForPrimaryA */ 0 -, /* mGridWaitForPrimaryB */ 1 -, /* mHoistLoadTaskInit */ 1 -, /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 512 -, /* mKernelTraits */ {} -, /* mLayoutA */ gemm::MatrixLayout(0) -, /* mLayoutB */ gemm::MatrixLayout(0) -, /* mM */ 256 -, /* mMmaK */ 64 -, /* mMmaKind */ trtllm::gen::MmaKind(4) -, /* mMmaM */ 128 -, /* mMmaN */ 8 -, /* mMmaTileK */ 512 -, /* mMockAllReduce */ 0 -, /* mN */ 256 -, /* mNumEpilogueWarps */ 4 -, /* mNumRegsCastAWarps */ 0 -, /* mNumRegsCopySfLdsSttm */ 0 -, /* mNumRegsCopySparsityInfo */ 0 -, /* mNumRegsPerThreadEpilogueWarp */ 0 -, /* mNumRegsPerThreadNonEpilogueWarp */ 0 -, /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 4 -, /* mNumStagesB */ 4 -, /* mNumStagesMma */ 1 -, /* mNumStagesMmaWithinWorkTile */ 1 -, /* mNumStagesMmaAcrossWorkTile */ 1 -, /* mNumStagesSmemSfA */ 4 -, /* mNumStagesSmemSfB */ 4 -, /* mNumStagesTmemSfA */ 4 -, /* mNumStagesTmemSfB */ 4 -, /* mNumStagesWorkId */ 3 -, /* mOutputDebugTensors */ 0 -, /* mPatchF2fp */ 0 -, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) -, /* mSchedHostTask */ gemm::SchedHostTask(0) -, /* mSfBlockSizeA */ 16 -, /* mSfBlockSizeB */ 16 -, /* mSfBlockSizeC */ 16 -, /* mSfLayoutA */ trtllm::gen::SfLayout(3) -, /* mSfLayoutB */ trtllm::gen::SfLayout(1) -, /* mSfLayoutC */ trtllm::gen::SfLayout(1) -, /* mSfReshapeFactor */ 1 -, /* mSparsityA */ trtllm::gen::Sparsity(0) -, /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 512 -, /* mTileM */ 128 -, /* mTileN */ 8 -, /* mTileScheduler */ gemm::TileScheduler(0) -, /* mTransposeMmaOutput */ 1 -, /* mUseCustomizedMma3xNvFp4 */ 0 -, /* mUseCustomMmaSchedule */ 1 -, /* mUseDeepSeekFp8 */ 0 -, /* mUseFlexibleClusterDims */ 0 -, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 -, /* mUseMaxTmemOverlap */ 0 -, /* mUsePerTokenSfA */ 0 -, /* mUsePerTokenSfB */ 0 -, /* mUseShuffledMatrix */ 1 -, /* mUseTmaStore */ 1 -, /* mUseTwoTmaLoadWarps */ 1 -, /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 0 -, /* mValidM */ 256 -, /* mValidN */ 256 -, /* mValidK */ 512 -, /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(0) -, /* mClampBeforeAct */ 1 -, /* mBatchedM */ {} -, /* mBatchedN */ {} -, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) -, /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 0 -, /* mGridWaitForPrimaryRouting */ 1 -, /* mIsStaticBatch */ 1 -, /* mIsUniformNumTokensPerBatch */ 0 -, /* mNumBatches */ 2 -, /* mNumRegsPerThreadLoadA */ 0 -, /* mNumRegsPerThreadLoadB */ 0 -, /* mNumRegsPerThreadLoadSfA */ 0 -, /* mNumRegsPerThreadLoadSfB */ 0 -, /* mNumTokens */ 0 -, /* mNumWarpsLoadA */ 0 -, /* mNumWarpsLoadB */ 0 -, /* mNumWarpsLoadSfA */ 0 -, /* mNumWarpsLoadSfB */ 0 -, /* mRouteImpl */ batchedGemm::RouteImpl(0) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(0)} +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 202640, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f", 640, "d5ad91f85bca4a1626dc91840c080c0a3c33ca779167e95c4805eab520b85463", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 202256, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 512, "aa9ba58a9a027d36b4803e69dbb2124ee67921836f02b65e7eaedbae9e3fc20d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -47835,7 +53466,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEltwiseActType */ gemm::EltwiseActType(3) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 @@ -47876,9 +53507,9 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumSlicesForSplitK */ 1 , /* mNumStagesA */ 5 , /* mNumStagesB */ 5 -, /* mNumStagesMma */ 2 +, /* mNumStagesMma */ 1 , /* mNumStagesMmaWithinWorkTile */ 1 -, /* mNumStagesMmaAcrossWorkTile */ 2 +, /* mNumStagesMmaAcrossWorkTile */ 1 , /* mNumStagesSmemSfA */ 5 , /* mNumStagesSmemSfB */ 5 , /* mNumStagesTmemSfA */ 5 @@ -47900,7 +53531,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 8 -, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTileScheduler */ gemm::TileScheduler(0) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 , /* mUseCustomMmaSchedule */ 1 @@ -47919,13 +53550,13 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mValidN */ 256 , /* mValidK */ 512 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(1) +, /* mActType */ gemmGatedAct::ActType(3) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 1 +, /* mFusedAct */ 0 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 @@ -47945,14 +53576,14 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 202640, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "5f1ba6239f0d92f906ca94935f2236a04bcd8255a36038f4486a302c045c2afe", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 209816, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f", 640, "c1cf95f0b227a0fe51f2e79820eafe63171a86392fa117f771e46414acf7ffbe", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 , /* mClusterDimX */ 1 , /* mClusterDimY */ 1 -, /* mClusterDimZ */ 1 +, /* mClusterDimZ */ 2 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) , /* mDtypeAcc */ trtllm::gen::Dtype(1056776) , /* mDtypeA */ trtllm::gen::Dtype(17826818) @@ -47981,7 +53612,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 512 +, /* mK */ 1024 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) @@ -47999,7 +53630,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 -, /* mNumSlicesForSplitK */ 1 +, /* mNumSlicesForSplitK */ 2 , /* mNumStagesA */ 5 , /* mNumStagesB */ 5 , /* mNumStagesMma */ 2 @@ -48022,7 +53653,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfLayoutC */ trtllm::gen::SfLayout(1) , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) -, /* mSplitK */ gemm::SplitK(0) +, /* mSplitK */ gemm::SplitK(2) , /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 8 @@ -48043,9 +53674,9 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 512 +, /* mValidK */ 1024 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(0) +, /* mActType */ gemmGatedAct::ActType(1) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} @@ -48071,14 +53702,14 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 202640, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 640, "a6bde4cefa7444d67921e3bc58976f9a74ced058bd43170bd5fe0aa89608c428", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 209816, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f", 640, "99bbb3ad975988948b66edc4878563c8777a92f94a6fa38e5ce63eeee74ea9e0", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 , /* mClusterDimX */ 1 , /* mClusterDimY */ 1 -, /* mClusterDimZ */ 1 +, /* mClusterDimZ */ 2 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) , /* mDtypeAcc */ trtllm::gen::Dtype(1056776) , /* mDtypeA */ trtllm::gen::Dtype(17826818) @@ -48087,7 +53718,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(2) +, /* mEltwiseActType */ gemm::EltwiseActType(0) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 @@ -48107,7 +53738,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 512 +, /* mK */ 1024 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) @@ -48125,7 +53756,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 -, /* mNumSlicesForSplitK */ 1 +, /* mNumSlicesForSplitK */ 2 , /* mNumStagesA */ 5 , /* mNumStagesB */ 5 , /* mNumStagesMma */ 2 @@ -48148,7 +53779,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfLayoutC */ trtllm::gen::SfLayout(1) , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) -, /* mSplitK */ gemm::SplitK(0) +, /* mSplitK */ gemm::SplitK(2) , /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 8 @@ -48169,15 +53800,15 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 512 +, /* mValidK */ 1024 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(0) +, /* mActType */ gemmGatedAct::ActType(2) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 0 +, /* mFusedAct */ 1 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 @@ -48197,14 +53828,14 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 202640, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 640, "9329e4104a13301d2dcc7ce1ed4f076eace049e25914bc6ff54eeba028b90437", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 209816, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 640, "75f03c86cdcc5ba341942297967cbf0b0300490a3fd9f765d335fb68de31d71d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 , /* mClusterDimX */ 1 , /* mClusterDimY */ 1 -, /* mClusterDimZ */ 1 +, /* mClusterDimZ */ 2 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) , /* mDtypeAcc */ trtllm::gen::Dtype(1056776) , /* mDtypeA */ trtllm::gen::Dtype(17826818) @@ -48213,7 +53844,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(3) +, /* mEltwiseActType */ gemm::EltwiseActType(2) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 @@ -48233,7 +53864,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 512 +, /* mK */ 1024 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) @@ -48251,7 +53882,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 -, /* mNumSlicesForSplitK */ 1 +, /* mNumSlicesForSplitK */ 2 , /* mNumStagesA */ 5 , /* mNumStagesB */ 5 , /* mNumStagesMma */ 2 @@ -48274,7 +53905,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfLayoutC */ trtllm::gen::SfLayout(1) , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) -, /* mSplitK */ gemm::SplitK(0) +, /* mSplitK */ gemm::SplitK(2) , /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 8 @@ -48295,7 +53926,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 512 +, /* mValidK */ 1024 , /* mWorldSize */ 1 , /* mActType */ gemmGatedAct::ActType(3) , /* mClampBeforeAct */ 1 @@ -48323,14 +53954,14 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 202256, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f", 512, "0733341d8eb04862927c527f4d85447f2fe327409241ec4130716da44edfb0a2", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 209816, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 640, "7d54939d510f8ee62047e9dc4305d2c327982ce3a23a4c7776e912dfb17c3e39", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 , /* mClusterDimX */ 1 , /* mClusterDimY */ 1 -, /* mClusterDimZ */ 1 +, /* mClusterDimZ */ 2 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) , /* mDtypeAcc */ trtllm::gen::Dtype(1056776) , /* mDtypeA */ trtllm::gen::Dtype(17826818) @@ -48339,7 +53970,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEltwiseActType */ gemm::EltwiseActType(3) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 @@ -48359,7 +53990,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 512 +, /* mK */ 1024 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) @@ -48377,12 +54008,12 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 -, /* mNumSlicesForSplitK */ 1 +, /* mNumSlicesForSplitK */ 2 , /* mNumStagesA */ 5 , /* mNumStagesB */ 5 -, /* mNumStagesMma */ 1 +, /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 -, /* mNumStagesMmaAcrossWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 2 , /* mNumStagesSmemSfA */ 5 , /* mNumStagesSmemSfB */ 5 , /* mNumStagesTmemSfA */ 5 @@ -48400,11 +54031,11 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfLayoutC */ trtllm::gen::SfLayout(1) , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) -, /* mSplitK */ gemm::SplitK(0) +, /* mSplitK */ gemm::SplitK(2) , /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 8 -, /* mTileScheduler */ gemm::TileScheduler(0) +, /* mTileScheduler */ gemm::TileScheduler(1) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 , /* mUseCustomMmaSchedule */ 1 @@ -48421,15 +54052,15 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 512 +, /* mValidK */ 1024 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(1) +, /* mActType */ gemmGatedAct::ActType(3) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 1 +, /* mFusedAct */ 0 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 @@ -48449,14 +54080,14 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 202256, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "5dbd34e5ab514329e2aefda61b5a2bea7f0582baddd263e1926ed1e19c59d774", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 213912, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f", 640, "e258a527e87f18b7d5e7a741ecbb9fcd2815912fbb1b1753fe1a2139c3ac6076", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 , /* mClusterDimX */ 1 , /* mClusterDimY */ 1 -, /* mClusterDimZ */ 1 +, /* mClusterDimZ */ 3 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) , /* mDtypeAcc */ trtllm::gen::Dtype(1056776) , /* mDtypeA */ trtllm::gen::Dtype(17826818) @@ -48485,7 +54116,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 512 +, /* mK */ 1536 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) @@ -48503,12 +54134,12 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 -, /* mNumSlicesForSplitK */ 1 +, /* mNumSlicesForSplitK */ 3 , /* mNumStagesA */ 5 , /* mNumStagesB */ 5 -, /* mNumStagesMma */ 1 +, /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 -, /* mNumStagesMmaAcrossWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 2 , /* mNumStagesSmemSfA */ 5 , /* mNumStagesSmemSfB */ 5 , /* mNumStagesTmemSfA */ 5 @@ -48526,11 +54157,11 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfLayoutC */ trtllm::gen::SfLayout(1) , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) -, /* mSplitK */ gemm::SplitK(0) +, /* mSplitK */ gemm::SplitK(2) , /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 8 -, /* mTileScheduler */ gemm::TileScheduler(0) +, /* mTileScheduler */ gemm::TileScheduler(1) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 , /* mUseCustomMmaSchedule */ 1 @@ -48547,9 +54178,9 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 512 +, /* mValidK */ 1536 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(0) +, /* mActType */ gemmGatedAct::ActType(1) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} @@ -48575,140 +54206,14 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 202256, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 512, "e6c882a43186ddf025da419f2a04ae45c8ee8352b2d294db7e0c8d11a785c2d6", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 213912, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f", 640, "3ad84b97b2e32771bf3aea3f23ba8b0e1895b792feaf4b4887931ce656c6f895", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 , /* mClusterDimX */ 1 , /* mClusterDimY */ 1 -, /* mClusterDimZ */ 1 -, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) -, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) -, /* mDtypeA */ trtllm::gen::Dtype(17826818) -, /* mDtypeB */ trtllm::gen::Dtype(17826818) -, /* mDtypeC */ trtllm::gen::Dtype(17826818) -, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) -, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) -, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(2) -, /* mEnablesEarlyExit */ 1 -, /* mEnablesDelayedEarlyExit */ 0 -, /* mEpilogueLdtmDps */ 16 -, /* mEpilogueLdtmBits */ 256 -, /* mEpilogueTileM */ 128 -, /* mEpilogueTileN */ 8 -, /* mFallbackClusterDimX */ 1 -, /* mFallbackClusterDimY */ 1 -, /* mFallbackClusterDimZ */ 1 -, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) -, /* mFuseLoadSfTask */ 0 -, /* mFuseUtccpWithUtcmma */ 0 -, /* mGridTriggerSecondaryA */ 0 -, /* mGridTriggerSecondaryB */ 1 -, /* mGridWaitForPrimaryEarlyExit */ 1 -, /* mGridWaitForPrimaryA */ 0 -, /* mGridWaitForPrimaryB */ 1 -, /* mHoistLoadTaskInit */ 1 -, /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 512 -, /* mKernelTraits */ {} -, /* mLayoutA */ gemm::MatrixLayout(0) -, /* mLayoutB */ gemm::MatrixLayout(0) -, /* mM */ 256 -, /* mMmaK */ 64 -, /* mMmaKind */ trtllm::gen::MmaKind(4) -, /* mMmaM */ 128 -, /* mMmaN */ 8 -, /* mMmaTileK */ 512 -, /* mMockAllReduce */ 0 -, /* mN */ 256 -, /* mNumEpilogueWarps */ 4 -, /* mNumRegsCastAWarps */ 0 -, /* mNumRegsCopySfLdsSttm */ 0 -, /* mNumRegsCopySparsityInfo */ 0 -, /* mNumRegsPerThreadEpilogueWarp */ 0 -, /* mNumRegsPerThreadNonEpilogueWarp */ 0 -, /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 5 -, /* mNumStagesB */ 5 -, /* mNumStagesMma */ 1 -, /* mNumStagesMmaWithinWorkTile */ 1 -, /* mNumStagesMmaAcrossWorkTile */ 1 -, /* mNumStagesSmemSfA */ 5 -, /* mNumStagesSmemSfB */ 5 -, /* mNumStagesTmemSfA */ 5 -, /* mNumStagesTmemSfB */ 5 -, /* mNumStagesWorkId */ 3 -, /* mOutputDebugTensors */ 0 -, /* mPatchF2fp */ 0 -, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) -, /* mSchedHostTask */ gemm::SchedHostTask(0) -, /* mSfBlockSizeA */ 16 -, /* mSfBlockSizeB */ 16 -, /* mSfBlockSizeC */ 16 -, /* mSfLayoutA */ trtllm::gen::SfLayout(3) -, /* mSfLayoutB */ trtllm::gen::SfLayout(0) -, /* mSfLayoutC */ trtllm::gen::SfLayout(1) -, /* mSfReshapeFactor */ 1 -, /* mSparsityA */ trtllm::gen::Sparsity(0) -, /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 512 -, /* mTileM */ 128 -, /* mTileN */ 8 -, /* mTileScheduler */ gemm::TileScheduler(0) -, /* mTransposeMmaOutput */ 1 -, /* mUseCustomizedMma3xNvFp4 */ 0 -, /* mUseCustomMmaSchedule */ 1 -, /* mUseDeepSeekFp8 */ 0 -, /* mUseFlexibleClusterDims */ 0 -, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 -, /* mUseMaxTmemOverlap */ 0 -, /* mUsePerTokenSfA */ 0 -, /* mUsePerTokenSfB */ 0 -, /* mUseShuffledMatrix */ 1 -, /* mUseTmaStore */ 1 -, /* mUseTwoTmaLoadWarps */ 1 -, /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 0 -, /* mValidM */ 256 -, /* mValidN */ 256 -, /* mValidK */ 512 -, /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(0) -, /* mClampBeforeAct */ 1 -, /* mBatchedM */ {} -, /* mBatchedN */ {} -, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) -, /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 0 -, /* mGridWaitForPrimaryRouting */ 1 -, /* mIsStaticBatch */ 0 -, /* mIsUniformNumTokensPerBatch */ 0 -, /* mNumBatches */ 128 -, /* mNumRegsPerThreadLoadA */ 0 -, /* mNumRegsPerThreadLoadB */ 0 -, /* mNumRegsPerThreadLoadSfA */ 0 -, /* mNumRegsPerThreadLoadSfB */ 0 -, /* mNumTokens */ 2 -, /* mNumWarpsLoadA */ 0 -, /* mNumWarpsLoadB */ 0 -, /* mNumWarpsLoadSfA */ 0 -, /* mNumWarpsLoadSfB */ 0 -, /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} -, /* mPrefetchB */ 0 -, /* mUseCMultiCast */ 0 -, /* mUseTmaOobOpt */ 1 - }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 202256, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 512, "f9c1f126bf693a5dd8d08839c9b5618b6ffa59ab212a8e227d4fbefbd99c45b5", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) -, /* mBiasDtype */ trtllm::gen::Dtype(1056776) -, /* mBiasType */ gemm::BiasType(1) -, /* mBlockK */ -1 -, /* mClcFastDrain */ 1 -, /* mClusterDimX */ 1 -, /* mClusterDimY */ 1 -, /* mClusterDimZ */ 1 +, /* mClusterDimZ */ 3 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) , /* mDtypeAcc */ trtllm::gen::Dtype(1056776) , /* mDtypeA */ trtllm::gen::Dtype(17826818) @@ -48717,7 +54222,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(3) +, /* mEltwiseActType */ gemm::EltwiseActType(0) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 @@ -48737,7 +54242,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 512 +, /* mK */ 1536 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) @@ -48755,12 +54260,12 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 -, /* mNumSlicesForSplitK */ 1 +, /* mNumSlicesForSplitK */ 3 , /* mNumStagesA */ 5 , /* mNumStagesB */ 5 -, /* mNumStagesMma */ 1 +, /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 -, /* mNumStagesMmaAcrossWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 2 , /* mNumStagesSmemSfA */ 5 , /* mNumStagesSmemSfB */ 5 , /* mNumStagesTmemSfA */ 5 @@ -48778,11 +54283,11 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfLayoutC */ trtllm::gen::SfLayout(1) , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) -, /* mSplitK */ gemm::SplitK(0) +, /* mSplitK */ gemm::SplitK(2) , /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 8 -, /* mTileScheduler */ gemm::TileScheduler(0) +, /* mTileScheduler */ gemm::TileScheduler(1) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 , /* mUseCustomMmaSchedule */ 1 @@ -48799,15 +54304,15 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 512 +, /* mValidK */ 1536 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(3) +, /* mActType */ gemmGatedAct::ActType(2) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 0 +, /* mFusedAct */ 1 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 @@ -48827,14 +54332,14 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 209816, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f", 640, "fd6a812279a8ba46eb54d090b36f4a754a66af1d25212e78d73bbbbb35c111ba", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 213912, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 640, "f0403bf6474ea1bb39d4989fc38463a61d86195e972b0b775e884386329ea24f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 , /* mClusterDimX */ 1 , /* mClusterDimY */ 1 -, /* mClusterDimZ */ 2 +, /* mClusterDimZ */ 3 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) , /* mDtypeAcc */ trtllm::gen::Dtype(1056776) , /* mDtypeA */ trtllm::gen::Dtype(17826818) @@ -48843,7 +54348,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEltwiseActType */ gemm::EltwiseActType(2) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 @@ -48863,7 +54368,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 1024 +, /* mK */ 1536 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) @@ -48881,7 +54386,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 -, /* mNumSlicesForSplitK */ 2 +, /* mNumSlicesForSplitK */ 3 , /* mNumStagesA */ 5 , /* mNumStagesB */ 5 , /* mNumStagesMma */ 2 @@ -48925,15 +54430,15 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 1024 +, /* mValidK */ 1536 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(1) +, /* mActType */ gemmGatedAct::ActType(3) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 1 +, /* mFusedAct */ 0 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 @@ -48953,14 +54458,14 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 209816, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 640, "f96be89c9b175fb8d261cd71e8f545c99300399289492d04fe01eecfc7e6a23b", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 213912, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 640, "1b1611b7b0e080532bd24445ec372468e1db6256deee0973ad647fa9fefe5882", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 , /* mClusterDimX */ 1 , /* mClusterDimY */ 1 -, /* mClusterDimZ */ 2 +, /* mClusterDimZ */ 3 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) , /* mDtypeAcc */ trtllm::gen::Dtype(1056776) , /* mDtypeA */ trtllm::gen::Dtype(17826818) @@ -48969,7 +54474,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(2) +, /* mEltwiseActType */ gemm::EltwiseActType(3) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 @@ -48989,7 +54494,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 1024 +, /* mK */ 1536 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) @@ -49007,7 +54512,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 -, /* mNumSlicesForSplitK */ 2 +, /* mNumSlicesForSplitK */ 3 , /* mNumStagesA */ 5 , /* mNumStagesB */ 5 , /* mNumStagesMma */ 2 @@ -49051,7 +54556,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 1024 +, /* mValidK */ 1536 , /* mWorldSize */ 1 , /* mActType */ gemmGatedAct::ActType(3) , /* mClampBeforeAct */ 1 @@ -49079,14 +54584,14 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 209816, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 640, "a8c2b9a87305e4db9bd307b211ce9a7e1b8de669d42bc1cbb0472494062b27f7", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 218008, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f", 640, "f4b5518071f1198191ec2411c0c81fbcc430c256518d90c6210a0fd91ab7e6f3", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 , /* mClusterDimX */ 1 , /* mClusterDimY */ 1 -, /* mClusterDimZ */ 2 +, /* mClusterDimZ */ 4 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) , /* mDtypeAcc */ trtllm::gen::Dtype(1056776) , /* mDtypeA */ trtllm::gen::Dtype(17826818) @@ -49095,7 +54600,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(3) +, /* mEltwiseActType */ gemm::EltwiseActType(0) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 @@ -49115,7 +54620,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 1024 +, /* mK */ 2048 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) @@ -49133,7 +54638,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 -, /* mNumSlicesForSplitK */ 2 +, /* mNumSlicesForSplitK */ 4 , /* mNumStagesA */ 5 , /* mNumStagesB */ 5 , /* mNumStagesMma */ 2 @@ -49177,15 +54682,15 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 1024 +, /* mValidK */ 2048 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(3) +, /* mActType */ gemmGatedAct::ActType(1) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 0 +, /* mFusedAct */ 1 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 @@ -49205,14 +54710,14 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 213912, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f", 640, "b11734bba44e6f5ed79d22dd39435b13822d5c77328b1fcc4012196e03e28600", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 218008, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f", 640, "e4dc0e64e6c9c6c32dec029b424c50ccd172b20be02ea249a0df0a66ed018da1", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 , /* mClusterDimX */ 1 , /* mClusterDimY */ 1 -, /* mClusterDimZ */ 3 +, /* mClusterDimZ */ 4 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) , /* mDtypeAcc */ trtllm::gen::Dtype(1056776) , /* mDtypeA */ trtllm::gen::Dtype(17826818) @@ -49241,7 +54746,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 1536 +, /* mK */ 2048 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) @@ -49259,7 +54764,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 -, /* mNumSlicesForSplitK */ 3 +, /* mNumSlicesForSplitK */ 4 , /* mNumStagesA */ 5 , /* mNumStagesB */ 5 , /* mNumStagesMma */ 2 @@ -49303,9 +54808,9 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 1536 +, /* mValidK */ 2048 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(1) +, /* mActType */ gemmGatedAct::ActType(2) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} @@ -49331,14 +54836,14 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 213912, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 640, "2270c140cdec40bf2431760668882ff95c6b0fe7a0c973be4c1a2b8cc20e7ae6", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 218008, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 640, "4eeb4f89971ff3efcb3281b3ada21dc77d79107a73422b4de4ec3224b15b1086", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 , /* mClusterDimX */ 1 , /* mClusterDimY */ 1 -, /* mClusterDimZ */ 3 +, /* mClusterDimZ */ 4 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) , /* mDtypeAcc */ trtllm::gen::Dtype(1056776) , /* mDtypeA */ trtllm::gen::Dtype(17826818) @@ -49367,7 +54872,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 1536 +, /* mK */ 2048 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) @@ -49385,7 +54890,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 -, /* mNumSlicesForSplitK */ 3 +, /* mNumSlicesForSplitK */ 4 , /* mNumStagesA */ 5 , /* mNumStagesB */ 5 , /* mNumStagesMma */ 2 @@ -49429,7 +54934,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 1536 +, /* mValidK */ 2048 , /* mWorldSize */ 1 , /* mActType */ gemmGatedAct::ActType(3) , /* mClampBeforeAct */ 1 @@ -49457,14 +54962,14 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 213912, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 640, "49d964be5275b46c2c12714bb37cd6231592551373ea2a9ee224b19918d7ffeb", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 218008, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 640, "6e90749b52a61d742477f06b2542de715a79aeef3da8e127a59c2789bfdc013d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 , /* mClusterDimX */ 1 , /* mClusterDimY */ 1 -, /* mClusterDimZ */ 3 +, /* mClusterDimZ */ 4 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) , /* mDtypeAcc */ trtllm::gen::Dtype(1056776) , /* mDtypeA */ trtllm::gen::Dtype(17826818) @@ -49493,7 +54998,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 1536 +, /* mK */ 2048 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) @@ -49511,7 +55016,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 -, /* mNumSlicesForSplitK */ 3 +, /* mNumSlicesForSplitK */ 4 , /* mNumStagesA */ 5 , /* mNumStagesB */ 5 , /* mNumStagesMma */ 2 @@ -49555,7 +55060,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseUnrollLoop2xForMma */ 0 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 1536 +, /* mValidK */ 2048 , /* mWorldSize */ 1 , /* mActType */ gemmGatedAct::ActType(3) , /* mClampBeforeAct */ 1 @@ -49583,14 +55088,140 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 218008, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f", 640, "0f9d41f79f0b633dc49b53969d065fd5c45b76bf23889b1c5fe38e87a040c3cb", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f_cubin_len, 162320, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f", 512, "15eb4c1668e341d323e5720f16d1addb8bd17c7bbdacfd410532a03b0adcbd46", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 , /* mClusterDimX */ 1 , /* mClusterDimY */ 1 -, /* mClusterDimZ */ 4 +, /* mClusterDimZ */ 1 +, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) +, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) +, /* mDtypeA */ trtllm::gen::Dtype(17826818) +, /* mDtypeB */ trtllm::gen::Dtype(17826818) +, /* mDtypeC */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) +, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) +, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) +, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEnablesEarlyExit */ 0 +, /* mEnablesDelayedEarlyExit */ 0 +, /* mEpilogueLdtmDps */ 16 +, /* mEpilogueLdtmBits */ 256 +, /* mEpilogueTileM */ 128 +, /* mEpilogueTileN */ 8 +, /* mFallbackClusterDimX */ 1 +, /* mFallbackClusterDimY */ 1 +, /* mFallbackClusterDimZ */ 1 +, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) +, /* mFuseLoadSfTask */ 0 +, /* mFuseUtccpWithUtcmma */ 0 +, /* mGridTriggerSecondaryA */ 0 +, /* mGridTriggerSecondaryB */ 1 +, /* mGridWaitForPrimaryEarlyExit */ 1 +, /* mGridWaitForPrimaryA */ 0 +, /* mGridWaitForPrimaryB */ 1 +, /* mHoistLoadTaskInit */ 1 +, /* mHoistMmaTaskTryWaits */ 0 +, /* mK */ 1024 +, /* mKernelTraits */ {} +, /* mLayoutA */ gemm::MatrixLayout(0) +, /* mLayoutB */ gemm::MatrixLayout(0) +, /* mM */ 256 +, /* mMmaK */ 64 +, /* mMmaKind */ trtllm::gen::MmaKind(4) +, /* mMmaM */ 128 +, /* mMmaN */ 8 +, /* mMmaTileK */ 512 +, /* mMockAllReduce */ 0 +, /* mN */ 256 +, /* mNumEpilogueWarps */ 4 +, /* mNumRegsCastAWarps */ 0 +, /* mNumRegsCopySfLdsSttm */ 0 +, /* mNumRegsCopySparsityInfo */ 0 +, /* mNumRegsPerThreadEpilogueWarp */ 0 +, /* mNumRegsPerThreadNonEpilogueWarp */ 0 +, /* mNumSlicesForSplitK */ 1 +, /* mNumStagesA */ 4 +, /* mNumStagesB */ 4 +, /* mNumStagesMma */ 1 +, /* mNumStagesMmaWithinWorkTile */ 1 +, /* mNumStagesMmaAcrossWorkTile */ 1 +, /* mNumStagesSmemSfA */ 4 +, /* mNumStagesSmemSfB */ 4 +, /* mNumStagesTmemSfA */ 4 +, /* mNumStagesTmemSfB */ 4 +, /* mNumStagesWorkId */ 3 +, /* mOutputDebugTensors */ 0 +, /* mPatchF2fp */ 0 +, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) +, /* mSchedHostTask */ gemm::SchedHostTask(0) +, /* mSfBlockSizeA */ 16 +, /* mSfBlockSizeB */ 16 +, /* mSfBlockSizeC */ 16 +, /* mSfLayoutA */ trtllm::gen::SfLayout(3) +, /* mSfLayoutB */ trtllm::gen::SfLayout(1) +, /* mSfLayoutC */ trtllm::gen::SfLayout(1) +, /* mSfReshapeFactor */ 1 +, /* mSparsityA */ trtllm::gen::Sparsity(0) +, /* mSplitK */ gemm::SplitK(0) +, /* mTileK */ 512 +, /* mTileM */ 128 +, /* mTileN */ 8 +, /* mTileScheduler */ gemm::TileScheduler(0) +, /* mTransposeMmaOutput */ 1 +, /* mUseCustomizedMma3xNvFp4 */ 0 +, /* mUseCustomMmaSchedule */ 1 +, /* mUseDeepSeekFp8 */ 0 +, /* mUseFlexibleClusterDims */ 0 +, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 +, /* mUseMaxTmemOverlap */ 0 +, /* mUsePerTokenSfA */ 0 +, /* mUsePerTokenSfB */ 0 +, /* mUseShuffledMatrix */ 1 +, /* mUseTmaStore */ 1 +, /* mUseTwoTmaLoadWarps */ 1 +, /* mUseTwoMmaWarps */ 0 +, /* mUseUnrollLoop2xForMma */ 1 +, /* mValidM */ 256 +, /* mValidN */ 256 +, /* mValidK */ 1024 +, /* mWorldSize */ 1 +, /* mActType */ gemmGatedAct::ActType(0) +, /* mClampBeforeAct */ 1 +, /* mBatchedM */ {} +, /* mBatchedN */ {} +, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) +, /* mBatchStrideInTokens */ -1 +, /* mFusedAct */ 0 +, /* mGridWaitForPrimaryRouting */ 1 +, /* mIsStaticBatch */ 1 +, /* mIsUniformNumTokensPerBatch */ 0 +, /* mNumBatches */ 2 +, /* mNumRegsPerThreadLoadA */ 0 +, /* mNumRegsPerThreadLoadB */ 0 +, /* mNumRegsPerThreadLoadSfA */ 0 +, /* mNumRegsPerThreadLoadSfB */ 0 +, /* mNumTokens */ 0 +, /* mNumWarpsLoadA */ 0 +, /* mNumWarpsLoadB */ 0 +, /* mNumWarpsLoadSfA */ 0 +, /* mNumWarpsLoadSfB */ 0 +, /* mRouteImpl */ batchedGemm::RouteImpl(0) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(0)} +, /* mPrefetchB */ 0 +, /* mUseCMultiCast */ 0 +, /* mUseTmaOobOpt */ 1 + }, gemm::SmVersion::Sm100f}, +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 202640, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f", 640, "280f003b6a4d520392f615b4c23fce3a6268d5453902206e085baa30247ce429", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +, /* mBiasDtype */ trtllm::gen::Dtype(1056776) +, /* mBiasType */ gemm::BiasType(1) +, /* mBlockK */ -1 +, /* mClcFastDrain */ 1 +, /* mClusterDimX */ 1 +, /* mClusterDimY */ 1 +, /* mClusterDimZ */ 1 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) , /* mDtypeAcc */ trtllm::gen::Dtype(1056776) , /* mDtypeA */ trtllm::gen::Dtype(17826818) @@ -49619,7 +55250,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 2048 +, /* mK */ 1024 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) @@ -49637,7 +55268,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 -, /* mNumSlicesForSplitK */ 4 +, /* mNumSlicesForSplitK */ 1 , /* mNumStagesA */ 5 , /* mNumStagesB */ 5 , /* mNumStagesMma */ 2 @@ -49660,7 +55291,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfLayoutC */ trtllm::gen::SfLayout(1) , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) -, /* mSplitK */ gemm::SplitK(2) +, /* mSplitK */ gemm::SplitK(0) , /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 8 @@ -49678,10 +55309,10 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseTmaStore */ 1 , /* mUseTwoTmaLoadWarps */ 1 , /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 0 +, /* mUseUnrollLoop2xForMma */ 1 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 2048 +, /* mValidK */ 1024 , /* mWorldSize */ 1 , /* mActType */ gemmGatedAct::ActType(1) , /* mClampBeforeAct */ 1 @@ -49709,14 +55340,14 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 218008, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 640, "ed0aa035dc9cd85be6daff01a91ca206df76da4cc225fa1d7cce96f15e0bb2b5", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 202640, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f", 640, "e8ca9ac801fb9797c231470ce424637e38aa762e5a42a2fcd4a0d668b2bdd765", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 , /* mClusterDimX */ 1 , /* mClusterDimY */ 1 -, /* mClusterDimZ */ 4 +, /* mClusterDimZ */ 1 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) , /* mDtypeAcc */ trtllm::gen::Dtype(1056776) , /* mDtypeA */ trtllm::gen::Dtype(17826818) @@ -49725,7 +55356,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(2) +, /* mEltwiseActType */ gemm::EltwiseActType(0) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 @@ -49745,7 +55376,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 2048 +, /* mK */ 1024 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) @@ -49763,7 +55394,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 -, /* mNumSlicesForSplitK */ 4 +, /* mNumSlicesForSplitK */ 1 , /* mNumStagesA */ 5 , /* mNumStagesB */ 5 , /* mNumStagesMma */ 2 @@ -49786,7 +55417,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfLayoutC */ trtllm::gen::SfLayout(1) , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) -, /* mSplitK */ gemm::SplitK(2) +, /* mSplitK */ gemm::SplitK(0) , /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 8 @@ -49804,18 +55435,18 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseTmaStore */ 1 , /* mUseTwoTmaLoadWarps */ 1 , /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 0 +, /* mUseUnrollLoop2xForMma */ 1 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 2048 +, /* mValidK */ 1024 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(3) +, /* mActType */ gemmGatedAct::ActType(2) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 0 +, /* mFusedAct */ 1 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 @@ -49835,14 +55466,14 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 218008, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 640, "844ad567857930d54e3495dabe4ba7f5579e2dfcec7b00279d101211bcd1d32c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 202640, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "70262e7f72c1faf21dfc307927b0ca601d7834518fce6393a8c3158e867ff4a6", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 , /* mClcFastDrain */ 1 , /* mClusterDimX */ 1 , /* mClusterDimY */ 1 -, /* mClusterDimZ */ 4 +, /* mClusterDimZ */ 1 , /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) , /* mDtypeAcc */ trtllm::gen::Dtype(1056776) , /* mDtypeA */ trtllm::gen::Dtype(17826818) @@ -49851,7 +55482,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(3) +, /* mEltwiseActType */ gemm::EltwiseActType(0) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 @@ -49871,7 +55502,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mGridWaitForPrimaryB */ 1 , /* mHoistLoadTaskInit */ 1 , /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 2048 +, /* mK */ 1024 , /* mKernelTraits */ {} , /* mLayoutA */ gemm::MatrixLayout(0) , /* mLayoutB */ gemm::MatrixLayout(0) @@ -49889,7 +55520,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsCopySparsityInfo */ 0 , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 -, /* mNumSlicesForSplitK */ 4 +, /* mNumSlicesForSplitK */ 1 , /* mNumStagesA */ 5 , /* mNumStagesB */ 5 , /* mNumStagesMma */ 2 @@ -49912,7 +55543,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfLayoutC */ trtllm::gen::SfLayout(1) , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) -, /* mSplitK */ gemm::SplitK(2) +, /* mSplitK */ gemm::SplitK(0) , /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 8 @@ -49930,18 +55561,18 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseTmaStore */ 1 , /* mUseTwoTmaLoadWarps */ 1 , /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 0 +, /* mUseUnrollLoop2xForMma */ 1 , /* mValidM */ 256 , /* mValidN */ 256 -, /* mValidK */ 2048 +, /* mValidK */ 1024 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(3) +, /* mActType */ gemmGatedAct::ActType(0) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 0 +, /* mFusedAct */ 1 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 @@ -49961,7 +55592,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f_cubin_len, 162320, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f", 512, "a6e8ad6d9deafbc1681352a2d7a1f1014e22936657050cde2fe7ad1aba04642a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 202640, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 640, "525ffa451d4421f878ab3586f0c4ae5cc01aa85470bf85b058813d7515c95445", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -49977,8 +55608,8 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(0) -, /* mEnablesEarlyExit */ 0 +, /* mEltwiseActType */ gemm::EltwiseActType(2) +, /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 , /* mEpilogueLdtmBits */ 256 @@ -50016,15 +55647,15 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumRegsPerThreadEpilogueWarp */ 0 , /* mNumRegsPerThreadNonEpilogueWarp */ 0 , /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 4 -, /* mNumStagesB */ 4 -, /* mNumStagesMma */ 1 +, /* mNumStagesA */ 5 +, /* mNumStagesB */ 5 +, /* mNumStagesMma */ 2 , /* mNumStagesMmaWithinWorkTile */ 1 -, /* mNumStagesMmaAcrossWorkTile */ 1 -, /* mNumStagesSmemSfA */ 4 -, /* mNumStagesSmemSfB */ 4 -, /* mNumStagesTmemSfA */ 4 -, /* mNumStagesTmemSfB */ 4 +, /* mNumStagesMmaAcrossWorkTile */ 2 +, /* mNumStagesSmemSfA */ 5 +, /* mNumStagesSmemSfB */ 5 +, /* mNumStagesTmemSfA */ 5 +, /* mNumStagesTmemSfB */ 5 , /* mNumStagesWorkId */ 3 , /* mOutputDebugTensors */ 0 , /* mPatchF2fp */ 0 @@ -50034,7 +55665,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mSfBlockSizeB */ 16 , /* mSfBlockSizeC */ 16 , /* mSfLayoutA */ trtllm::gen::SfLayout(3) -, /* mSfLayoutB */ trtllm::gen::SfLayout(1) +, /* mSfLayoutB */ trtllm::gen::SfLayout(0) , /* mSfLayoutC */ trtllm::gen::SfLayout(1) , /* mSfReshapeFactor */ 1 , /* mSparsityA */ trtllm::gen::Sparsity(0) @@ -50042,7 +55673,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 8 -, /* mTileScheduler */ gemm::TileScheduler(0) +, /* mTileScheduler */ gemm::TileScheduler(1) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 , /* mUseCustomMmaSchedule */ 1 @@ -50069,25 +55700,25 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mBatchStrideInTokens */ -1 , /* mFusedAct */ 0 , /* mGridWaitForPrimaryRouting */ 1 -, /* mIsStaticBatch */ 1 +, /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 -, /* mNumBatches */ 2 +, /* mNumBatches */ 128 , /* mNumRegsPerThreadLoadA */ 0 , /* mNumRegsPerThreadLoadB */ 0 , /* mNumRegsPerThreadLoadSfA */ 0 , /* mNumRegsPerThreadLoadSfB */ 0 -, /* mNumTokens */ 0 +, /* mNumTokens */ 2 , /* mNumWarpsLoadA */ 0 , /* mNumWarpsLoadB */ 0 , /* mNumWarpsLoadSfA */ 0 , /* mNumWarpsLoadSfB */ 0 -, /* mRouteImpl */ batchedGemm::RouteImpl(0) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(0)} +, /* mRouteImpl */ batchedGemm::RouteImpl(2) +, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} , /* mPrefetchB */ 0 , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 202640, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f", 640, "d455ef510ce22e1e5d0da160d8fa843140837fab205b2ee344848e67c584dc5e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 202640, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 640, "6da00071ecb658cc6d66cfcc1f4ba0a1e323be56b2744da1304eb0b7e5c4b27f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -50103,7 +55734,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) , /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) , /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(0) +, /* mEltwiseActType */ gemm::EltwiseActType(3) , /* mEnablesEarlyExit */ 1 , /* mEnablesDelayedEarlyExit */ 0 , /* mEpilogueLdtmDps */ 16 @@ -50187,13 +55818,13 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mValidN */ 256 , /* mValidK */ 1024 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(1) +, /* mActType */ gemmGatedAct::ActType(3) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} , /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) , /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 1 +, /* mFusedAct */ 0 , /* mGridWaitForPrimaryRouting */ 1 , /* mIsStaticBatch */ 0 , /* mIsUniformNumTokensPerBatch */ 0 @@ -50213,7 +55844,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 202640, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "577cd363c3fe17e885d6e989221c2408c9af34f164be4fe257b66e8dca289207", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 202256, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f", 512, "0f0026e618f31c9a706eb360fc593352f673a07ce786e78413379cd78a46547e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -50270,9 +55901,9 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mNumSlicesForSplitK */ 1 , /* mNumStagesA */ 5 , /* mNumStagesB */ 5 -, /* mNumStagesMma */ 2 +, /* mNumStagesMma */ 1 , /* mNumStagesMmaWithinWorkTile */ 1 -, /* mNumStagesMmaAcrossWorkTile */ 2 +, /* mNumStagesMmaAcrossWorkTile */ 1 , /* mNumStagesSmemSfA */ 5 , /* mNumStagesSmemSfB */ 5 , /* mNumStagesTmemSfA */ 5 @@ -50294,7 +55925,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mTileK */ 512 , /* mTileM */ 128 , /* mTileN */ 8 -, /* mTileScheduler */ gemm::TileScheduler(1) +, /* mTileScheduler */ gemm::TileScheduler(0) , /* mTransposeMmaOutput */ 1 , /* mUseCustomizedMma3xNvFp4 */ 0 , /* mUseCustomMmaSchedule */ 1 @@ -50313,7 +55944,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mValidN */ 256 , /* mValidK */ 1024 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(0) +, /* mActType */ gemmGatedAct::ActType(1) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} @@ -50339,259 +55970,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 202640, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 640, "61ab3c3eb6daa69463f1083f89dbbdee8cb194173b7a236a52c8216d4de57914", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) -, /* mBiasDtype */ trtllm::gen::Dtype(1056776) -, /* mBiasType */ gemm::BiasType(1) -, /* mBlockK */ -1 -, /* mClcFastDrain */ 1 -, /* mClusterDimX */ 1 -, /* mClusterDimY */ 1 -, /* mClusterDimZ */ 1 -, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) -, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) -, /* mDtypeA */ trtllm::gen::Dtype(17826818) -, /* mDtypeB */ trtllm::gen::Dtype(17826818) -, /* mDtypeC */ trtllm::gen::Dtype(17826818) -, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) -, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) -, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(2) -, /* mEnablesEarlyExit */ 1 -, /* mEnablesDelayedEarlyExit */ 0 -, /* mEpilogueLdtmDps */ 16 -, /* mEpilogueLdtmBits */ 256 -, /* mEpilogueTileM */ 128 -, /* mEpilogueTileN */ 8 -, /* mFallbackClusterDimX */ 1 -, /* mFallbackClusterDimY */ 1 -, /* mFallbackClusterDimZ */ 1 -, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) -, /* mFuseLoadSfTask */ 0 -, /* mFuseUtccpWithUtcmma */ 0 -, /* mGridTriggerSecondaryA */ 0 -, /* mGridTriggerSecondaryB */ 1 -, /* mGridWaitForPrimaryEarlyExit */ 1 -, /* mGridWaitForPrimaryA */ 0 -, /* mGridWaitForPrimaryB */ 1 -, /* mHoistLoadTaskInit */ 1 -, /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 1024 -, /* mKernelTraits */ {} -, /* mLayoutA */ gemm::MatrixLayout(0) -, /* mLayoutB */ gemm::MatrixLayout(0) -, /* mM */ 256 -, /* mMmaK */ 64 -, /* mMmaKind */ trtllm::gen::MmaKind(4) -, /* mMmaM */ 128 -, /* mMmaN */ 8 -, /* mMmaTileK */ 512 -, /* mMockAllReduce */ 0 -, /* mN */ 256 -, /* mNumEpilogueWarps */ 4 -, /* mNumRegsCastAWarps */ 0 -, /* mNumRegsCopySfLdsSttm */ 0 -, /* mNumRegsCopySparsityInfo */ 0 -, /* mNumRegsPerThreadEpilogueWarp */ 0 -, /* mNumRegsPerThreadNonEpilogueWarp */ 0 -, /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 5 -, /* mNumStagesB */ 5 -, /* mNumStagesMma */ 2 -, /* mNumStagesMmaWithinWorkTile */ 1 -, /* mNumStagesMmaAcrossWorkTile */ 2 -, /* mNumStagesSmemSfA */ 5 -, /* mNumStagesSmemSfB */ 5 -, /* mNumStagesTmemSfA */ 5 -, /* mNumStagesTmemSfB */ 5 -, /* mNumStagesWorkId */ 3 -, /* mOutputDebugTensors */ 0 -, /* mPatchF2fp */ 0 -, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) -, /* mSchedHostTask */ gemm::SchedHostTask(0) -, /* mSfBlockSizeA */ 16 -, /* mSfBlockSizeB */ 16 -, /* mSfBlockSizeC */ 16 -, /* mSfLayoutA */ trtllm::gen::SfLayout(3) -, /* mSfLayoutB */ trtllm::gen::SfLayout(0) -, /* mSfLayoutC */ trtllm::gen::SfLayout(1) -, /* mSfReshapeFactor */ 1 -, /* mSparsityA */ trtllm::gen::Sparsity(0) -, /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 512 -, /* mTileM */ 128 -, /* mTileN */ 8 -, /* mTileScheduler */ gemm::TileScheduler(1) -, /* mTransposeMmaOutput */ 1 -, /* mUseCustomizedMma3xNvFp4 */ 0 -, /* mUseCustomMmaSchedule */ 1 -, /* mUseDeepSeekFp8 */ 0 -, /* mUseFlexibleClusterDims */ 0 -, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 -, /* mUseMaxTmemOverlap */ 0 -, /* mUsePerTokenSfA */ 0 -, /* mUsePerTokenSfB */ 0 -, /* mUseShuffledMatrix */ 1 -, /* mUseTmaStore */ 1 -, /* mUseTwoTmaLoadWarps */ 1 -, /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 1 -, /* mValidM */ 256 -, /* mValidN */ 256 -, /* mValidK */ 1024 -, /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(0) -, /* mClampBeforeAct */ 1 -, /* mBatchedM */ {} -, /* mBatchedN */ {} -, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) -, /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 0 -, /* mGridWaitForPrimaryRouting */ 1 -, /* mIsStaticBatch */ 0 -, /* mIsUniformNumTokensPerBatch */ 0 -, /* mNumBatches */ 128 -, /* mNumRegsPerThreadLoadA */ 0 -, /* mNumRegsPerThreadLoadB */ 0 -, /* mNumRegsPerThreadLoadSfA */ 0 -, /* mNumRegsPerThreadLoadSfB */ 0 -, /* mNumTokens */ 2 -, /* mNumWarpsLoadA */ 0 -, /* mNumWarpsLoadB */ 0 -, /* mNumWarpsLoadSfA */ 0 -, /* mNumWarpsLoadSfB */ 0 -, /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} -, /* mPrefetchB */ 0 -, /* mUseCMultiCast */ 0 -, /* mUseTmaOobOpt */ 1 - }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 202640, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 640, "a193bafea5abf0e74bef83f59f93ab63d3319a7c106b4b479d58c10cb2f135ce", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) -, /* mBiasDtype */ trtllm::gen::Dtype(1056776) -, /* mBiasType */ gemm::BiasType(1) -, /* mBlockK */ -1 -, /* mClcFastDrain */ 1 -, /* mClusterDimX */ 1 -, /* mClusterDimY */ 1 -, /* mClusterDimZ */ 1 -, /* mCtaSwizzleType */ gemm::CtaSwizzleType(0) -, /* mDtypeAcc */ trtllm::gen::Dtype(1056776) -, /* mDtypeA */ trtllm::gen::Dtype(17826818) -, /* mDtypeB */ trtllm::gen::Dtype(17826818) -, /* mDtypeC */ trtllm::gen::Dtype(17826818) -, /* mDtypeMmaA */ trtllm::gen::Dtype(17826818) -, /* mDtypeMmaB */ trtllm::gen::Dtype(17826818) -, /* mDtypeSfC */ trtllm::gen::Dtype(1050629) -, /* mEltwiseActType */ gemm::EltwiseActType(3) -, /* mEnablesEarlyExit */ 1 -, /* mEnablesDelayedEarlyExit */ 0 -, /* mEpilogueLdtmDps */ 16 -, /* mEpilogueLdtmBits */ 256 -, /* mEpilogueTileM */ 128 -, /* mEpilogueTileN */ 8 -, /* mFallbackClusterDimX */ 1 -, /* mFallbackClusterDimY */ 1 -, /* mFallbackClusterDimZ */ 1 -, /* mFusedBiasShuffleMode */ gemm::FusedBiasShuffleMode(0) -, /* mFuseLoadSfTask */ 0 -, /* mFuseUtccpWithUtcmma */ 0 -, /* mGridTriggerSecondaryA */ 0 -, /* mGridTriggerSecondaryB */ 1 -, /* mGridWaitForPrimaryEarlyExit */ 1 -, /* mGridWaitForPrimaryA */ 0 -, /* mGridWaitForPrimaryB */ 1 -, /* mHoistLoadTaskInit */ 1 -, /* mHoistMmaTaskTryWaits */ 0 -, /* mK */ 1024 -, /* mKernelTraits */ {} -, /* mLayoutA */ gemm::MatrixLayout(0) -, /* mLayoutB */ gemm::MatrixLayout(0) -, /* mM */ 256 -, /* mMmaK */ 64 -, /* mMmaKind */ trtllm::gen::MmaKind(4) -, /* mMmaM */ 128 -, /* mMmaN */ 8 -, /* mMmaTileK */ 512 -, /* mMockAllReduce */ 0 -, /* mN */ 256 -, /* mNumEpilogueWarps */ 4 -, /* mNumRegsCastAWarps */ 0 -, /* mNumRegsCopySfLdsSttm */ 0 -, /* mNumRegsCopySparsityInfo */ 0 -, /* mNumRegsPerThreadEpilogueWarp */ 0 -, /* mNumRegsPerThreadNonEpilogueWarp */ 0 -, /* mNumSlicesForSplitK */ 1 -, /* mNumStagesA */ 5 -, /* mNumStagesB */ 5 -, /* mNumStagesMma */ 2 -, /* mNumStagesMmaWithinWorkTile */ 1 -, /* mNumStagesMmaAcrossWorkTile */ 2 -, /* mNumStagesSmemSfA */ 5 -, /* mNumStagesSmemSfB */ 5 -, /* mNumStagesTmemSfA */ 5 -, /* mNumStagesTmemSfB */ 5 -, /* mNumStagesWorkId */ 3 -, /* mOutputDebugTensors */ 0 -, /* mPatchF2fp */ 0 -, /* mPerTokenSfDtype */ trtllm::gen::Dtype(1052672) -, /* mSchedHostTask */ gemm::SchedHostTask(0) -, /* mSfBlockSizeA */ 16 -, /* mSfBlockSizeB */ 16 -, /* mSfBlockSizeC */ 16 -, /* mSfLayoutA */ trtllm::gen::SfLayout(3) -, /* mSfLayoutB */ trtllm::gen::SfLayout(0) -, /* mSfLayoutC */ trtllm::gen::SfLayout(1) -, /* mSfReshapeFactor */ 1 -, /* mSparsityA */ trtllm::gen::Sparsity(0) -, /* mSplitK */ gemm::SplitK(0) -, /* mTileK */ 512 -, /* mTileM */ 128 -, /* mTileN */ 8 -, /* mTileScheduler */ gemm::TileScheduler(1) -, /* mTransposeMmaOutput */ 1 -, /* mUseCustomizedMma3xNvFp4 */ 0 -, /* mUseCustomMmaSchedule */ 1 -, /* mUseDeepSeekFp8 */ 0 -, /* mUseFlexibleClusterDims */ 0 -, /* mUseHoistTryWaitForCustomMmaSchedule */ 0 -, /* mUseMaxTmemOverlap */ 0 -, /* mUsePerTokenSfA */ 0 -, /* mUsePerTokenSfB */ 0 -, /* mUseShuffledMatrix */ 1 -, /* mUseTmaStore */ 1 -, /* mUseTwoTmaLoadWarps */ 1 -, /* mUseTwoMmaWarps */ 0 -, /* mUseUnrollLoop2xForMma */ 1 -, /* mValidM */ 256 -, /* mValidN */ 256 -, /* mValidK */ 1024 -, /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(3) -, /* mClampBeforeAct */ 1 -, /* mBatchedM */ {} -, /* mBatchedN */ {} -, /* mBatchMode */ batchedGemm::BatchedGemmOptions::BatchMode(1) -, /* mBatchStrideInTokens */ -1 -, /* mFusedAct */ 0 -, /* mGridWaitForPrimaryRouting */ 1 -, /* mIsStaticBatch */ 0 -, /* mIsUniformNumTokensPerBatch */ 0 -, /* mNumBatches */ 128 -, /* mNumRegsPerThreadLoadA */ 0 -, /* mNumRegsPerThreadLoadB */ 0 -, /* mNumRegsPerThreadLoadSfA */ 0 -, /* mNumRegsPerThreadLoadSfB */ 0 -, /* mNumTokens */ 2 -, /* mNumWarpsLoadA */ 0 -, /* mNumWarpsLoadB */ 0 -, /* mNumWarpsLoadSfA */ 0 -, /* mNumWarpsLoadSfB */ 0 -, /* mRouteImpl */ batchedGemm::RouteImpl(2) -, /* mRouteSfsImpl */ {batchedGemm::RouteImpl(2)} -, /* mPrefetchB */ 0 -, /* mUseCMultiCast */ 0 -, /* mUseTmaOobOpt */ 1 - }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f_cubin_len, 202256, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f", 512, "5e6eafbc9ad680c492939a0232e620663c296966bc4abdbbe168ee3210b1cdeb", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 202256, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "91e50932eb4666ff70364ca177a99a8a1a35e1b5bfcc82275509fd2bbcf884a9", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -50691,7 +56070,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mValidN */ 256 , /* mValidK */ 1024 , /* mWorldSize */ 1 -, /* mActType */ gemmGatedAct::ActType(1) +, /* mActType */ gemmGatedAct::ActType(2) , /* mClampBeforeAct */ 1 , /* mBatchedM */ {} , /* mBatchedN */ {} @@ -50717,7 +56096,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 202256, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "10ce69b50d326f27acf8e8f4b693cfd93828d58126aaa6875af68eaa1660b870", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 202256, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "24a8f8a1f638e8ff9543e6ae1aac00128df88f42a8903f999cb376bee40dab5e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -50843,7 +56222,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 202256, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 512, "fa350271b7344b051e3272de089634f2da917e2bf0824eed396438ac3c83a917", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 202256, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 512, "f3f5dabc8dfabbc1860f0743ad04abce8861bb4afaed8cc31a7ffdc0f506004a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -50969,7 +56348,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 202256, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 512, "0bc68323c7b397db47d89002662187052228e685a82b5225601587c0f6018d82", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 202256, "bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 512, "d8cf17651942549da3bea2b1d355f3339f90ae936bee5d0c58cc31c9f61c9af5", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -51095,7 +56474,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin_len, 211608, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f", 512, "fc51c0a95cfd3e802b38209cdbf66dcf6a38b386d0c471353b2f19b5e29f6054", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin_len, 211608, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f", 512, "ec51053286dbce2acc884b44d03d949c4c1171815e83b500df6e40928b9cd0ad", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -51221,7 +56600,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len, 211608, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f", 512, "6afdb9bdba1a85243cd120e70fa559be922e164cf400d832b0888452f9087da3", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len, 211608, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f", 512, "00ca60dca083632669a1044ddfaab2a10795dde742d59cf6d548aa2ffd59e318", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -51347,7 +56726,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len, 211608, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f", 512, "7a3f93f23989911d91a1a4ab50371583bb8759ba59200ada92df327054e83c20", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len, 211608, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f", 512, "1d7dd5a0f1c67de7920b6b775c0a49f5678c26bebe71719a24058c5a6413d50f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -51473,7 +56852,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128u2_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128u2_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin_len, 211608, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128u2_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f", 512, "c09b3b2295588859844185cdef51519eb3ecac84bd9da0a30554a7c073e3081e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128u2_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128u2_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin_len, 211608, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128u2_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f", 512, "52591da0e217e6a4900366dc99ac170f36e3a0b8573112e9ba6d05a666c615e7", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -51599,7 +56978,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128u2_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128u2_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len, 211608, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128u2_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f", 512, "a8ad26148d175ab05b90dd1a2ed864478c3e73275fa3e97b3ab575ed4a038937", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128u2_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128u2_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len, 211608, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128u2_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f", 512, "83df2fb1a6a6a9d622a8c77d334537a443410a976ce7cca0f19348799020187b", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -51725,7 +57104,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128u2_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128u2_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len, 211608, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128u2_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f", 512, "150f8c85ce34d0b9cb1beb3cce40c6d03089e283146cbb72690576ec22a95a5f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128u2_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128u2_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len, 211608, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128u2_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f", 512, "88f7eb11db3dc73d1716aea543ae25d7b0b3acd1276e704f9fb7d0df650e9341", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -51851,7 +57230,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 227984, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "8f76f0096f094e61312a1e8d51ce214a0f49e99002c9b69fcbe3e0e944ae1f37", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 227984, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "585b7f4b35f3806aa0c941ee9760597d15a2fae59e08da7e572fced3f3415d61", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -51977,7 +57356,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 227984, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 384, "7d614055d7f11311509b2fbbf7242ef282e03e0e5c1c2ef8d5de7b097b6b58c7", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 227984, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 384, "0b4a16e153bf3f1a503d9ce9f989c05a47a1b7d39e09a1d5c4c30d9e5a84b678", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -52103,7 +57482,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 227984, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 384, "80cf474086aa5b979ceb07f952810ad94cf47b70305ec21f8fcf02c87a0d5738", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 227984, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 384, "7b6f62534001f027ac22d8c07a370a919ec090334dfd8946f1f139eea1b35e92", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -52229,7 +57608,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 227600, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "39245cc67700b334e47c548603fa467a4556d276c4e6390194d5622ae476cbda", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 227600, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "caf79706766b81eb56efac5d13d866a2a054b2a288843755d1c175e1f7681c41", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -52355,7 +57734,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 227600, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 384, "2266fb7136ce1b15805f753f65f58efa66a3a1c8176ccc9f383288d39485f128", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 227600, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 384, "9fd5d67212f207fc48991932094b0284d0583d8f95ff549b9b2490ba003bee63", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -52481,7 +57860,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 227600, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 384, "d7320c17701cd6fe804d88afbd5c75717148f510ce1aebab28b2ead935ce89d3", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 227600, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 384, "7a99e4d608882e05280967f7acf3b396273aba6df87e677f082a17a6f18df0c3", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -52607,7 +57986,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 227984, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "ddf919e2754b6656a8073293f7f76bf35e71f11f343a25892022daade06ef55c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 227984, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "ea6c7ef13a86f0e140ba8fe5e787cd699bc6fc9355f407882fc17a89e4d7e744", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -52733,7 +58112,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 227984, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 384, "49c8cb5b367c213d2e46b0a058e076081a4ac3a04312234879159a36463935e2", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 227984, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 384, "7b6a10f0d78c25e88cb7ed9b20bd42b5d69de38822a6e47091411718e1ad50bc", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -52859,7 +58238,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 227984, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 384, "d07e26cf4e81049355cc54424b570fc8ca0c5db5c8b4b5e4352f6a7fa3d03893", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 227984, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 384, "0daefaadd757d0d7566fe810280318bbdefbe9f0602e968cb6451e0b4e6ef4f4", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -52985,7 +58364,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 227600, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "cab262e5c0325a9e883e1009861383edc174b3188b82bbd7737efabde477e002", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 227600, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "73a4f316f6a1bd7318f733d4d43bf31c8afd3c291e248442d6cd1af591418152", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -53111,7 +58490,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 227600, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 384, "36a9a420bf4979f362dd6cd6ebaaa1c716b4e5d4d66cfca88673df202fb12273", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 227600, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 384, "d1f955decd351099bf61fe861934f05032e04607e83c1ae69cc5ccbca4d9c1ca", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -53237,7 +58616,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 227600, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 384, "4a3fcf2da0bc5c5e77f9bfa9593570a7fc4a8d4da9c893014d1b67e5b0b182cb", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 227600, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 384, "0e1a9a921dd80f71bdd5a44d884c74946864dc2d127da9bbcf25c7e8ad1b4c61", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -53363,7 +58742,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin_len, 212632, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f", 512, "76d4d29c78f890ddbb84efaedd7f5ad78c6ddb87e800326c47c195a084ac523b", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin_len, 212632, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f", 512, "3d288aeb3792d968329187ebbd51a77a3ee850321b5fd057b1aaf4abc81edeb7", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -53489,7 +58868,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len, 212632, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f", 512, "03772bb38101eb93b72db615f5b2f5a33942df0af665b37f229b34f72b67ae5f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len, 212632, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f", 512, "e8281fe4ad0e111f553687abaabe677c399a38334cc9fd0c8fc88b4f21bb3fba", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -53615,7 +58994,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len, 212632, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f", 512, "946fe09d34006b24361bcaf97718bd15a74cad663b31ce93d5d1d4dfcdd29fe0", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len, 212632, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f", 512, "d4db7f835769c385171ceb5b918996bb954afa96ca671afe9184422d7df9e6ee", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -53741,7 +59120,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin_len, 212632, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f", 512, "7077c6bc9f285fa109bda99fa5d2ee717fb5d80bead361d82ab5f581e410e8ac", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin_len, 212632, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f", 512, "d34eb055f217fba0ab85aeef239960b358b1bc98c217e558b4537dbfe3415a1a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -53867,7 +59246,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len, 212632, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f", 512, "bc52798304db4812a0b753052fc106b1a68c187d0b33c00477f2e964314e95bc", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len, 212632, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f", 512, "90a184d1fe48ec19c0f46a2f529c82ae77fa5fd88150159a453d2403152891fc", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -53993,7 +59372,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len, 212632, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f", 512, "65f82e2bf3233efd02611a1ae7fea44c55f20ea1d0f502c73b0d922e45b2b9c0", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len, 212632, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f", 512, "4f4c14e879fa4f6f65180b234a90551371eb591bccd1ece1f68c17bb267a2824", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -54119,7 +59498,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin_len, 221848, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f", 640, "fdc9014b785bcfac8cfd0e26b78b17c04cb9741c6e4b3b385856db58ae7305d6", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin_len, 221848, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f", 640, "162dcf05a068d2b0b4038afdb2af5ed6454972ed7f9de059870a7a3c53fd824c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -54245,7 +59624,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len, 221848, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f", 640, "2e92fe6a367588a5d8c92808e8ef1efdd97df734cbf25b59cbca3d8dabe2ba32", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len, 221848, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f", 640, "9e9474548b7366a0e604c2befe3db3e479c04ced78c39d1683344789e6caf195", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -54371,7 +59750,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len, 221848, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f", 640, "61dc36f33872049253d07fb80460cb5f698979981e0a6f2b8644c430aab03e5a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len, 221848, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f", 640, "cc811310e8bbae8c6474901a54a060f646f89a7596ceab3c88414abcad388454", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -54497,7 +59876,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128u2_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128u2_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin_len, 221848, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128u2_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f", 640, "7e2b4908512566eeeaed1597de374d362963fe051d87561731a06683a22c9541", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128u2_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128u2_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin_len, 221848, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128u2_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f", 640, "33b9fb3e7db3bc9f8671c81e727c8fbdfc1d8905bdd4976cec471c1feb16de3e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -54623,7 +60002,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128u2_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128u2_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len, 221848, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128u2_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f", 640, "ad9eef10e895de07089df2a669414eb185839e868bfa85b0a8796dfa4eda0921", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128u2_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128u2_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len, 221848, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128u2_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f", 640, "b853f43692a79ffae0430cc48932eb57975a1b602d59e595937ff554bbe3fd7a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -54749,7 +60128,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128u2_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128u2_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len, 221848, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128u2_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f", 640, "24e99ea5bd232c97435d1b68f404f2e5d986668f5073b2035ebcd93c8fe6c975", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128u2_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128u2_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len, 221848, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128u2_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f", 640, "a49f26e3387491a9ba7e514821dec65b56d2d2af1d2efdb7e5a6d7e864f491c4", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -54875,7 +60254,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 213648, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "37cf107cd46d1eab1a55a1bcdd54b60c63b21964fa2c79bfc0942247235d6dfb", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 213648, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "b7ddc1b25174d85eba2a5059b31d7b31dd559cc4b860a91701c4b885b5debac6", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -55001,7 +60380,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 213648, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 384, "56d855c64f058519843810ecc01b67b5870da43b227ecd480ad673bc7fa19a26", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 213648, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 384, "758077c581eee7a2845382913a8eb9a5ec5d40008305f9ff90d3e200accaba06", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -55127,7 +60506,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 213648, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 384, "c1e2d42ecaa83eddd697b58768f3213e74674ba4c3fb25d75fa24ec9bd86dec8", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 213648, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 384, "bd99c069dc3ccb9a9bfd38281ae76eb3bb338f6530ea2b1160cfd6149878641f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -55253,7 +60632,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 213264, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "e7610fcd50a47baedf228bc2a5ae1275112b24d887dee5cc986f8a6f14b740a4", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 213264, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "926ac0d967c19a755693374879f760120994c62a6797857f7bd50f8604355ac9", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -55379,7 +60758,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 213264, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 384, "4e807026b0aa8b698a50227a3f7943b07656ee840b130780771719a39849257c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 213264, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 384, "b94923d70449ecd29bf4662e278ab8159b1aebc4653d73c78ca61df32a6842de", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -55505,7 +60884,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 213264, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 384, "9085c9f40f11b2a693022c9fca020ae3c9dc603992c2ad6cd5b975a9c390323d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 213264, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 384, "6eeb8191ed87cf8ae05df93f9ce1f4d0be77444739309846c68c5e6cc5633d13", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -55631,7 +61010,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 213648, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "5fc2e7e21046490b05d9af031aefc0951e439372b8781ac89079628dbfabaa49", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 213648, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "588ca0eef05a13a32c19ce50e97ba11466b7cbd659c62a3400a50789c67e1b22", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -55757,7 +61136,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 213648, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 384, "a7eb061f80e9a9de17ee0143ec8575bd495bf8918737153fa86bc967357487eb", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 213648, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 384, "25a6995d689868081d7e2a1285b15e7704e4ecb0b0fa54d71e44f11cd2e38f06", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -55883,7 +61262,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 213648, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 384, "d876bb0bad5856bb485155efd9452ab69002f2dd19cae0e12bd27eddc501ab50", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 213648, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 384, "4328d03d9017b6858ffd30c7dd645b659f6394e5e19dd7c67d7ec8e6f266596e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -56009,7 +61388,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 213264, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "9e4d2fd79819ff453e8d0e9001d48f7dd613b490050ce6038e0d350ce778cbbd", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 213264, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "393cb543955eee4b900b1e5241b79f439821565daf07cbb3fa5ae7bd9488a476", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -56135,7 +61514,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 213264, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 384, "79aec59ac0d1b0ee3164f5380d097aef638b2fe9804b2987c7b7b7d3b38815a1", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 213264, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 384, "10759df4736e56c37f40e9fd1ac3163215167fdfe10628a4629f0423109e60da", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -56261,7 +61640,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 213264, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 384, "391deb3926e48e02d7430b62ff2eda73879708ad737af9510de4df0959a08499", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 213264, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 384, "3a1668e96ce24763a9dcf4c026b4310c8b4feee7759e586351af5a0a49d9e7cc", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -56387,7 +61766,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin_len, 218776, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f", 512, "a12ebc1ebd821213c4dee283035d6fa83d431b2c30e0597483c36c3adbb88e7a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin_len, 218776, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f", 512, "f699f4ee06901a3e3e238b4739d4eba49afc2180aab288fdfab3b5b38ace616b", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -56513,7 +61892,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len, 218776, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f", 512, "31a1f45c0a05a017d0cfd239b803ac09b76590dc11da7cae293aaa5c3bde3296", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len, 218776, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f", 512, "02aab39fd362c785ff0ee02d2e768fc3a6224f59799652abb8bd3987cd22e928", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -56639,7 +62018,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len, 218776, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f", 512, "cac912408944b824f9c609f91900eca3896c8d3f58c3db0853f8383014c81ed6", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len, 218776, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f", 512, "4b4e25cb11e5fdc0ac0dae8b1c180777ee3af627bc88b6abc830a61bc96e5e07", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -56765,7 +62144,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256u2_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256u2_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin_len, 218776, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256u2_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f", 512, "857d48b24d6eee9f9182b8126072dfb795f86cb03e4d6c7f79355606eeabbeed", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256u2_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256u2_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f_cubin_len, 218776, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256u2_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f", 512, "b7dbfba88c4711dc3a8a093d4f71580fe74ac32ae3a142232c563ef563be8448", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -56891,7 +62270,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256u2_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256u2_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len, 218776, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256u2_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f", 512, "aa4a8ee86a152a34253e141ef00d0ff4e03b8ddaf38b3f54777a0266d71a48de", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256u2_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256u2_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len, 218776, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256u2_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f", 512, "c52925f378affce9be8b33dffa6a700987eeff63e7e17b97881cca75f3a7e53c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -57017,7 +62396,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256u2_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256u2_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len, 218776, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256u2_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f", 512, "077a992d85bf706fb7077cda80522956e818b51e5167e6f60a999af8789dfc71", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256u2_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256u2_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f_cubin_len, 218776, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256u2_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f", 512, "631c5334b138a7bcd30cfb52017d62ecc78b7329ea29b9cbd47c88911a7c8568", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -57143,7 +62522,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 214672, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "5b261a22492fce7866eced26de55c54eeb0809b7bff68af57ccc042cb571e851", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 214672, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "4e72ed3aa2034b053910c3bae5e15adfe25b5a4efe86a862a20c5b2103d15f2d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -57269,7 +62648,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 214672, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 384, "d100e6c020454b2778d3aa48d37e7add2ac953b18dee78c11bb46248319378df", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 214672, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 384, "16e1f2e8954001b6b0001f978a187d6a99542586f7edaf3a2f6fe278fbb316a0", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -57395,7 +62774,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 214672, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 384, "a57a01790bedb57d817f0d3fcf757a76e76d506adc4ca96bdbe17a0518de498d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 214672, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 384, "83464adfa1ee948b932f6125d6bfdf331e72123173642b3d48c70e125896d8c3", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -57521,7 +62900,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 214288, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 256, "b41ba6ad67e121f7f585d846cfa115e68b5f1a66a277503f559fa44734c51aec", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 214288, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 256, "08d86704b463cecbe7c219aa64868dddaef53621defaff171722c9aa2fbad444", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -57647,7 +63026,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 214288, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 256, "da869d348ad807a58b05e17e15d7161a2aae734d298fd1fcedeab0f29d9e7a6e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 214288, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 256, "df42d5e7d0124263833bf168636001128a713540c45344ddea1b4da51f24a0b8", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -57773,7 +63152,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 214288, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 256, "614e84f7ae5b54079c93f17a3eb5cf9646fa16cd9fc148cae71e3d8d34ff8114", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 214288, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 256, "4887284c990b88f82f23602647890edc393edd403345f688ba304d1e83334e4f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -57899,7 +63278,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 214672, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "7d2f20893e0a8290d0550f4082eaa70a13ce896a16da3007cb5c3d3e4b2f6f5c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 214672, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "49526b56ae636a8391fae15c28bc160e40381e76b59d89b1deb37baa3d1089b8", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -58025,7 +63404,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 214672, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 384, "7f7882e82b41eb02631f79ef08d5f602d1ae1a1a0f555badaee7a21debb1d351", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 214672, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 384, "3a31a05bf793063956433e9ef5e700739a142367a03d4de2220d07543b1147f8", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -58151,7 +63530,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 214672, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 384, "2aeb1112f5aa9c28c2dcb15c07a8750e14c397da6723f50a4167c4c3632dee66", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 214672, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 384, "4f4dc5697e1fddf41c783d2506b0c977d323f7f997d63ae30b611d22716eadb6", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -58277,7 +63656,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 214288, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 256, "2bfaf6596001e5cee1146b9d505637ec817a747467e86083133ba01f138cef3d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 214288, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f", 256, "aa8f68d109f636a773861f210aee09bec7c9a3292caae57e90f37fda1df7ed45", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -58403,7 +63782,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 214288, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 256, "340995ff338f29ba51877545102e40f17e675c2abeb004c96721e493c259e626", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 214288, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 256, "f7b1168765a310c0b436acc91f63b726c1f0c3592d07f017eedd9459649b9d27", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -58529,7 +63908,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 214288, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 256, "bd71e58feb4eaadc7a21b8f48e13856a5429488a6778e446b3c0bd856ae9f076", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f_cubin_len, 214288, "bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f", 256, "b98f54f6ae522bdf3599945476859f9ca42455876258964ced88dc462d371d8b", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -58655,7 +64034,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin_len, 190640, "bmm_E4m3_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f", 512, "2e9ce134618efa3af5bf0a262c6a4f3f6ca0418e5ddb62df58e67e977cc9e66e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin_len, 190640, "bmm_E4m3_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f", 512, "5210d1de9ff1f35b596ccc4316911457a61cb7bd59772bb4fc5bd5aa509ac9c1", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -58781,7 +64160,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin_len, 190512, "bmm_E4m3_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f", 512, "2d78d641c3f5ce6db17758e95bfb385fed331c52c99736571275f37e14a36d85", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin_len, 190512, "bmm_E4m3_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f", 512, "27ef80aab43393145654ff4622b701e8039279e224da72ed006bc583dc4149c5", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -58907,7 +64286,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin_len, 190640, "bmm_E4m3_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f", 512, "693e46559617e718787478386e64aaa2e724e0d08737512d9865422ab53fa4a4", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin_len, 190640, "bmm_E4m3_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f", 512, "8522a34b25a63c7daa09c817ac317c3cac2f6e39144c7f7c2ed2ed819f9c760c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -59033,7 +64412,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin_len, 190512, "bmm_E4m3_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f", 512, "0cf25b4921cc625dcf8969549bdc3f98c6d9374385ce00c36d1814ee37ec2897", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin_len, 190512, "bmm_E4m3_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f", 512, "b6f7c936daad888d3f85e62647d2a777191f32c3da20834ef2780b916d0475fc", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -59159,7 +64538,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len, 119216, "bmm_E4m3_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f", 512, "1122f6b890a381a4c89f559f1f705eae99e60a08e90ac30b9740a20bf7e27457", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len, 119216, "bmm_E4m3_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f", 512, "143178aeb025534a337f2c6056769424cdffe14abffeca84fbfe5852754d983d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -59285,7 +64664,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len, 119088, "bmm_E4m3_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f", 512, "fad190b3f3d5bfb8afe036f0191034dc013c42f416c389d884268fac776acb55", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len, 119088, "bmm_E4m3_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f", 512, "d4edb911700b52705b13e4c0e9f39bdb544334b39d4401a4cbb79a2584eaab6e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -59411,7 +64790,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len, 119216, "bmm_E4m3_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f", 512, "bd93e2de6838acf1452469e616e630aec20516501b208f59e3954f35ca2fdf67", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len, 119216, "bmm_E4m3_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f", 512, "1098c9bc9cbb593f203ae54e4315590a5515c466b0456f6005eff564c2e1c77c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -59537,7 +64916,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len, 119088, "bmm_E4m3_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f", 512, "07bf927187a4a1ea04c898725c464573d462196dafe7b602dea0fa59c6fed7a4", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len, 119088, "bmm_E4m3_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f", 512, "4f6323ece0e184d3a9445bc0af7909e55def0e9184bf7b58c27d63c73adad340", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -59663,7 +65042,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len, 133552, "bmm_E4m3_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f", 512, "b6e0ebf9c8018d828c165c7508a5908000392a1eb5fbb3cc39b02c2ceae513b2", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len, 133552, "bmm_E4m3_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f", 512, "b2f7db3f347cb72dd1922bf511807814fa57670c27aecd3d2c8d54e71f530f84", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -59789,7 +65168,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len, 133424, "bmm_E4m3_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f", 512, "169f826d3a4120b737d7062c92efa80371c4159cedfedd5e007368126cea5774", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len, 133424, "bmm_E4m3_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f", 512, "7db24cd707ad88c57f36d876f4432f8ae4139ecf0b0121f7bf2dccf1a649de11", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -59915,7 +65294,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len, 133552, "bmm_E4m3_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f", 512, "19beda82bfcdd907cdc2f86073509944facbdf980f649ae8c81cfb954b80bfe9", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len, 133552, "bmm_E4m3_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f", 512, "25cac2c1f02662414a278ec3b415d0aefe5ded27f8664dc354b11780231347e4", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -60041,7 +65420,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len, 133424, "bmm_E4m3_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f", 512, "ba7f46aa74978bed02fa28f71ea0a0aedebcda0d065d8c881918d453f20b42b9", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len, 133424, "bmm_E4m3_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f", 512, "7de619c26e4efcd6a85be33bc6ec711168c8672b77e9e3e67041c7a5e4b73993", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -60167,7 +65546,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin_len, 164016, "bmm_E4m3_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f", 512, "c304b5eb779f1b0ed2e7e76ed21286f81ddf55fb61546ae2d1f69c7348e7df04", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin_len, 164016, "bmm_E4m3_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f", 512, "582fe7ef4539910de8b36a98d4a1aa97f007626ced8f9aa9e84d8ae9c593775f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -60293,7 +65672,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin_len, 163888, "bmm_E4m3_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f", 512, "106e90cda78af26a70325c0342df7e7e26908621f334c1dd3a82f13b4b411f56", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin_len, 163888, "bmm_E4m3_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f", 512, "1e0110c29e2d43b9ac2c102a2233c22486f9331e2a2b3696f492a8694841a126", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -60419,7 +65798,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin_len, 164016, "bmm_E4m3_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f", 512, "80091b9c1df2163cd4e911356c5e901c199e652fde7a42f33b79fb88bcbf66ad", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin_len, 164016, "bmm_E4m3_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f", 512, "83c1e95ee5b8f6fa912cf80b8e661761faac2acacd4cc620c3f7ca54811aae11", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -60545,7 +65924,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin_len, 163888, "bmm_E4m3_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f", 512, "46bb858d70bff24998ab9fac472f3a02371abc3f85a2f637169aaa206e56e740", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f_cubin_len, 163888, "bmm_E4m3_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f", 512, "eca412e174a3688d4346babff491e8b678ff9bd826d50b5204361157d208e030", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -60671,7 +66050,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f_cubin_len, 77600, "bmm_E4m3_E4m3E4m3_Fp32_t128x8x128_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f", 384, "bf97736c6eca1938d1aa6b138934e6f0af4630121c94daef32a02f694297aab1", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f_cubin_len, 77600, "bmm_E4m3_E4m3E4m3_Fp32_t128x8x128_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f", 384, "7825b1716c77b21d4cba3a5b77c7de0939e6ee43cc9a8f02f10cb006b74859c4", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -60797,7 +66176,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len, 148400, "bmm_E4m3_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f", 512, "9735f4620f7a49872a1d8fbc70121e44772d31b501c7de756194cb908475f19c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len, 148400, "bmm_E4m3_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f", 512, "2dc8f02c05481658c03609e28e506911ac0be3f5a250efdd2552f34560d10eef", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -60923,7 +66302,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len, 148016, "bmm_E4m3_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f", 512, "af89772453aa3810c357ec97cc78993fbc0f4795ef8f1c571f8bdd575ffd278f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len, 148016, "bmm_E4m3_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f", 512, "54b1725b099dd2b3abecdd43102bc76b9c9a71793d856e48238726845ffe9bf8", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -61049,7 +66428,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128u2_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128u2_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f_cubin_len, 77600, "bmm_E4m3_E4m3E4m3_Fp32_t128x8x128u2_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f", 384, "f8896c75af18524fc878cff4d1b88ab07e930a8705136b8446d6c9d116cd8b29", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128u2_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128u2_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f_cubin_len, 77600, "bmm_E4m3_E4m3E4m3_Fp32_t128x8x128u2_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f", 384, "22f6b9b8a134cba2abf6d29d6d9c3b44397d025b1712c1ebf676ba0b41ff2b00", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -61175,7 +66554,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len, 148400, "bmm_E4m3_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f", 512, "794eac7e536e460d8bbade89b6a887cb9b9dcbaf58df7376272b1620ac78e154", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len, 148400, "bmm_E4m3_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f", 512, "d723dcfe488a155f30d90337ffb5a6357a00ce67258fe9f3d74bfa2ffbfa08f7", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -61301,7 +66680,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len, 148016, "bmm_E4m3_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f", 512, "74130c217679c26d03c899ff8001748e72369ffd8ae52a35d9119204d913bbc9", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f_cubin_len, 148016, "bmm_E4m3_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f", 512, "2ce27b9563bba3629393b638e32f882ff068f986b76941a73fe91f2bef85fbf4", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -61427,7 +66806,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f_cubin_len, 214288, "bmm_E4m3_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f", 256, "10b00a2d507f6376c0f3a7f5b07e226989986d1b8508b7940e651fcbf0e14093", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f_cubin_len, 214288, "bmm_E4m3_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f", 256, "e262b75e1101999575fcc625ab34cf53ac20dec8a93b101596019493153044b5", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -61553,7 +66932,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f_cubin_len, 214288, "bmm_E4m3_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f", 256, "5f2b48d498411d03a85bb972301f0ccf8f5161539db326e5ffde98bc07e6d029", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f_cubin_len, 214288, "bmm_E4m3_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f", 256, "ebe7ca0cd0b6aa2b9d90604d6c813c9dc4c11b39eed61d3f60b026d88c48d67a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -61679,7 +67058,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f_cubin_len, 214288, "bmm_E4m3_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f", 256, "98015ea893543e8f48f8f30312c8da20e9e9cc026757d270937740ba1045eb36", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f_cubin_len, 214288, "bmm_E4m3_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f", 256, "7b12ffd321898f943669c81158dfe46bbd1a48d7037a51278b5394e3ba9e1ab7", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -61805,7 +67184,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f_cubin_len, 214288, "bmm_E4m3_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f", 256, "f847eb6bc3ba1b77ffe51b830f6a3ff46fd5b7d504c7f9255e77acd3814d4296", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f_cubin, Bmm_E4m3_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f_cubin_len, 214288, "bmm_E4m3_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f", 256, "1a1edaf7ab60223c651a084c0821092bfd4f890bd823f17377c5075cbdb38624", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -61931,7 +67310,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 197264, "bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "b9176ed1e8ecbb4ec9fd34c66dcec28263393cd67b4cf01f44e49fbe435dbf52", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 197264, "bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "4df7f90aca2297a29bcd5075625817f1a4b0de9b2ffc833dc14ad5507737c53d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -62057,7 +67436,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 197136, "bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "7efafb8e5d6a0f891efd6ec4d0e6b99a335cc463b21ddf3abc7938a49d659880", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 197136, "bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "b512de2a9ff7f98c235f9784f03c688b182bc6ae9d4c3101559b6d482900efb1", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -62183,7 +67562,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 197264, "bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "a9c6b4468d9118a9d2cd3704723b92fb995d40f32b2d01aefb8455d82a76f625", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 197264, "bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "f0dff292edb0e4e60cb96af164810542c30480d560ffc52d38b45e594e2ffdb9", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -62309,7 +67688,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 197136, "bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "d1daec13fd45ee77f5e8ca27fc45dbd6afd4b4ee9e928bda3171758595665fe7", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 197136, "bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "ab5213183fe1f76400f9519370e262408ca7b6965ac9e011a64cf7e7f9a7445e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -62435,7 +67814,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 219792, "bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "fc1ca5433f76492353e9f716a2c691ff15e48aeb156d6312783bd4e754284125", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 219792, "bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "9bce2391095bd8a05fc780031363de047b3820ca88752dc35ca480b2198c4dda", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -62561,7 +67940,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 219664, "bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "abbdff57e604e3a3438e53d567a0ab5bb8f848f0e7d213a8c6e648ac0af303b6", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 219664, "bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "1d8c7f5b417f04ca1d9559dbe26130e5e9798d4bf1ac4d8ce3b9634669038243", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -62687,7 +68066,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 219792, "bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "ff2349beeecd60b8a4f70866ddb80e58372d603a1972f97bc9d9add585072fc2", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 219792, "bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "7486566276fe5902bea09c90d43d0b74a8e28779f7bd6e45d1da6d3c5ef2ecef", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -62813,7 +68192,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 219664, "bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "01a39e60531aa46ccce80294035a8eb9eb623a880a30df2e5e8d0a4cba653eb6", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 219664, "bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "109fabd2e3b21a2413554c30f2abd84ab03c9dd34632b911766f7999b4bd1224", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -62939,7 +68318,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 214672, "bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "43b89cdd540a4d30470b8a68383b70bb45ed8eab0f86cbc98955f4b96bb485f4", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 214672, "bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "1d5a81c218280f717bb4ee164765f3347f887dd50ab399be35c8daf3fe023301", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -63065,7 +68444,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 214544, "bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "087ef901e031017bbe42c1887b044ed0c58ba3fde9f141cf6bfefa44bd442983", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 214544, "bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "5365efe10b9dd40073f1f8a49cd7647bc4e03c5c5e2423f75685a6319d6cc282", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -63191,7 +68570,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 214672, "bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "e66f6fd87e374d25e54b3f663ecbff37326993cddf1209f7e56dce6102322609", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 214672, "bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "078db36ccbe053e5cfb4d9c8da9ac913bc6c2bc24c4f888172c0d1eb49b06bd3", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -63317,7 +68696,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 214544, "bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "421cdeb05cdd8841d98b131d81013246d3c41d26dd4610ea79e6929bdb2ac46a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 214544, "bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "a2a35171ae2850e621594cedee3578951ebb698f3a891585751d9665db9e57ca", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -63443,7 +68822,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 186000, "bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "38f7a35eb3071576c8646ec5d8015ac07637414551823239de8febc2eda23030", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 186000, "bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "3304fffa39fb15284fc713cab4ff019a4fed31b23e41e193797b04ec90c61130", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -63569,7 +68948,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 185872, "bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "b59a6f402ad5d671e3171f3a93000a5b81adf11710e91c487800352a15c84340", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 185872, "bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "a91ae4e3ce9509e846980fb537c9e181657b6187325e7c3215f99df2c0e738a2", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -63695,7 +69074,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 186000, "bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "c22415f5c02ce88e5e50580b1c316a63d8fc1af0ee0d0318e48d7cec8713d66e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 186000, "bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "3a7d6489ed0f30b6d6d2a592d75754bb6d16ebccc23bbb09c535f654f86cde99", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -63821,7 +69200,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 185872, "bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "b24161e6870621828c26b892fa4b3eb25e2961d30f2a7fb58803122804063153", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 185872, "bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "95ee857d27439f2338bbaf4fe89432f97010821a66ac1b92e49f143e768c0e70", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -63947,7 +69326,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 221840, "bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "cc33375f30da72c6adcb1fa1d01af12d374c4b73e35147484ade96e45e7a7b7f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 221840, "bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "11008f2b4d117f38f86e4e2068077fa696ad3d5c408136b9a74559dff481f258", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -64073,7 +69452,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 221456, "bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "ecf9077063770fdfe2e8400bca946e4652e993645769b078b85dc9bcf44bb7c3", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 221456, "bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "845ef101a90b77c215a5d8e7ce95c1e59f683a7d604f6a6e00efdf77cfa9226f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -64199,7 +69578,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 221840, "bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "bd27f8e5b0267e76dd774f5d1341c9a0ac9c5a7ead094bff30b6c603dc224c3a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 221840, "bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "1d0d9332d817eb347c5969657f21dfc13d589c6309992972025b256f36b745dd", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -64325,7 +69704,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 221456, "bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "b33396900a11a3f382088e91d069ff4c73570afed9457946e6b80e999aa069ba", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 221456, "bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 384, "f3430b46cc21cb322a6a89f08319ab9e12b494920f35ea9195e7070d72b69a16", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -64451,7 +69830,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Fp16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f_cubin, Bmm_Fp16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f_cubin_len, 163344, "bmm_Fp16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f", 512, "c6f451d32798efa5176013225f5f16fb47871368a7f5fcf6136e01e87fa77f05", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Fp16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f_cubin, Bmm_Fp16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f_cubin_len, 163344, "bmm_Fp16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f", 512, "d05eea21cbd9c06758aae6667157ae4557e6e1a53bbeb2e9bf8c764ab6ee4e0e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -64577,7 +69956,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Fp16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f_cubin, Bmm_Fp16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f_cubin_len, 163344, "bmm_Fp16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f", 512, "f30a522a3b2611527238e547699ebc4bafecf067fb2aa9b686f555f098d5e8ed", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Fp16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f_cubin, Bmm_Fp16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f_cubin_len, 163344, "bmm_Fp16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f", 512, "3aa2a71784dee187b52edccea80343a53fa3d65e59ffeaa61ed92fe02a1cf04b", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -64703,7 +70082,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Fp16_E4m3E4m3_Fp32_t128x8x128_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f_cubin, Bmm_Fp16_E4m3E4m3_Fp32_t128x8x128_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f_cubin_len, 77600, "bmm_Fp16_E4m3E4m3_Fp32_t128x8x128_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f", 384, "222a8605d41afbf8544ef43605dd3db148c674733cea55792716898133ab0836", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Fp16_E4m3E4m3_Fp32_t128x8x128_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f_cubin, Bmm_Fp16_E4m3E4m3_Fp32_t128x8x128_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f_cubin_len, 77600, "bmm_Fp16_E4m3E4m3_Fp32_t128x8x128_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f", 384, "cb9de38ae88d3b96341e86e8f0039e65c2de2a51d115b3fcb5aeb72ed6f8f3bc", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -64829,7 +70208,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Fp16_E4m3E4m3_Fp32_t128x8x128u2_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f_cubin, Bmm_Fp16_E4m3E4m3_Fp32_t128x8x128u2_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f_cubin_len, 77600, "bmm_Fp16_E4m3E4m3_Fp32_t128x8x128u2_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f", 384, "60e6793839ae1faf6e1586a798ce2c23a03681099f5f5de6d3f34718a151c94b", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Fp16_E4m3E4m3_Fp32_t128x8x128u2_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f_cubin, Bmm_Fp16_E4m3E4m3_Fp32_t128x8x128u2_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f_cubin_len, 77600, "bmm_Fp16_E4m3E4m3_Fp32_t128x8x128u2_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f", 384, "3f0367b79e3d2d0e975095d1377827d328389788f0e7c0dbdc4f08d065770a24", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -64955,7 +70334,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Fp16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f_cubin, Bmm_Fp16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f_cubin_len, 215312, "bmm_Fp16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f", 256, "d4c81d413ee0008dfab9876f9fa187fb491ebba0658f05d0808bb71d7c7ee841", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Fp16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f_cubin, Bmm_Fp16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f_cubin_len, 215312, "bmm_Fp16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f", 256, "b2392fc7261e797ab2eb853bf6c076151de140c1dfc57d60a4166d82b0c8d0c9", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -65081,7 +70460,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Fp16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f_cubin, Bmm_Fp16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f_cubin_len, 215312, "bmm_Fp16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f", 256, "770e5c35da7a98e7fe15407e614f52755dc8173f0b83aabbdfe6556f278d0be8", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Fp16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f_cubin, Bmm_Fp16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f_cubin_len, 215312, "bmm_Fp16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f", 256, "776720f4689bf896764ed3f14db6be3b435aebe1b39300f4e979327648ff82d0", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -65207,7 +70586,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Fp16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f_cubin, Bmm_Fp16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f_cubin_len, 215312, "bmm_Fp16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f", 256, "5b6536b38365461fdcddd19a9efc0f10a92e410b4274c6cdfcb84e5673e6b63c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Fp16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f_cubin, Bmm_Fp16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f_cubin_len, 215312, "bmm_Fp16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f", 256, "5bac0ae01570c2347a2a28c211519c438167c194360267b3725a392e40ac883f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -65333,7 +70712,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_Fp16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f_cubin, Bmm_Fp16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f_cubin_len, 215312, "bmm_Fp16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f", 256, "a2b98a99fc23fac19dfc19c7b71cf3061d2715d69c6db8b189830eca751ff87a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Fp16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f_cubin, Bmm_Fp16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f_cubin_len, 215312, "bmm_Fp16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f", 256, "1adb3f26de1e4ed1f193dc85efd3541e338be1b4231f73265f469b23f865acd8", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -65459,7 +70838,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len, 190360, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f", 768, "41c49b5c29da4462a33edd8af0ea689fbce4fdba08f0c02fb989ea37abb7a6c0", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len, 190360, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f", 768, "7e4c070d2fb94627aa0485f27b5e2582817d1b76f1387cf158da48b87a8b7b94", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -65585,7 +70964,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len, 191384, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f", 768, "8d1bbde613872a43a2b7280c74d3b31a0fd4e23dd8f9df5533c983bb7d066897", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len, 191384, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f", 768, "1a800ce7ed21d55b866d34d211f103c463bd156aad58bca94d1ea58da62dddb1", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -65711,7 +71090,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7x7x7x7x1x7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_lsfbW4_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7x7x7x7x1x7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_lsfbW4_dynB_sm100f_cubin_len, 191384, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7x7x7x7x1x7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_lsfbW4_dynB_sm100f", 384, "e523ecebd69225c3d675a4c823b046cbb3483d9cc926b3dd8ded662ba5893dcb", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7x7x7x7x1x7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_lsfbW4_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7x7x7x7x1x7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_lsfbW4_dynB_sm100f_cubin_len, 191384, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7x7x7x7x1x7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_lsfbW4_dynB_sm100f", 384, "c64912b90db6619374f57ac285a07ad4eabccb3b1b588797fa0e3ec16ab86aa8", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -65837,7 +71216,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7x7x7x7x1x7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_siTuGlu_lbW4_lsfbW4_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7x7x7x7x1x7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_siTuGlu_lbW4_lsfbW4_dynB_sm100f_cubin_len, 190360, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7x7x7x7x1x7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_siTuGlu_lbW4_lsfbW4_dynB_sm100f", 384, "54dcb526aadfa88f5546f868cda57f955efdee3e00f98d22b0e363de09d29287", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7x7x7x7x1x7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_siTuGlu_lbW4_lsfbW4_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7x7x7x7x1x7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_siTuGlu_lbW4_lsfbW4_dynB_sm100f_cubin_len, 190360, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7x7x7x7x1x7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_siTuGlu_lbW4_lsfbW4_dynB_sm100f", 384, "9a667bdd6477db658aa71e80ab62c5d8159088f213e9ecc2c84c8885d84363fd", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -65963,7 +71342,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128u2_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128u2_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len, 190360, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128u2_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f", 768, "98b5cd405aa5825a01083fa5ac9717eea6795764772ea4bdf41b4f692e7438a7", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128u2_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128u2_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len, 190360, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128u2_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f", 768, "bf957abcaefc3111c2ad919906564d6df61ced2d960b349c829a54f30d098144", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -66089,7 +71468,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128u2_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128u2_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len, 191384, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128u2_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f", 768, "a84b7e5b283c22af508aed4e3b1c57c7b95dbf3a10f21a44c9b05141f66de78c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128u2_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128u2_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len, 191384, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128u2_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f", 768, "8604b42d4a134a2a8f59c73140720cced5fad3d439f7c9fd182ea0b68fa55ada", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -66215,7 +71594,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len, 214680, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f", 768, "2def5a9167d42081c328ff38c2078b1a56f22b9ba95ef6ee15076560e4fc0121", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len, 214680, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f", 768, "0b88244b36b2b7fb2604c07396a75eae0320d3d6da81eaf6effc58df1d2e66d4", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -66341,7 +71720,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len, 215704, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f", 768, "1024acf28e0bb05e3aed6e64a9775f0e2290fd028c328be6e84efe217a1c59e6", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len, 215704, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f", 768, "f8dd8af0cdd7dad68a8eccd2765285ef05612fdcf838ede861c946ba57ab5917", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -66467,7 +71846,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4x4x4x4x1x4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_lsfbW4_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4x4x4x4x1x4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_lsfbW4_dynB_sm100f_cubin_len, 215704, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4x4x4x4x1x4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_lsfbW4_dynB_sm100f", 384, "1ebdfe613e3ac9f3e2bab6fd2aecebecdf19414dedd2374248175cc1c2b1b0cc", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4x4x4x4x1x4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_lsfbW4_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4x4x4x4x1x4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_lsfbW4_dynB_sm100f_cubin_len, 215704, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4x4x4x4x1x4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_lsfbW4_dynB_sm100f", 384, "87af0c317e282067db1fbe7ed4239b3cc1420b9b2da464729bfc616b3763fa44", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -66593,7 +71972,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4x4x4x4x1x4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_siTuGlu_lbW4_lsfbW4_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4x4x4x4x1x4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_siTuGlu_lbW4_lsfbW4_dynB_sm100f_cubin_len, 214680, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4x4x4x4x1x4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_siTuGlu_lbW4_lsfbW4_dynB_sm100f", 384, "fed222b86391c7c34a24f149a76f924a0226e34b3e450040a49c97e82fb9334d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4x4x4x4x1x4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_siTuGlu_lbW4_lsfbW4_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4x4x4x4x1x4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_siTuGlu_lbW4_lsfbW4_dynB_sm100f_cubin_len, 214680, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4x4x4x4x1x4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_siTuGlu_lbW4_lsfbW4_dynB_sm100f", 384, "644619c0e65b43d48c93b156eef1ae2aca14dcf03979d35e52cc8df537fc9cb3", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -66719,7 +72098,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256u2_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256u2_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len, 214680, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256u2_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f", 768, "37af651f4b874418585f208a21d8f89bb69d211cb4ffe37ddb09b16e74a0946c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256u2_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256u2_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len, 214680, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256u2_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f", 768, "a4e3f6ef4a1872a5a1c6fe4005cb8634220f61bdba2c8c8c1f5b43705c061a79", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -66845,7 +72224,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256u2_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256u2_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len, 215704, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256u2_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f", 768, "3aeb51137ba1f1f595d15df7e6cbda62f715d5eceab772f9b35e52c2f8a6aeac", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256u2_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256u2_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len, 215704, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256u2_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f", 768, "a20e24ca50a56788161f95bd81974d48f79bdd96427767c1fe0d25b955580dab", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -66971,7 +72350,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 197520, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 640, "5bfdd8cd202d3f33f62029a0c0008ad3c2b682be38575a8f6384a68dfe3c0d8b", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 197520, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 640, "d806568649d9244baecd1a5f8aa0b5de8cd024cd4c44ec445d547d85df9f4f4a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -67097,7 +72476,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 197520, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "0245ae9b677c31722373472f7dce52595865aff37079c15f4354aa0fed8ffc38", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 197520, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "67167fffe14fbc121f50a7d99e8a087a0463248b44811846013638716fa4fcc1", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -67223,7 +72602,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 197136, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 640, "b807bf53f5272a0c48fffb35a55c4364a1d9a0aeb94b880ab6ca688149971111", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 197136, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 640, "622c45380fa688204b0538d12042adc32f54ce9f7cd67ca4473e93f053abcbd8", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -67349,7 +72728,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 197136, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "e7c3707ecc45a56252b7a6412916d0e311441fd351864e9803c1f62fae248d0c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 197136, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "d509ca9f1e4857ac210b4d254b13063667be1b80daf45cd4659adb126f01f803", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -67475,7 +72854,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 223128, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "3f2a93dba4bfc40006c609808e12329f0c33a01963e97d72c6657b1e2e70e302", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 223128, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "3a6670afa73908091cd018dcbeee3de35030427c647aabf2d635cf794758f893", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -67601,7 +72980,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 223128, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "563a2d87e62b4629843aea9e92838883779557940c42039e5623a18dd1ad91e2", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 223128, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "ea91ca62eccb3a9c741106a3f8c86dfef7e0a6ec04c386ce7723e7b2673f1d1b", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -67727,7 +73106,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 223000, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "e53f7bdc5beb779215da83ed06f8b2ff7fdd371093c036e1c228dc7520132bb5", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 223000, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "1b81c8b72efe2c1ee4168cfa93ff886ceb38aa58cac47ebcfe1b2d965d4a9bda", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -67853,7 +73232,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 223000, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "2e106095c57d7750c2d4e3b8b0c1ae240dfdbbfcc2b5af6f8a0f8c96477d6e94", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 223000, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "a53b2bafda1a6cace0dfde7a6e51ca9fda829fb385fc3a5f4d12fb2e00f1debc", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -67979,7 +73358,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 197520, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 640, "7fd2d8906fc0c13a304612162c55e76360c2dbbf412d424e42b7de9f511f44e3", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 197520, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 640, "def263bfba5cc3d8602ae7d3eca05de6c46cee1c2c191c1b8c7e29af80365615", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -68105,7 +73484,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 197520, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "891acad6ccbea3f0047c96feee6b142333cbfcbf511a2a452f8f9a1df541c96b", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 197520, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "b6bd78a21147b3c6a3343215e2e902442107226a1ef5d30583c971454e4f1c12", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -68231,7 +73610,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 197136, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 640, "798ac5b1f13b50f3326a7de1988b2b918344d1f3b7bbf97082a1783d4094a4af", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 197136, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 640, "3d4c54ca0f7c28edbda0e90a9df277e3205d996727577222a6ea9ac2a8aefd17", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -68357,7 +73736,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 197136, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "489e8f4b6743eaea6911e8258dc64c4d962b9558ed91b793dce6faa3554d2a1f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 197136, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "8846e9cd489036c36e0c637d9a7435d8ecab4ffa77a5147e218deb6a904680ba", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -68483,7 +73862,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 223128, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "3ab8eca687f3f073b82ced294d3d9fe3f04f03779ba53d7c8dee35df0df6f9d0", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 223128, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "3e11c3a0864bef4d6c0fa3a1c7a9d25ecafa0d535e96f44e0520481f3a25a2ff", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -68609,7 +73988,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 223128, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "efd313e300227160bea6f87373b7cca358acfcac4a9c904fd835321f405b29a0", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 223128, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "c3d6e8a0d3ca2f225945f18f4359a2c7263221c7edaf6dc249d83dae9a5381b1", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -68735,7 +74114,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 223000, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "b1bd0c055b1673061d8edb54df5d0a664c9e7d39ccfd2382952eaaf6ba665f96", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 223000, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "49535aacf3800526afee1b1446ddc34c8ae3fcdd35f68f17e3aefcb4769237f8", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -68861,7 +74240,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 223000, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "3b6582cdb0bca52a8541b65362a697223d603e2ef2b5ed575d52375c151d3d8e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 223000, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "fc3f2709ed84e9c1557ca5c0e6199dcca1e9123b33e3fd2adfe38e0bfa750fa0", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -68987,7 +74366,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x128_s6x6x6x6x1x6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x128_s6x6x6x6x1x6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len, 222872, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x128_s6x6x6x6x1x6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f", 640, "453174759486d5cf69905d999afdc218ca339cb24ad2a7f90146751588222c3b", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x128_s6x6x6x6x1x6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x128_s6x6x6x6x1x6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len, 222872, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x128_s6x6x6x6x1x6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f", 640, "a9669f49158d8311513479ab3ecb02cdf7aa0e2fabb07d778f96743dcaf1fe41", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -69113,7 +74492,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x128_s6x6x6x6x1x6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x128_s6x6x6x6x1x6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len, 221848, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x128_s6x6x6x6x1x6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f", 640, "eb804c975c9070017f70d777a78da26c1ddb7f24f0c40c349a0867256d06e65f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x128_s6x6x6x6x1x6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x128_s6x6x6x6x1x6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len, 221848, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x128_s6x6x6x6x1x6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f", 640, "660f48dcf5bc0bcf43859eeb46f4abe69856a66a26ff9ab7d8433ab99c588184", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -69239,7 +74618,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x256_s3x3x3x3x1x3_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x256_s3x3x3x3x1x3_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len, 222872, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x256_s3x3x3x3x1x3_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f", 640, "e387feae1e1ca9b2c93bc72d54cd035b4f9796b6380c458a3d615a82141633e8", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x256_s3x3x3x3x1x3_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x256_s3x3x3x3x1x3_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len, 222872, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x256_s3x3x3x3x1x3_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f", 640, "125bcc8302761d86a294735beeb41b9eb18a258189f598d9cae6170a96c0d793", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -69365,7 +74744,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x256_s3x3x3x3x1x3_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x256_s3x3x3x3x1x3_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len, 221848, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x256_s3x3x3x3x1x3_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f", 640, "43daaae59700c6261cfff6fae7cbe91056f4e61452170608db2f80b625b6090d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x256_s3x3x3x3x1x3_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x256_s3x3x3x3x1x3_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f_cubin_len, 221848, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x256_s3x3x3x3x1x3_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f", 640, "2466b633476fc8ef0c2a9663802527e1a58cc0773a41f7520565d3daed42962f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -69491,7 +74870,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 221072, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 640, "509722e9ec11421b2513f7cd2d3faff5bc473434b1cbb0da19c1f7c019950237", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 221072, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 640, "9b58c69722a0d259dc988250c641ea96c6fa17c069446d22aaa517b966180cff", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -69617,7 +74996,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 221072, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "535f3fe8da9d47dd1988bd950f780b4fdaca568757b2b3a41cd2347f21e08292", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 221072, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "bb12633eb0e7d270a0087dacc8c32ca81682bce3279e81e1c3761fd2266c2c76", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -69743,7 +75122,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 220688, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 640, "e1de93fa95104b4050c53238c35ad3be46eac580d026df11dddae77c1165816b", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 220688, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 640, "76d904498ef9f4d50de5e3f440a19689ab9ae2dd8566557bc2436a02bd5de7b5", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -69869,7 +75248,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 220688, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "68ed4c77bda741a7fd9e5200e945a0686d69bc27d39f88c09b640fe65433f058", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 220688, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "a78cad22c19d8cac46da73dc8232d717e7104b9c8af30fab5b4e68acb18cb8b6", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -69995,7 +75374,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 200600, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 640, "446a3b442242e4524b8db27b529c31a78bd0a5284c75b1e110c867008496f8bf", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 200600, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 640, "34fb50e87ec90c3cf328450888919c74cd24c5505b36705bfcf044eb7f821387", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -70121,7 +75500,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 200600, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "79b4e3f1a0c50d0de7243a468b27d5a5d19fea41d1de624427712eef10430d7b", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 200600, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "f4a9693b0266e795d272a05c0512a07e2dc6d951c3e7463447c0c726c63ecc26", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -70247,7 +75626,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 200216, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 640, "84976dbf9331b270bedc382984a5529d903072f96e98316228b9868bf29326bb", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 200216, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 640, "a4dd4a3293f1833533385ddfe8531b6bdebcdfe8b8bfb45a4dca4e20c53ada73", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -70373,7 +75752,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 200216, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "0df5949855db1a0d57d07f1000b22df5dd15fece821da9875b8c8bee950fb7b9", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 200216, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "1ede8db421d6fa6110f3d42aa0a3df7c99ffaa7098de688ae87c2ae96fc8fa42", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -70499,7 +75878,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 221072, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 640, "5f57e2b74c8c324aa1f3de982b3c67cfd0237a59aacca7308c1dabca68359952", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 221072, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 640, "03dda35f47b490d5f3ea31ee2dc6037a806bd2b84061b539f6c6444d47d581d9", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -70625,7 +76004,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 221072, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "2030724ccadf3ea36dde52962cc9e00ee9aaca201659a010e8cd1d9dd6580eb3", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 221072, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "4b5f43415cc7b419a81af9eb65eafd8408160508ecb9be70093d623f850ed36c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -70751,7 +76130,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 220688, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 640, "000296a9befea9ec01b16d1bc984d9d9d858fec1471696e9410434b333fce0f9", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 220688, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 640, "232a8dafad21fa937276a5323146b6bf655e7d0ab2d351078fa68c3758d60352", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -70877,7 +76256,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 220688, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "32c35df4172c523d329b3fef001e6d531937b67b9152c7577ca9d8ab3dec4d78", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 220688, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "2e4e6f75a525a7195332f9a7654a4516b4938a851c5739004bc0b050ca2a92ec", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -71003,7 +76382,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 200600, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 640, "3122cf721fc10ab2547d5df61f8951db840322e0a3e245d7e08cab99a28b7d6f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 200600, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 640, "3a39f2c79e78bcbf3878ba11323779d01945308decd675980f58f068c18ea1ab", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -71129,7 +76508,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 200600, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "d6a0b34d4655de8521e8361f05351431d9d3e5c0537d3f1ee1594ea04200740a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 200600, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "bbd7a6e1048da437004b599bae7c355038b53415a294fc8ad31516de9667ad75", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -71255,7 +76634,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 200216, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 640, "72cba0eebe2d931ddfcdd1419bb57f7dfad07cce74c665170a793a46046bb0e6", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 200216, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 640, "09651dee13d4e379718673a73dd7ea8bb397d04ffbc08b0eacf50fe6c00d2885", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -71381,7 +76760,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 200216, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "b1a3443eacc4496bd1757ee2a7da68a37068b3eb002ecfcea9a953c9f6ff2045", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 200216, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 640, "b54fbee520998d3ffc5ef2160975ce336bb96511e98b6aa621e223ecd367f200", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -71507,7 +76886,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f_cubin_len, 162968, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f", 512, "c10896220d4932d9d0a6b88694203f537050af37b542c2ed2cf57093ac36d54b", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f_cubin_len, 162968, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f", 512, "1402958a2475b660048b3520c20740951330476e7bc0372a55e2c8834dfa2b83", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -71633,7 +77012,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f_cubin_len, 163992, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f", 512, "34b8f2bf422d8a04996c43ca98d2d34ee40492c0f412e99d7a1d5ef811139358", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f_cubin_len, 163992, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f", 512, "f68dd89840bd9d3eed7b1ba66c66df5d450b9d09bf9e02b29adc49eecdf96c8b", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -71759,7 +77138,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128u2_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128u2_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f_cubin_len, 162968, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128u2_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f", 512, "9a1046b197f953ef89caa22a96791b12f7fbf3fd9e9e88505c8fa9f1f17207ca", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128u2_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128u2_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f_cubin_len, 162968, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128u2_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f", 512, "098a6c1f2972217c9204b9ea1ddfa3a1ce3ebdc0b4b816ae8af877323cd50a51", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -71885,7 +77264,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128u2_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128u2_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f_cubin_len, 163992, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128u2_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f", 512, "bb64c69a2995d59f55d5a4e9ff876c6afcb99860e94743454e14a766db5752d3", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128u2_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128u2_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f_cubin_len, 163992, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128u2_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f", 512, "ec2bc6d1c96ea9c4e8aa8b920fa0e1eba511ec0e0aa30972bb41349319b6b027", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -72011,7 +77390,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f_cubin_len, 182936, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f", 512, "4cb6e06e25c2958b33903d82eb25de5951b2ec1f9065e976fbff6e76e995096c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f_cubin_len, 182936, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f", 512, "71f0669667575b6aee126b928a54d10fb6b1da146440f6f334c87c7bcb2a23b1", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -72137,7 +77516,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f_cubin_len, 183960, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f", 512, "bd21ef8a85152bffc3bf9d532431d8bd7c43e7e06f0aa73206b4c4456a8f7e55", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f_cubin_len, 183960, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f", 512, "f6939cbabe0e5ece56e54a06c0fcfb21a3aeba706f01f096c6f7efa1b5b48d0e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -72263,7 +77642,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256u2_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256u2_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f_cubin_len, 182936, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256u2_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f", 512, "8d7c80646237114d2a5ca5c8dfb032f3da8038aa8fbadf703c591ffd505eae50", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256u2_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256u2_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f_cubin_len, 182936, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256u2_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f", 512, "5876facfe39c5c316a9d9945625eb8733d7ee83da999b1ca73b180f9404d2109", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -72389,7 +77768,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256u2_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256u2_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f_cubin_len, 183960, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256u2_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f", 512, "b990550c50bc69dced298f4bea3542c265729a3973713ac7499e26c2a82755ed", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256u2_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256u2_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f_cubin_len, 183960, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256u2_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f", 512, "0b1062c1265483234bd48af0a4c7e464901c16065b76fa8096c4bcb90362f1e1", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -72515,7 +77894,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 222096, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "a71a666d4eb660092b018280d41d6d3abf3c172d4c4aa9c4dc35045a7c0ce305", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 222096, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "bf354321ddf5550fdd4c0a257479414c2e8c3081669fa0bc8ff0ec43b05fe66c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -72641,7 +78020,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 222096, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "07a055ef85e904d22689ec7e248731ec9358e84799d4bc60b26872137d72f9b8", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 222096, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "c203b9daf0881d2c1622bf97e4be11b5bd9a96b5be9f9ebfa06ead4fc05c5eac", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -72767,7 +78146,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 221968, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "424faa689f530844e9916d7124b0e21d81766a8e78b256816972309795766182", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 221968, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "6a566bb8fa350568027ae318c19b8da3f50cf8737b47695777ad9b07d8d9dfd3", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -72893,7 +78272,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 221968, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "d6962e99fde2cf9211a083f11a062664dcca57194d46b6abb5d0b9f8c31bd32b", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 221968, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "f51e37ba205a35106a01c9973a39c2461811611b49126e38687e7349bfe57ae9", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -73019,7 +78398,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 229272, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "d0c04ad40bed365b29dc419528d0ace66b60d05911da1b404a7499c271d2f3f6", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 229272, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "b74ac770a227fbed385bf645597b3c982f3dbace33d1facb3ba8b5bb79a6dcce", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -73145,7 +78524,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 229272, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "b2da343403a27e0bcf97bf7afd44bc7babe463661df8002d4f802d10f17bf83e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 229272, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "de89ff526174d1d2cc77c1de8cf2efcf38ac356ce48bf3036ff5c7c06380dbec", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -73271,7 +78650,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 220952, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "7e5e80b0a30d4aa40f847a4ad2e032cb4e9545516407b5bbceab25f21acb4e29", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 220952, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "588de6fe971ae76f37c5d4b4789f7d605e338acb994bbe17cfe95992f43a3186", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -73397,7 +78776,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 220952, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "7c062c8271b765c785821e7a8a2f6c0f97dffe8e46b8b288e348ca69881ff5df", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 220952, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "0692282a6b0e01da0576e5b153ad6a88714f1705df622d769155dd939d1ebf93", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -73523,7 +78902,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 222096, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "7abdb9582e2d1fc00828ae028dc9f6d37619c49bd9f9fd2021350b076ecb1c7a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 222096, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "85cae16cd09f7dc19add035213aaa20fd9ba15596fc743f73aac9c02fc0e62b3", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -73649,7 +79028,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 222096, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "53474c90b1714cf9ed273c68e43c1b2be5ae9b4612f523f10be7e88b577bd30d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 222096, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "e15c4c1ba2d5d878724c14a82fc1aaf2ad5c60806ba0d7fdba48da2bac09d266", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -73775,7 +79154,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 221968, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "e07a060965d5f7483ef97c78aaa350603acee827b350da4ef5a4e7be011c5935", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 221968, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "015124ad9ccfc9bd2df9ec891420fd245410de2a35172b4d0eb94bf222631384", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -73901,7 +79280,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 221968, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "2b6748125850b92b8fc220bcda73d14c28c53a95c7f5fa2e92c4ac751372a034", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 221968, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "ea4a1b8d7a1c9c588df2901ae62968996393de6f64a219345e54ccc95021343a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -74027,7 +79406,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 229272, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "60a0ca4bb8bd8b77e68a324570b357e92f26d67577990b5e0e662ab1a56d7fcb", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 229272, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "02991cc0452c4a68ce4f4e4fa3ea718eb0ed69c8018ee63e86b252643e18c874", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -74153,7 +79532,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 229272, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "42dcc62d512a0e26161cf3aecc8ffab70b437e2990b8e13c3912c8d251f424f4", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 229272, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "6da6864bcf3c725776e7ab199c3ac6ab312e070d3533c1ab466285fdc8d96e33", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -74279,7 +79658,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 220952, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "8c425a2ffd03008fadb054b8dc75a1b7a2a8596b3defe81c4edc3a1e4f5867e6", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 220952, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "f6e5768b0cb20cbb12ad99ca059a10099225e3854d5e80d3879937c799b46a0d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -74405,7 +79784,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 220952, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "1a9392eece3afc1b8dd4da26998c85d9f598a2d098ab96de97eab2ab9a0af035", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 220952, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "65653049cada1836d5c4be78b7baa1f6b485aba452288e92e371922d39ba2cc6", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -74531,7 +79910,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 221840, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "d332e5598f37e6c678d1b76290f06fb775d7669aae02855db58b14ea1fbaa1b9", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 221840, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "625e8f6154572ff191b7acba5f0853f29d3dc9e1dd12680bccf9dbd8b1117681", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -74657,7 +80036,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 221840, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "774916dc6f096cb494f14ce9501f962ac35fa1ed3a785256d64afee6bab661da", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 221840, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "4c691848f6191600c069dfa8a0a8fe533573fcc50875eacc02d557a58bf97ea8", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -74783,7 +80162,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 221712, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "40b596f363fdecd4645e7d748540a4b2899e4922bcf1e084149857c6495d5056", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 221712, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "8e055dc5925f1cad34c169b180fd237ec02326b7c3385405c242342ca27b1a30", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -74909,7 +80288,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 221712, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "9f3b43f9e4d30b9baf4631ea25b87cdda9cfd262657bd34f325c1fc87ee9462c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 221712, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "fe91020e633d91e5028669534c90743448b56bf10fc0f66f5bfa665afdbbaf36", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -75035,7 +80414,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 229016, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "d862d8d888775abd117a86faced9b029d3fa8caff245e9b8c2c2e291566eded8", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 229016, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "ce39e70918710a7055c9422e6abca1dd1ea122676e9e50bb9ef5a451deb42b05", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -75161,7 +80540,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 229016, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "3626a50a93aa53dd237c3d7e8ef6f0827ae8bd5deadbb7902bc023f73a405237", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 229016, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "1381d0faa0db058ad952adba77c4602afdc7026b319beec32c6cc499f2837171", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -75287,7 +80666,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 220696, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "9dcd8f6398c371c1430a165162aeac426551277ac0e23402a87dd94828e3674e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 220696, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "6732ccddcd4f8161fabb6f396d4d5edb44c5f39bfc0e9ed131fb9d3b4d40525c", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -75413,7 +80792,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 220696, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "1720f6123c92e49216693c6f04a91c106b1bd9ab55cb2fe7843ab8a722be0127", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 220696, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "69ea0c012fd19df158335b20e1aedfb57d5162c302e092a8d87994afb05ccf14", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -75539,7 +80918,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 221840, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "66d3753fea1dfeafa5fe7f107923fb20380a9cd153979c0bbfa19e58edec1767", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 221840, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "97e2685e267bed8361f5f71379206bbccf9b04f3ebfdd4b6212ef64b06b2dcb4", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -75665,7 +81044,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 221840, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "75473500f82ba217367c76c935411d7ff28c35d3d614ece97a84f236040f7f30", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 221840, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "1c88977dc73c87bc5c74f926d07df63a4f2b30bb4088d2425e96d2c45a725a04", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -75791,7 +81170,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 221712, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "f3303c9bae3db25e963f01a79aa9d2e25719dbeb0684bdeeedbf720f19a6d749", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 221712, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "4ac4e7fd65fbc41be7c1eb4af7f0dad2c6ef9530931dd702dd6240162423d550", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -75917,7 +81296,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 221712, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "dfa5825e824237d81c091c669c32defbda27a8a7c7e96d9bb1aede6b3368adde", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 221712, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "44f606735232d8fca44cf828fbb2e3869c8cd3f04fe148c783e8bab25c728b5e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -76043,7 +81422,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 229016, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "e9c21220289b892c23c385261784818548c578a5841e46b87db89718b1082ea0", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 229016, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "7b575d3bac5cec22f424954809f578e097757b9ca43c6c5cfdb07c27f1cdcbd0", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -76169,7 +81548,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 229016, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "37634a02f09307d6de4db49a9cbee816f8ad510e1db529c7bb9e1aa8ae329a53", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 229016, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "45bca4a7e76a544d1974cafa277406013bb2792b8c0b9c7e8cbea2bc4f6a2271", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -76295,7 +81674,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 220696, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "dfe8d7374bcd31e819845aa180bc2a490210929e2bbb11fc5c48a1f6359ccc12", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f_cubin_len, 220696, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f", 512, "03ae9d9071259cb3ab278db950604f224383ae072000050682885939e2b7c055", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(0) , /* mBlockK */ -1 @@ -76421,7 +81800,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 220696, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "cb7e360a0bf8605e38dfb63e228576c4e425c022ff0e49faf7ada4b0d4bc91c3", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin, Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f_cubin_len, 220696, "bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f", 512, "7d1e80802c3dc184edb08b72e758244780009834f921c95306c14a10bf4451ee", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -76547,7 +81926,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE4m3MxE2m1_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7_et128x64_m256x128x32_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_laW8_lsfaW4_dynB_sm100f_cubin, Bmm_MxE4m3_MxE4m3MxE2m1_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7_et128x64_m256x128x32_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_laW8_lsfaW4_dynB_sm100f_cubin_len, 193688, "bmm_MxE4m3_MxE4m3MxE2m1_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7_et128x64_m256x128x32_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_laW8_lsfaW4_dynB_sm100f", 896, "6c2e767ae62c7da7581d1f9797c45168598de080002b8390645335a9b47502c0", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE4m3MxE2m1_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7_et128x64_m256x128x32_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_laW8_lsfaW4_dynB_sm100f_cubin, Bmm_MxE4m3_MxE4m3MxE2m1_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7_et128x64_m256x128x32_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_laW8_lsfaW4_dynB_sm100f_cubin_len, 193688, "bmm_MxE4m3_MxE4m3MxE2m1_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7_et128x64_m256x128x32_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_laW8_lsfaW4_dynB_sm100f", 896, "befc3b4912e9274b170f8a84ca6fdb63bfaf1eae0ad7e2db05bbbf7900da9d00", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(2) , /* mBlockK */ -1 @@ -76673,7 +82052,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm100f}, -{Bmm_MxE4m3_MxE4m3MxE2m1_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4_et128x64_m256x128x32_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_laW8_lsfaW4_dynB_sm100f_cubin, Bmm_MxE4m3_MxE4m3MxE2m1_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4_et128x64_m256x128x32_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_laW8_lsfaW4_dynB_sm100f_cubin_len, 217752, "bmm_MxE4m3_MxE4m3MxE2m1_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4_et128x64_m256x128x32_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_laW8_lsfaW4_dynB_sm100f", 896, "149f48cfa67eee8c1615df071852d054eb7f6296dbf3e1bae5c645496ce441e6", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_MxE4m3_MxE4m3MxE2m1_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4_et128x64_m256x128x32_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_laW8_lsfaW4_dynB_sm100f_cubin, Bmm_MxE4m3_MxE4m3MxE2m1_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4_et128x64_m256x128x32_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_laW8_lsfaW4_dynB_sm100f_cubin_len, 217752, "bmm_MxE4m3_MxE4m3MxE2m1_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4_et128x64_m256x128x32_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_laW8_lsfaW4_dynB_sm100f", 896, "699858b362241a5c57ed34de339ba1cd9f8eefbbb6e49073f5c70c5adf64df64", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(2) , /* mBlockK */ -1 @@ -76801,7 +82180,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { }, gemm::SmVersion::Sm100f}, #endif // EXCLUDE_SM_100F #ifndef EXCLUDE_SM_103 -{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len, 116368, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a", 512, "07dcafd30f2e349ea818e3d3b816ec40bbdd0a2f261665c3209726508ec8c975", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len, 116368, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a", 512, "4c4e7a6f4fd7002423593183a3d567929b1bd73345327dd6e258286986fae2e1", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -76927,7 +82306,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm103a}, -{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len, 116240, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a", 384, "38cd91f274f1cfed2c53dfdd68ca2faae8d11f47d158f24271d4297d8c405b90", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len, 116240, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a", 384, "7906881dbcf278faa6f29222290259eff2ee82a6d5e55ec5c23bf6ca53bb84c8", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -77053,7 +82432,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm103a}, -{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len, 116368, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a", 512, "00aa03ce3bcac5ac050aabc7af37a829f073a005f5bdbc33b7daa28f15e09b27", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len, 116368, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a", 512, "231c6f0dd1498128622411301f46235b0d6ddf2bc6ca10f85b9e387532026e00", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -77179,7 +82558,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm103a}, -{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len, 116240, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a", 384, "fb28f2289161b81d1dc6dc067406ae796bb1d944fef6fd6f55b1f9bd64fdd804", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len, 116240, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a", 384, "595504f873cd3b04f04a771a79aa0647e86339b098f1d13d494e345a48e349dd", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -77305,7 +82684,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm103a}, -{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len, 140944, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a", 512, "bc5cda33cbcd55ceae874ce9d799f062a0365d6798895ba0b2033a35a9c3a7ab", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len, 140944, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a", 512, "3704a353f8247460759457aad41e771ad3298fa64c777126229fd44f9f925dbd", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -77431,7 +82810,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm103a}, -{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len, 140816, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a", 384, "9a1d894e4393939e9eac9315efade0459a57365cf47f0d558728225fd504947a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len, 140816, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a", 384, "92a851d3bff42baaa004cf948ccd5191c1338a5c97cce512316f2f6f2921a0dc", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -77557,7 +82936,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm103a}, -{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len, 140944, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a", 512, "a5a1c68e10e8872edf445fb486844f7e8e18589c43c0d84b88209e8ec10f27e6", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len, 140944, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a", 512, "b6e6803e83fc4ec2533e4cc7b92f37f35ffd6d72ce5999b511bb4ff0e2cef822", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -77683,7 +83062,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm103a}, -{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len, 140816, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a", 384, "50f45c442887c874f0efc6cebad0e64ea6fea93ad85ce8b16a841c82f8df14c8", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len, 140816, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a", 384, "e4f5510a0d5544b713721476e1d905c43a0db13a874517abd00999f176a031ad", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -77809,7 +83188,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm103a}, -{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len, 157328, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a", 512, "e65fb4c0f43c6640ec059f63fb12c0d600916e1fd3086dd200c1ca8ab743a4ff", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len, 157328, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a", 512, "2b106f9db8d475af9914de79e48942a1a3b08c2b0fce15a5781addbedf96d36a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -77935,7 +83314,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm103a}, -{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len, 157200, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a", 384, "2af12ba09fda0fe483397879d84869a26678a144ba99885830fae722a632a30f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len, 157200, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a", 384, "995a3cdc105cc1da6b9a09249ca148a4a3859961e33342f2a1792d781578cb58", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -78061,7 +83440,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm103a}, -{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len, 157328, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a", 512, "468139f4660b3e9c9d69f5daaf7b0db00c1a77a49d456f81f298eb29735a4bbd", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len, 157328, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a", 512, "1834ffd2e9cea27620d005582e94dc5474f40cc1d944aa1333d639e9063e94f7", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -78187,7 +83566,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm103a}, -{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len, 157200, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a", 384, "8422c894b48a73264779e2f3deb1cb411e0dbea0f33842a50374214d7be2ddce", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len, 157200, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a", 384, "61d35ef452548a2600a3deb242948d57d1c000bb2b399d56134c0007f859d557", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -78313,7 +83692,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm103a}, -{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len, 104080, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a", 512, "a718a2fe4af1844acab365ced53da58d9f153d04e72310b3293d147b104b55e6", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len, 104080, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a", 512, "3538cddc9700d39ab457c390c781733c0e5f2a753953be20ddfc10d81f549499", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -78439,7 +83818,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm103a}, -{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len, 103952, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a", 384, "d4c46bec43fece4d731780c563dfd45db28edabba6c35af26d9d05f700e78d0a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len, 103952, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a", 384, "1bae59f2e9bb85d01d0e6117175fcc09daa5d3f77efa65a4ed2945135f6ca067", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -78565,7 +83944,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm103a}, -{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len, 104080, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a", 512, "8517303b490220192a8a65cebeab77f45295fad7c5a6f4675b34f8d3ff5508d8", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len, 104080, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a", 512, "8cf4742b185361aa29abbe121cb4155d23dc14235dd8a2c0f22cce8b7dbf06c7", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -78691,7 +84070,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm103a}, -{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len, 103952, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a", 384, "af2f6d683560dac68f537efc9ff98af8fc060027c7c8b64974dee742495fff67", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len, 103952, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a", 384, "196dbbcea756939803eb86ebfa4cae98f9abd8ff97aba062e986d3acce232fb8", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -78817,7 +84196,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm103a}, -{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len, 123536, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a", 512, "f94527d95515605d664db19e06705ad88c6799559780ed938d6548b5b762c8ac", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len, 123536, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a", 512, "b0283b4449af436ab7c89c1ce8cb74bd537490db9b7a13a645459ab8f1c302eb", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -78943,7 +84322,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm103a}, -{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len, 123152, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a", 384, "ce12eddee469fcb4732b85aab09e474886ea24c1ee6afeb5a7fbc405aa0d5efa", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len, 123152, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a", 384, "35460110e59fc813d480728d7a021bcde2e279f4f753203b7323ab60fee4deff", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -79069,7 +84448,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm103a}, -{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len, 123536, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a", 512, "ead53d4e43bd0757cc0365e1cc406b2db4cc0ff73f9cc2ca5e64d2fe0498509e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len, 123536, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a", 512, "3d098a4122a8f31408506877a3ba59ae0a87f73b17ee1e6bc78a1ba75c50c81a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -79195,7 +84574,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm103a}, -{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len, 123152, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a", 384, "82339e09861f14db2d1fad2dc77731439c87664aa9ad0f73c0540549a189118a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin, Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a_cubin_len, 123152, "bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a", 384, "017483a7a28275474dee4aae31f28b303326b4eb9c2ab21aee18baa0696f9f78", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -79321,7 +84700,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm103a}, -{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len, 114320, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a", 512, "260495bf013585f26af55fa99ae45a7c40c4aec1cb8bdab53f5240272f947dac", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len, 114320, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a", 512, "62e09e19c5007265347c6ed547063697abd73e031cbccf827960ac1e4ceb5d3d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -79447,7 +84826,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm103a}, -{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len, 114192, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a", 512, "10e2c9b2fe1f4e1df09f67da29a63d8cd17e8deb26c77f8b31ad61ff79cbcee9", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len, 114192, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a", 512, "8e8e951de77bb240277f78c4ea54a38dadda9a90600c9979cb13998e81bfecb4", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -79573,7 +84952,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm103a}, -{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len, 114320, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a", 512, "c7e4e00a51fab39a148cbeccb82fe3fb41b0a8a02dc22352aa5d1c9aeee80951", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len, 114320, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a", 512, "8998cc34f938c09dac4974406c091143d55f78c79ee28adfc7771c657a2ded6a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -79699,7 +85078,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm103a}, -{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len, 114192, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a", 512, "e5c9d4fb2e0455801021377d9ae75cc134eb8a9f5012ca2cb75474c021822b23", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len, 114192, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a", 512, "f3b34afe93be1f235c18c128de783ffea4fd1d4937652736a00843d64c47b0fd", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -79825,7 +85204,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm103a}, -{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len, 136848, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a", 512, "9ea1e535a600e70aadc2431a148ee0808e352600cc6a48f1b88edbea1ba3ad44", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len, 136848, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a", 512, "ec6f50df4e30d3228a853b671f61126797706754c17d2bc1c8fede256ec79e0f", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -79951,7 +85330,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm103a}, -{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len, 136720, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a", 512, "77ab2ba8abe56046ca5141a0675c8a7ccbe1ffa1d7e9cfa5b1e584b3be52f476", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len, 136720, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a", 512, "7bbcc7547de03175f0ef64e5591638103a6a9257e36c97f6dc0d7b07b38860fe", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -80077,7 +85456,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm103a}, -{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len, 136848, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a", 512, "a3147ed18dfc064b1bce8cb2f98ecd8733227ba79926e801517c6c62c380d61b", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len, 136848, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a", 512, "c43fec722e3117fbf04b2fc58eda3845672a27b39b622ff3eea3c102b6be9b1e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -80203,7 +85582,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm103a}, -{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len, 136720, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a", 512, "2ff7c54209005beb4038beb12b4cc8f506784afaa8ab8d3fed8233b5a6f5e646", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len, 136720, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a", 512, "d4b5c3ca032eee42d3ab6b4b42705d6b0b6a71fb3d701857238d97dc41be5d36", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -80329,7 +85708,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm103a}, -{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len, 149136, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a", 512, "29783f8f9bb85c176340f9cf212f7830e1f406941898cfd40411bed17feeef77", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len, 149136, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a", 512, "5b1decefee2ee17ce198e3c385ec43fea6650e404d7e29bf72c7ce31fa1788bf", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -80455,7 +85834,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm103a}, -{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len, 149008, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a", 512, "ed1dab836fcd8a9a62ce77a3f60e1a8478b7fe62a6e3d5d6b559d291b22c53c3", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len, 149008, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a", 512, "73b7026add1a62b756eae13d847dc221f9cccd20c1229b1aa55b95a3a82e9385", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -80581,7 +85960,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm103a}, -{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len, 149136, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a", 512, "a1390455e653639d438bb626808e704a7aee4753787d58edd063484f2bdbdbed", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len, 149136, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a", 512, "19dbc021a7917f38b488c4870e7e7042f031d3265c15ac4cdbde2153a2206ac4", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -80707,7 +86086,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm103a}, -{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len, 149008, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a", 512, "37686a0f0b1c832f774f803f6b84fd46163c3dfd945c9d38bd02e3a962cf0baa", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len, 149008, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a", 512, "b7f96a0e03e31cedab90ac6ecf317f70934a8fdedafc35ab7c0a263930777033", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -80833,7 +86212,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm103a}, -{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len, 103056, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a", 512, "8824a03e5abf2e1e80eaa34a6eb405b0003da1756f1b6f5c3785e1ca49ab4ff7", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len, 103056, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a", 512, "d6c26744f4cde9ec1d0aad665713eb26bd7114532776282fa19c57d07a4b5aea", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -80959,7 +86338,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm103a}, -{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len, 102928, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a", 512, "2d931660bf65d2c3d7b6dcc3f8f8916ab91ccb20f431e3086c98514a5a70e60e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len, 102928, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a", 512, "ed3bb5d0fabdb1686de53ffb836d502938bdbf891c6e61f987b1bec24011132e", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -81085,7 +86464,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm103a}, -{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len, 103056, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a", 512, "479f512d502398812f6b9fbd4f7c29822fada3be8d5f4a7c3469106035dcae4d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len, 103056, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a", 512, "ddb9612e7b4c6342f15e5e8dc4f477b0c3a10955b597a9302a2a3bc81c9cfd36", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -81211,7 +86590,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm103a}, -{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len, 102928, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a", 512, "6f9b00a7256107a70758b8726ad0703ee2ef75d22954773ba0d2f763dbe70034", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len, 102928, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a", 512, "bf6d254176bd63f7f7cab5f730674cdf2bb89001ffdce697a364ff4fb690e24d", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -81337,7 +86716,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm103a}, -{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len, 122512, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a", 512, "349cb6ad0ea69064545143ffd05def39871f06965835ff303c89b7673a48dc67", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len, 122512, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a", 512, "6872050f04924487d1c469faf4adc60e1cba8a2a6ee3b3c7ee3ac7c1ab022ed1", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -81463,7 +86842,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm103a}, -{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len, 122128, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a", 512, "d311e7dfd8d23393e285fb98d92a47221b75346b6ed1ece3af6c3a786a33dede", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len, 122128, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a", 512, "5d850d2920fee534e084d5cad22cb3929164dad2ede4abc4921757b69fde81a4", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -81589,7 +86968,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm103a}, -{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len, 122512, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a", 512, "4d4ddb5fd540d75e0ff63a7a13752c07019d87c9b5ceced557c467664aadb76a", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len, 122512, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a", 512, "a0b558247ee633addc04a0a1c4980b25734cff282d139c2492f2ff0308f1b6bd", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -81715,7 +87094,7 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { , /* mUseCMultiCast */ 0 , /* mUseTmaOobOpt */ 1 }, gemm::SmVersion::Sm103a}, -{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len, 122128, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a", 512, "9344afab994a47ea13ccf248aad2151f2e43de2005a826f18924f3ccacfe87de", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) +{Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin, Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a_cubin_len, 122128, "bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a", 512, "e7f6fdd2f687338332a529d60e0b621f87c39f2d005809ded172bc0e3af38a17", "", nullptr, nullptr, nullptr, 0, { /* mAllReduceAlgo */ gemm::AllReduceAlgo(0) , /* mBiasDtype */ trtllm::gen::Dtype(1056776) , /* mBiasType */ gemm::BiasType(1) , /* mBlockK */ -1 @@ -81845,5 +87224,6 @@ static const batchedGemm::BatchedGemmConfig tllmGenBatchedGemmList[] = { }; // clang-format on } // namespace kernels + TRTLLM_NAMESPACE_END } // namespace batchedGemm diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/KernelParams.h b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/KernelParams.h index 7650961d2f57..3983699dd43c 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/KernelParams.h +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/KernelParams.h @@ -232,7 +232,7 @@ static auto makeTmaShapeStrideAbc(GemmOptions const& options, int sizeM, int siz // Create the TMA shape/stride for A/B block scaling factors. static auto makeTmaShapeStrideSfAb(int mM, int mN, int mK, MatrixType matrixType, int tileM, int tileN, int tileK, - tg::SfLayout layout, int sfReshapeFactor, int32_t const numEltsPerSf) + tg::SfLayout layout, int sfReshapeFactor, const int32_t numEltsPerSf) { // The outer dimension. diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/KernelTraits.h b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/KernelTraits.h index 462658a518be..077a73cfa60c 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/KernelTraits.h +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/KernelTraits.h @@ -215,6 +215,8 @@ class KernelTraits : mMmaKind{mmaKind} , mFuseUtccpWithUtcmma{fuseUtccpWithUtcmma} , mUseMaxTmemOverlap{useMaxTmemOverlap} + , mUseTmaStore{useTmaStore} + , mUsePersistentScheduler{usePersistentScheduler} , mUseCustomizedMma3xNvFp4{useCustomizedMma3xNvFp4} , mFusedBiasShuffleMode{fusedBiasShuffleMode} , mNumEpilogueWarps{numEpilogueWarps} @@ -609,6 +611,10 @@ tg::MmaKind mMmaKind{}; bool mFuseUtccpWithUtcmma{}; // Whether use the max TMEM overlap trick. bool mUseMaxTmemOverlap{}; +// Whether the epilogue stages output in SMEM for an asynchronous TMA store. +bool mUseTmaStore{}; +// Whether work tiles are assigned by a persistent scheduler. +bool mUsePersistentScheduler{}; // Whether use customized MMA for 3xNvFp4 bool mUseCustomizedMma3xNvFp4{}; // Which BiasType::Mn preprocessing steps are fused into the kernel instead of the host. diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/config.json b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/config.json index a414d177c1ef..1bd90637b8bc 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/config.json +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/config.json @@ -301,6 +301,11 @@ "geglu", "none" ], + [ + true, + "situ", + "none" + ], [ false, "swiglu", @@ -414,6 +419,11 @@ "geglu", "none" ], + [ + true, + "situ", + "none" + ], [ false, "none", @@ -497,6 +507,11 @@ "geglu", "none" ], + [ + true, + "situ", + "none" + ], [ false, "none", @@ -563,6 +578,11 @@ "geglu", "none" ], + [ + true, + "situ", + "none" + ], [ false, "none", @@ -610,6 +630,11 @@ "geglu", "none" ], + [ + true, + "situ", + "none" + ], [ false, "none", @@ -663,6 +688,11 @@ "geglu", "none" ], + [ + true, + "situ", + "none" + ], [ false, "none", @@ -713,6 +743,11 @@ "geglu", "none" ], + [ + true, + "situ", + "none" + ], [ false, "none", @@ -760,6 +795,11 @@ "geglu", "none" ], + [ + true, + "situ", + "none" + ], [ false, "none", diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x128x256_s6_et128x128_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x128x256_s6_et128x128_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 38a41a94e3fc..9097a5dd8282 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x128x256_s6_et128x128_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x128x256_s6_et128x128_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:72865a03c95dc195560b5370cfaa527a7e66fe8f22522db0bf07f3d52b140c76 -size 56714 +oid sha256:3a3654b3080b6b3c1ea80255b6a20392d57850c84707765958267c8df73b1eab +size 71313 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x128x256_s6_et128x128_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x128x256_s6_et128x128_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 65a1d0fdc5c5..3fdcd312ec4a 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x128x256_s6_et128x128_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x128x256_s6_et128x128_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:1a28dac8facd8aef6c2d317418842830439b72b641a2a5734a6df3a93c8fd46b -size 59752 +oid sha256:bdb1108e466df4867eab319f98eded4bbe4bce92fbcd974ebd29d1791574ac56 +size 75626 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x128x256u2_s6_et128x128_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x128x256u2_s6_et128x128_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 95c2816621ff..077a6998e52d 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x128x256u2_s6_et128x128_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x128x256u2_s6_et128x128_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:c8476135c22f23c5da3d32b311882410b13a50e313ea3c149baaa1e5594f96ed -size 59702 +oid sha256:097e07bfb9281c775d994e4647224a04aa0b0ef675fec71759acf39b858979c8 +size 74850 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 943cdb969146..bfb21a3d71c0 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:12eafca1284eca5dfef2341f6006afc6750a3e22745a64abc277e73da4516b62 -size 45047 +oid sha256:12678bb5025d347354adab890deea619571c243a675b2395fdc6806e73549f42 +size 54971 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 0c671369e762..cbfee22b7258 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:821d8daa90c238a88573c19bc58587170874d8930dceda6d7801fa9c9983572b -size 31181 +oid sha256:0534cc9fabeaf02d33fb68bc7e7bc30b87576e2311af18aac9bc9ffd1f3f2861 +size 38045 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index d98dc14c30b1..b260003aae21 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:9a92bcbbea1c255499de3a054d5486814109fd0516b13307982c388c528fa821 -size 47539 +oid sha256:19c9da058647b60a8ef700503dc332e8cc76ad17f0588dd0f7afaf4ab803ed8b +size 58804 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 55fd7d890258..c52345608be3 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:4760aa7b80aa9bcee2833a41e978e8e505f33544bd77e78fc766f75e900b0e89 -size 34419 +oid sha256:3bcae85a5639f83758dc0dc9256cbcbae30bfec1369f7226057bc79357bc6dc6 +size 42123 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 18c57fb19ed7..12b111b5ef52 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:74e31f7c293b7767d030ed596d6869d8bd12884e47d798b3c193d26a7e7e0132 -size 52217 +oid sha256:6b73728f210f9a1908a7c0f62027db404483804dee91478ff99f1491eb46d967 +size 64550 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 1f5825b525ed..4e8686ffc198 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:5e877a928963b033fdba07cb9fbe3ca0af4a3d6677f15b8e4432ec35b1574d9d -size 38463 +oid sha256:6e73986869601a8c53c32ee95358a70d497bfc28ff16ac5c07b39316949be56d +size 46984 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index ff0b22783338..a3b6cffb4231 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:55fb411e9104e4010e9c41c524fbb365a2d34cc2060b5742101bae0cbd14af0e -size 56398 +oid sha256:8921db828ca8cb01bee4ae6aa3f74d705d77a8991b2200e091c7d98b9fc5451a +size 69875 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 7f62072b8ec6..16ba65302451 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:84ecaef85326a5b15242a46c803c295615b6c8c09ecfa41f504cd88406345803 -size 42367 +oid sha256:99bb135a5aa7502ba615967e63981f67a8b5e8ba8d86cb0fc9466842c29cb298 +size 52477 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x256x256_s5x5x5x5x1x1_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x256x256_s5x5x5x5x1x1_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index f200a2f81162..f4d78eec7b00 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x256x256_s5x5x5x5x1x1_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x256x256_s5x5x5x5x1x1_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:5d7f298dcd0b65e0d9660df5e5ea0fcfb4f8543e4b9668abc88ba6e7dc5ada86 -size 59278 +oid sha256:ade6fd5855637b8f5cf05c1a9df2cb3ca29d17c1a86768c62422198992a66fc1 +size 76273 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x256x256_s5x5x5x5x1x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_fCp_tmOv_bM_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x256x256_s5x5x5x5x1x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_fCp_tmOv_bM_rgTma_clmp_dynB_sm100f.cubin.tar.zst index ec7e48bd0b3e..87754202ed30 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x256x256_s5x5x5x5x1x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_fCp_tmOv_bM_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x256x256_s5x5x5x5x1x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_fCp_tmOv_bM_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:0149eccae89ac422e36c915e980eaaf54eb9a50bd87072ed03e15250e189e67b -size 69349 +oid sha256:9101e7104ee11dbd049aab122e38c30c2489f9c55519de7b2d469d074b0d31a2 +size 88831 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index f8d41cf9379b..2ef6439e9730 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:c4351aaa4c8f52308c675bfabd7d06b82ef48b66a642d0fe7e5ebdfeaa263714 -size 45957 +oid sha256:9c2fca727baf35a00650fa01e486151252c492833156de3a5041eef35f12b47b +size 56360 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index d4d9966b8728..79d540396c10 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:7e8a4fc4a1efaf58fb321d9616548b10b59e9341d56867d6f344a9cc66b9a2bd -size 32547 +oid sha256:33ca69e20a8fa2321191f2624d9db9a7a9a5cef7e7954ac051048daf4ec60b1b +size 39792 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index c051daf411a6..32699cebd8fb 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:4a9eb1e0a5dba47adc591bd35ec14bffd5819cc11d66cc96fd6c1146e72ffb84 -size 48605 +oid sha256:6d557be33a0d9bc84b26f78d1ff3f50ac34c6a4460d222f7db2514cdfa4735f2 +size 60107 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 9a15dfc88b23..33e8c5ecf78e 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:fdb40c9e1ca7510051627750905ff5a12dd19e4b5eed584ad8f13ce8d503fa89 -size 35641 +oid sha256:ada3d6e44e7958c85f3d2b886138ee58e753204ade5e6250f4dd866b9bade2ea +size 43860 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index ae2031e3ae32..37fb6746bb82 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:b105f6e14087a91775da03a85e4340a11296ba6f5b52cb0013269cba1f22ce20 -size 53377 +oid sha256:8959357c32c44e617dddf0075e88a0c9066b751951b8e861901f980bc5becfa2 +size 66280 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index e3e0890f7e83..75a29e07fc98 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:0a76fe04064e1e5607c63c4bad0a657a51aa406d62a6b903f11e36a64ae78a33 -size 39556 +oid sha256:d9ffde2de37cf23b402f2c3aa4077278951dc2bdc280daf7f12370f48493237f +size 48383 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 3e912d3fafe0..e0af7d87fe10 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:7584045facf596d373b5f0c79b9c3b887eba25c1b0fce0e80b2ffeaf9de4d88a -size 57254 +oid sha256:9c51161aa1c5a8c671460a39ca3dab6db8fe9bfa713e5410591f54cbbf46faca +size 71530 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 4db5bb79b901..d018fcf33ede 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:ac2d9dcbfa012af67a197c53d4499e74e492fae16312bd14e11f1347667f959f -size 43528 +oid sha256:2a2fc6592cfb1434cba990b4fdecdf6d9e6b1bc0e2966c913f43a96e3b9e2fd1 +size 53882 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x64x512_s4_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x64x512_s4_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 1a8feb3f53de..495c1f66f628 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x64x512_s4_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x64x512_s4_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:a84bb2dfc2f8f86bc1d4dd24a358c2cc406f5bf4eb987c8a0ab1a2142c98146f -size 55439 +oid sha256:911f68baa2bcc2525600491ba26a6781f364d967d88724fcdef6b614e3dbef7e +size 69112 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x64x512u2_s4_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x64x512u2_s4_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index a87895fab23f..ab7a64d06018 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x64x512u2_s4_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x64x512u2_s4_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:2d81fe64a56da64c5971c2a540b2a67c0cc4e240dddde43ac4d94b01f21715f9 -size 59277 +oid sha256:fd56cc4810c054a828427681354103e66b16534db4df3b8bb4165762454bd5e5 +size 74116 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index b6d34d34b672..8d796ec518e3 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:c0326c868044643adeb64487431833e655dd1f6166908c3f486f8b63fb1d3dc4 -size 44522 +oid sha256:67111b59c13c3c238671e8dbb0f95fb9d2f629c47a66568c900040f76a5a988d +size 54445 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 6af3439cb12b..a7e6954bc019 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:e74e7d7092d907f27c1bb88a0c124cdd96fc37eaef0f83e5aa82413e677c90f9 -size 30579 +oid sha256:0803138e214b9ff277a67fecbb204a8d926f2d34f7466cf0ad1d8dac3a503bd1 +size 37009 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index b457a6c55055..6103797b0439 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:4430512c817c8ffc7076ed62a8702b6ae5b3b315b6e31dd83eef3f64066f89c5 -size 47187 +oid sha256:dc3aedf1401d2b1635d564c747e37a002045a19183b796917aadf1c397a6e64d +size 58388 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index b6e09d40bd4e..b14814d28db6 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:7c211df05929d1ed697cc050ff64bb078154a12ff0cc20a5e3b263e20572adc8 -size 33617 +oid sha256:3a4c3bfa480bfc7e74d8d830b674303957e86f0708a1b4cfaacdf53af6d7472b +size 41188 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f.cubin.tar.zst index b97f07143033..b9e0419cccfc 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:4986a4038939d3c0a8e7d463ff4273ed63904ac27983ac213b0218bc1647d882 -size 31733 +oid sha256:e4931f90aea05e4f6505b103be61e8ee1c4a0b14c733c12ebe492538b476365d +size 38507 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 2cb128a8605c..881ff9d314c1 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:9b53fbb1350459902cdaab9fe40db07b37c58e96374f40dc0c74f7570a58785a -size 46009 +oid sha256:8e20d48f6c3a5beb5be809dd2fd0fb6cc4fc931196f565d32322d628eec9e2d4 +size 56381 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index ce8530f1b5df..4316ccc6d825 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:b1a49873eade749350d6da1ab0b531bd88754c08a0bdfed998c174eede50e13f -size 31986 +oid sha256:dc250987eb677c77c7424113ee10d2b868c02af2cb28602ae44c041c46cef9c2 +size 38993 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPdx3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPdx3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 244473ca0949..77ce79c02624 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPdx3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPdx3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:c308b17364776352412f3f2cc59255e5deb8e401e0dc611db27b2077b94a96f4 -size 55848 +oid sha256:fcbe3ce5165cbde782576c23d63b4d4a45b9797db4c52bf8e65a146d2d3035dd +size 69687 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPdx3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPdx3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 02362b07e040..daa320c2a918 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPdx3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPdx3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:9931b594d0b7d4c43a5cbec103d6fc18d00ee017c59654d4ed2089bc856ad418 -size 56166 +oid sha256:1a41c2d0b4636ee0b9c136d248c9f2a1306ce1d308d5630b86e3297b1cbca272 +size 70410 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPdx3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPdx3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index aaa3876f0fca..00af59f77aa2 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPdx3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPdx3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:5a363c1918406d891e2122a4bf602f4b5c34da0e51a5e981e3b797070d55f73e -size 56065 +oid sha256:b4059eb8f23d2d6f440cabaf335c9dcbf63af46771de3ef5683c80b72176fbd8 +size 70471 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f.cubin.tar.zst index 65f35cbdcb94..471f67b7d16b 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:58fbb3626034a9238223cf3a434500fecb9ad70d0725cf2e4c8464eb9baf8177 -size 35574 +oid sha256:b5277c258621210ad4e3ee3d42f30ef641aaf4ebaed09cb5a6e2e40ed374dda8 +size 43676 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 8689edcedf01..f30e01217f3a 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:0a249ff45e649574a88a04303f97552f3a81a35e9938482e8a2f34d30f8d9027 -size 49499 +oid sha256:421037d5b4184a25e2bf4bff0409a0804a5773e234b8cf940def3e6b0b1f7c7c +size 61585 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index f3eae5c94ce7..f0909f16ed7b 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:605b8ab73db7742035df540bbabe0cf33d7d87b52438a6fe7d40c98f41db8dad -size 35784 +oid sha256:198fe6ec3f0f4f61891a3898ae918dfdbee2fe9f240e145b651672c46a33c4d6 +size 44056 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst index 59a87afbba27..71cf90fccfb2 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:4085a54c25a50b0696825aeccd6f41124e1fb962edb16d13aa0fbf6e8e714c83 -size 46353 +oid sha256:75c6476485f675a76db80ca5d9990175e059792a631548332e4a132e8d52d6bd +size 57038 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst index f0534460a6d9..8334f87afae3 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:74afc2e3cbca778ca7801e183f850fb9e28a4613514aa5c903bb59b0755cee92 -size 46209 +oid sha256:4fb87cbc23c98c6edc1df6cc598964b644ec91753b8e080b69a6e7c4b7f161c8 +size 57037 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst index 16a0975d25d9..42df9682e3f8 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:51e8fee541b65e1d7535aa8aa43babe1c5f0fa9da5c97d05f861a52a13a9c1d6 -size 33869 +oid sha256:8db14aeaa6286d2aa3b88be3d79aa8b0200e7ceec8224992323d3c2f704d9f06 +size 41217 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst index 37eac6947636..3e4ea38e88a5 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:b47a26d6eca7cc389694755dfbe41c6303adeac2caa2eaa1e4548a9ffdde7fb8 -size 33885 +oid sha256:fc95e3ee5fa482b97f288eb14a8e5da1c8ffad151085fa1d6939acd08e049ac5 +size 41175 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst index 3cfeb078f116..4cd46fb36e44 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:22c64397b9e044eadc6a64afc055dd3a7646bfa02f19bac064ef34f5aa43c181 -size 48643 +oid sha256:5428178259d8bed19d5effbb97e54c81f98d50763952a290fffd2449aea20a09 +size 59944 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst index c1b1e12c39bd..46458f984833 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:bcd430b11921983a1ea9fd0847be7d4b2af58c35e20faf015bd5a520e78624c2 -size 48683 +oid sha256:53a9e80404a6f33d4948894ddbd59f1eb6e515b5d6bbd4158ffecd502946c426 +size 59981 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst index cc6f4c8d98f5..70c85cc1bb87 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:0cf5864162fc935c4be88fc654349f2f1cba7ba013d4159571bc374af40860d0 -size 36201 +oid sha256:35d6428c97994e020074c6467d68a183b7fd41be9ee448f91037acdb6946400d +size 44445 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst index e6c6cce2e735..f48e329d1215 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:2978bf3fc06cd3230fdb7ffd287d79d57e9796b343dc8b3d5e5df947e44cad02 -size 36190 +oid sha256:9d0256b0dae725ed45f9e81a368f5cf9e971591e9aeeb9ec141006e24a17b622 +size 44468 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst index 61780e5631ab..088d4d0a0dce 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:d35d6c7ae9501a1cfb1eb9f8bd17cc60edde66d97e68e3721c4e92d81cb7c76e -size 47478 +oid sha256:ac18d8eb02da40ffe92284d1d0dee2e299fbdf0232af561ce77e016622d3d059 +size 58413 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst index d22c7bd067fb..53c80d1881f8 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:a55cd59af31012b085548131c97d2b6f8b0c373d6aff44031cf0d222e60cab97 -size 47450 +oid sha256:11752faa4ad0f49f1ecce87a115637a61f7949cc5f548db70beab3cdd9b00c75 +size 58402 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst index 3aabf957c9a1..669c1716d209 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:0fc97c441c33c8be29206a33ad8a17b834d1fe4ceb044a32b875670809d0d6fe -size 35135 +oid sha256:86e455eb43c3cc7b7d553e8e16a4052d9df7d62eaa1fd2a35a8f1e220e4a963f +size 42915 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst index 002fd87b0e77..3ff28a100773 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:2445181254406fe2b465847bb13ed2a9b9e32354146d1f157116576b447ced6a -size 35152 +oid sha256:dfbc4cdc32287365adb1c54b41e5a6c3122050016c3b7d1babdc3230fea8cbee +size 42890 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst index 8a5e74d5492e..9073ffc97da0 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:c84cfc9c0358b4a6b84cdaf91864c3bdbf341e34f0c9aba2b1523575a3f4159f -size 49660 +oid sha256:2eaf86c1c297e2da631c93211fc456017488ad42406ca0aa51becb82b8d0fe15 +size 61549 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst index 3201c700a48b..09bc34ec8381 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:77b654462e57588a6aedf0472b51a989992c098e2c6c5a1fb759d4c065e3dbf7 -size 49698 +oid sha256:7c8ed500968a120702ff9c382333528e153a9ee87a66cbae519f1f4005e28911 +size 61642 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst index 9193d37bcd8b..04ecc7c1d079 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:f30d0f6609fdbe199a7fef1012bb65517b4234bb9c889f8c019d1529d336b5b6 -size 37446 +oid sha256:d6b8de79e1c797ce5969ab077ad40502229d29b2be96bf634b1156d2d796cd2f +size 46005 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst index 027843084233..22bcd0fd909f 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:9aa9eb30390baa8bddd3a377182009b2c607b2ca23a9a4c35d944e4b4f55aff3 -size 37449 +oid sha256:3721e13d97f5a892b26f46f1bb39a0a2638928805d35c1d5dcd3f6cb459832a9 +size 45992 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst index 2009e3c46fac..8f08f0ee2a69 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:b2e943b1bb5515e55424645e6ead5566f2ba78480b663e53a22aa336c9eadd4e -size 49020 +oid sha256:e1f2bdcd82f942e32730ae26540dbb0c2e127bf27924323565284dbc3cf8d62f +size 60859 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst index b9780c3592ee..f8846da66f13 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:954983ec994893f0d3fef11a31e5192304826db34abbcf9464736bfeab54ea9a -size 48956 +oid sha256:f9a6353af88b97fe49dea671f472c3cd5c7564e990ad0ef1fed7e40ab1c8173c +size 60849 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst index 51a6a96b12be..71a682115497 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:7e50908288be50474f4ec82df57b1e6b741dcb3e77435325d7977f24b9fc54b8 -size 37267 +oid sha256:b7b9950bf10b1b5d1ffe450f9b42f73ecd0b64ccd41b9863d9bed1d6266f92b1 +size 45772 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst index 22f8bb871d92..3c63474f7b02 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:9db16d128a153218546fc268b5bb0e35594059dc38ddd8fadb50b032b8be9590 -size 37292 +oid sha256:db39114f4153556354a42ed61904fb43770957ad9c0aa3bd3849e87c8a0a8f49 +size 45825 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst index 2ac9886c6f10..6c505f788e58 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:53c8bcf247ce0723561c58a673ca4db65d68445d57767e8c3b6b5482b78997bb -size 51549 +oid sha256:7364eb2a14b9eace2b011c12a3ba06f1fca25a06bb98e43888ee88bd686e98c7 +size 64409 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst index 323568770340..6a8d61f5b00f 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:ab7a306e9412a6754402f69d37e583b0101aec506cd0e4c82b19d60dc4b9eb5d -size 51591 +oid sha256:914d6bcef9b8eca3c9a0b3789d01a8104738b09c6fed079209f0c379d28430a1 +size 64439 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst index ffd14b91db9b..219b95d95bc0 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:9a25d7170d5e57e4349b3086331275a3c01e4cec0662ae800fa09a365aaf4b9d -size 39489 +oid sha256:34ef0a4a073b063caf0d3502d6b6fa08f28a64842b942f32c188123c7a6ecca0 +size 49223 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst index 9fcc16deb41c..154d4ae73d73 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:d2cfa1fb837cebe8ad782e8d5c85f776422d070a612fb34c2f51f8483536e1c3 -size 39587 +oid sha256:139a9e5e1ff399f4dc79c63b4d0c6567673ebdf99a8e6f4af0ccde20ada6e590 +size 49235 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst index 5ced41f42f8c..a433a07eb478 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:ea580334e1b17bf10706f1524a6abe81077f6555e439c207229817d6790edaaf -size 45922 +oid sha256:7734adf1c9444141a44fa6f4841280c57a1cb137f8e7506a97ae34d81292a038 +size 56372 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst index c1ce19a1c173..1f10485c3fd9 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:8aef1f9e4d2be7ccf788c8966d664eaa1fc52c369119edbe08588743dd021054 -size 45958 +oid sha256:80464cc46381d74f92c422ee79663dc5c95aefa449f646f92eb29928b887865d +size 56304 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst index bb9ac4de4927..6f330c27b1dd 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:e21cc69fb82cab5f5411dd6236a43b11c314ab44c67047d77aa8fdbadacc6c74 -size 33221 +oid sha256:56eeb0eeda6c01189138672caa1aae5245615100af467b48f79b3316941e10bd +size 40413 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst index af450e669356..dd2583ae49f7 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:d260314f379bc33a7984b87d76fd5fc1c4c50478751e05dcb1fab97f98b49910 -size 33240 +oid sha256:3ad66e1a0b90646e3fed44ff2efed5c4c1e9ff8bd5701a4570ff7803d06026bf +size 40321 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst index a7d37724d632..b30e0143117b 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:5e6f86a12f19131edc603fc871c8571528f2b50c027293a425f3b9f5282a215f -size 48099 +oid sha256:56c354e3574c1556412ce2760ef6a3a5a9e26c0d9804cdc43cfcc76f649334f7 +size 59434 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst index 9042dc200c55..863141fa985a 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:041aaef5aa30bbc2d91503742baecfd33c7466d8857aa2e501d911b7f8327bd3 -size 48097 +oid sha256:708ab0dd6e1bc43d019c977b51cd48e1109a369daa3cc56893e9e082c84813df +size 59387 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst index afcc7fa32610..e134b0431e9c 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:821d6e32a729d709c0ce7b7171ebf616622b2f12a1970200f408011252d3a7fd -size 35645 +oid sha256:e6c91f51212d6ffa72224e346b3002c8815b3455fa0d89d6e9535ec11d9217ff +size 43742 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst index 032d31be3aef..838b8bb577d4 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:d0919cb16c2e02ab182c32d74f30d07cffe5c3ade55980f2aece9d7b180beead -size 35669 +oid sha256:7f11c9d2118cdfff3b2058ce55308e4cdbc14eb1022e70ce79e1c67dbe4a43d1 +size 43710 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst index 761e9c7427e1..e92ed325cbf9 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:de7bba494cce340a4fc6d612660bc46de3477a1a856cf73434c9fee98dc64b4d -size 55563 +oid sha256:91546e36389e51d71d8031edf0bdd06dca8e4341324701b5bbe3064cbbbbcbe9 +size 69705 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst index daedda5e1cb7..d54a7cf948f4 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:ce0253883dd68d8f59559859830293e76ae32c86a37af740fb21d7a132274973 -size 55722 +oid sha256:dfa58e8040a6066b4a3ede1a963be83a0c3327b1d64c4065902b85b5da27db9e +size 69754 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst index efb82c1c4f58..bcd39fc08390 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:e36437c143a10c49325073fa5f0b8dab5762fea1563efc11dfd5742641bbcfdc -size 43943 +oid sha256:13f7cd53c3240468cff6e4f45176936fc7f47bbf7bc9aae985ed0bf82833eab1 +size 54643 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst index 539bf0789367..108a426b3cf4 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:0377812eb2a7ebafa405da2232d978835c0f35bf99443bd1ee2fac82043233d1 -size 44033 +oid sha256:eb100224cf1a93be3df8b1b943ac8a4d834b89435d64a0cd77823e73454a7916 +size 54658 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst index 9041f34bb2aa..2dcdef593ea1 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:4d40939eb06160189b757fc89d33a177d0c8cde322eb19e1f12659dd3043674a -size 58742 +oid sha256:fe46562981bffdcc910c7afa2ce150fb4752e79a2b8496e59b44fa6557e31878 +size 74011 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst index e2fcb9a99a11..afeb749e49bd 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:5a6d84b69a643e74328ac112623bb911fff0d6e06aabda69c7561d18a943ae84 -size 58754 +oid sha256:dab12639c44c610fdb4ad645acfb3cff7c8b1638fd79e7568e74fff0ade2764c +size 74039 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst index 71cf7e733095..71b9b6c6bb9a 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:ee9d67a4bfab5b9460d129e1ef7325382094b055150b73e7000d0b1da20015a2 -size 46781 +oid sha256:2ee6cf0cda8fcfaf38c5ebbbf96aa1ef56d7a637b0ad64b5cc1501911d7168b9 +size 58398 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst index eb7a68c068e2..1bbea755cbcc 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm103a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:9904b954fb0f3d4b40971412455bed076a3e0515cbc12a9b645637d5d8ad6092 -size 46796 +oid sha256:eedcc30c3910dca740c49abe7e934237bedd1deefdcb5c7b046f71a36d92ddb7 +size 58399 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 317453c88009..9b93cfcef02f 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:41503ae9701d16fa86271618b2c75147d151eda0c5c85426f7bb602ab3fdbdde -size 65762 +oid sha256:8e920808588978e12dea1ea70aa291d5cfc44718d61d99db47b8b2035e28164c +size 84450 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 679390d485fd..7cbe99e3b109 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:82e70a5a03ce758c30aaa416bc080ecae5e50c270bf5e38060bd818e37d0ce10 -size 49736 +oid sha256:ad984f55e6d5de638d4b1381f575dd1be3d4d6a20468936b94951168a03f0e26 +size 65629 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index ee9e347e455a..91e7199cef7f 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:b5754ad584207d68fb2450df5c02a900eab35e16da49a65e901c77073a60f98b -size 39937 +oid sha256:67d12a0e2596361479ca82c42184613c0ef842006dff943d25f29425f544121a +size 49266 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 165d8eea580a..81bc979937bb 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:f2b8cccdc87b7d6244d58b9f55eb3b7dda0b347f3e1c9d5402798628e3169c4b -size 68925 +oid sha256:5efca6582e57e24a9ed8fa91f7eea08a458c29412a5d90cff719a1d1dd05ac99 +size 88320 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index ea4cbea7fcdb..8884fb4faf59 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:99efc98e6803c952f8aee9f2def44b3b000b90d5da0f02173d6537dff8232ce5 -size 51826 +oid sha256:fb288c2b6d2d52eb15ca02b530c94eef1f87e5004d2d99ca70b5e3298ff6bd5a +size 68260 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128u2_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128u2_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 192712b504f1..c3ca36977f83 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128u2_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x128x128u2_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:1df3ce74e579e8a00f639f0bbef9759b2632f293bc7283bcc1a960d728b9a766 -size 42131 +oid sha256:bee07ee799e9981749aac47225ecc723aacb8e409372d0f0b2b93d2b2d66f593 +size 52009 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index e1ca2388ff3b..1bbce2ded80d 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:123c660513796308c3fcfd08aeb6cb7197202fbfad4655da9cab46f94438c949 -size 43862 +oid sha256:b16dc09a70bce709c083440299a39ce1ef876b053f18b610c2450122f18741a4 +size 54325 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 5e5968fc31b9..1b949c05f954 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:7005c2851d7963014d50c57138f18fe837bf1a5aec8c213bcee7c2c408c2945b -size 31937 +oid sha256:f351427217a88eeb8e9eb43c36ee2809e931dc5aac8061781f32d3b165eb3239 +size 39071 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 5a6fcb263944..6940401aa407 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:dd3a61f7f95b3f0e32fcc77d243d6afb94d084ffbb42b0f8e663049a533f7887 -size 47283 +oid sha256:c7ace63e5c53137db5da56967b028f2cedc671694ba4cafbaf33e3956ff6d42f +size 58703 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 420b9e4d300c..b52ad8be8c6b 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:ba66ed0ac7ecc929a5fd93cc50f4f881ef2cfc478bbc46bd42b7a446a1501fa4 -size 34292 +oid sha256:5c54db4fe107cb2145feb9801e87c1d7bb78a2a12b3c9b34e428a4dc4087accd +size 41939 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 743e891a9f32..aded780be128 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:04ec7ad8671b72cfb042476be9a113d2558bd1615750fa3e2b625cd9225e141f -size 30337 +oid sha256:eeaf714477117aeb02a2fcacb6989ae20645d8c0e1b13da1d28f8d9b94695bd3 +size 36425 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 263b5d38c4c2..b905971b7997 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:d983eb70b9044b556245598686247da47c1122813c605b9d825810ccc7d40895 -size 20475 +oid sha256:11bad53549e67e9adc20d6e9424f0d349552ecd20a6317e73d32e861455ec3ab +size 24308 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index ee319e642aec..fae9ce1cbc7c 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:e4afb0445a404e4f8562d9119af6771fc1a2c03e948be2dd16faa80eed74a4df -size 32713 +oid sha256:e9017dcd6fdd287c8622a90329ab6493f8fb66487e333730f8dfa796a8cc28f4 +size 39595 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index f02d5955513d..5b71422a601f 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:99b344497f7cd0582ce99123c35f48dce598ef24f98086d2d0374d8e85d53439 -size 22996 +oid sha256:2174213f8d77951a3d7f80ec06917a5b3a7111bcd526f3ff38c23c043227dc61 +size 27270 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index e508e4f13452..ae6fb74bc911 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:b46a67b2dad8a506583179833ee856896fee9402a5b4837aba6c6f4b162de19d -size 45045 +oid sha256:f6fb3662d41e597026a53f58220f3e1c2504f6bb18d57507600e33af535a1673 +size 56284 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 4bf8169258ce..f0dd6a591de9 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:9f64ae745037f52d8cc3e19bbd3a39bb1435556dbe1e30c0585dc1c1f477d1b3 -size 47024 +oid sha256:7016362d851c154ae79a7baf4a7ac7862d6ef3f075b0ee4e017fe1af755c740d +size 58454 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x256x128_s6_et128x32_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x256x128_s6_et128x32_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 8cca9cdc7f92..c4df8305b9c1 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x256x128_s6_et128x32_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x256x128_s6_et128x32_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:d906e232bf910b01dce31a543b19d2e4e900afe27dbf6f4c987a0950e91abc6d -size 40693 +oid sha256:417a5455647aed04743bb6e1afac6e6765d94664dd4a0c2ca5f61da704db6a66 +size 50058 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x256x128u2_s6_et128x32_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x256x128u2_s6_et128x32_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index f7e5abf4e786..ba52b7c4691e 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x256x128u2_s6_et128x32_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x256x128u2_s6_et128x32_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:55748f0ab0451bfdc9f5c68f72cdfc2d6ee11a5527511f8a13cc20db7266c94d -size 43007 +oid sha256:5f340111ccff71db4ca1c2604f3cffe843279d693e3bbfd7dba7b34f71774eda +size 52941 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index dd0ccae778a6..65e2a678aae6 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:284880cf41ffcd0f93b17ff5489e133011f2c9dc1ee6a8b1296a52947c3395df -size 46785 +oid sha256:d0dea51d99795fd896824d8a221799f54d04922a803bcd0ab24dcd8a291ca010 +size 58562 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 2fe52739b5f2..7d73af827dca 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:7c60cf771da96ffa94ba2edf39f7791b031f40de4072060f59d68caccd0afc6c -size 34892 +oid sha256:7242c17bdb08392419b466531af86a8116e71e24dae19ab2bb47709d7f50471c +size 43211 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 6d97214c0c72..d61916e2b4c3 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:8b5f0b752dacc0c17488fd0a311183b92fc46a0c52fbc4cca14a8eb531da1de5 -size 49791 +oid sha256:289dbd2de176f298479c4febe12956bb63f6ce5807d175f5eb2b056c3113a838 +size 62623 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index f6bd84389824..ba9d09a6cab8 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:8efabbae6d7e1c2780c7eeb078f76b12e3eabe4c640a890d35f3a7d95ea61a92 -size 37349 +oid sha256:08706354b3c627a90e86dbe552f8a2c1712e87984914752fc277083c85724b64 +size 46442 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 96968153af4e..a4c111cc5758 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:38cb94f27d6704b7b6610e5efdf1260369464fafaa57f1d7f2509508f8ed0464 -size 31144 +oid sha256:fa472bce0d48be79b4fb210ae67e25b7d5fc1736b3d892430089f7301c42fa12 +size 37549 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 6ac6e2b6a1e8..b1b5c1368f0b 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:d4304b677bdfd7353184f17fe5dcea216566cd24be689f814752e488797bb2b2 -size 21698 +oid sha256:3d04c2698e35d925c8345869a6e69b8fad731c6064c59fbaad66193c36877fc4 +size 25815 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index e12a386c803a..35e3d5f3c77f 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:4e5a66f64370cc91ba2c7854e7a8721130d1c3e633c33746615a7714a3edda79 -size 33599 +oid sha256:b00ca15d57ef8d497dc5282e9cacc913bf4f9276f86679815cf068083649448e +size 40744 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 5c38c52aa9c8..5063cd51f7d9 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:eb191fe3b932bca060ca66789a9e8cef702daf1e8e1228317a9871d13c7d41a4 -size 24094 +oid sha256:b26a0ea24090de21ee6e9c5bba5c4b42d6857509ebae5d29c7b750fa002b6553 +size 28659 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 995f836ad6de..e9b2050ba0ec 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:d01dc0b71e66069340d7f26222bb4ebb79400a911d0a7db691edc1a99df2c10b -size 53952 +oid sha256:1b2564b24d308337ab3c9396f85187951f307e97006f5894904f192492fb0fb0 +size 68199 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 48a73ade59c2..0178d2a53d45 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:7562324fbbc588023fda02d25feebea738d6df18d659e08f6ab110603c8863ee -size 40063 +oid sha256:d1e3707debb9d5b0ef4e42f16c8cf4cd8af093399b33b62b65c0b26512d35ef9 +size 50982 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128_s8_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128_s8_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index a6ccaac8a100..1b62fa30b20b 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128_s8_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128_s8_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:6b758c3ac0ac190c3f99dcd99e485b068e95e29e752f65ea614a5600633eec1e -size 36328 +oid sha256:06c3192aba0e1bfe895397abba7fbca5f9cbd35f145b0850bf3ec79efde41c11 +size 43967 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index bb6a2b4b8401..4e5dee2ff58f 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:40b114954a8613d9f6bc513d82a975cc178f1dd0be301a3fde26340bbcef70a8 -size 57196 +oid sha256:de613b15c3c0a49399c72eda1f7e638f7bfcdc9729ba31c02a48c4be3364098d +size 72767 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 68b96783637e..7969b4810de1 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:ecc8f1782df676d8465b1652787d7eb1e0efd8eb521f1648f1982326929eef15 -size 42514 +oid sha256:f3b3dcdd31bcd19276498b99edf283e2d998e2731d7145f39e3785424e717c84 +size 54464 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128u2_s8_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128u2_s8_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index daf76c2e2c1f..82acbfbbdd9a 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128u2_s8_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x64x128u2_s8_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:02d656ffcf358474c9b57de55e08cdfc922f54e160ac243c4acfeb4b376589c3 -size 38332 +oid sha256:8737e4745ce414b070b9eca7bba6bb24ca80ebdbddaf9266d8cde40c996daeef +size 46437 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst index b68bc8a5590e..5ac1718b3f97 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:712ec6fa6bd6a461f330e1242ea05e77669cae3ff62e5ea116b9b5b1de43d5dc -size 29787 +oid sha256:f68a98c63a5707c1b285cde888575df5f47a9b7f1de7517aab2e6797acaff7d8 +size 35943 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index f50fc87678eb..20baa6efb1a8 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:67af5dccd88d4232e8079bd047f290986eafed8bd06e8c4ecac3bc95b4954037 -size 42534 +oid sha256:7ecc045dd4cfc01ed83eae9e63f8f96386e4863223175a0c82533d2fdbe4f739 +size 52149 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 4d43412abf77..b0dda482a30f 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:7ea33d3c6b911c2b38f35baf91649db11c7b19c65ab49eb1dc9a3003316502c0 -size 30376 +oid sha256:e79d71aa893d7ae0f6b22caed7a8ff344aa660a74d674b676fe8c425d91d4c2f +size 36812 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128u2_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128u2_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst index c5b1e66223f6..66e973bc6058 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128u2_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128u2_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:d6c3e37b34cf7f31f1e38148ce8e98c2b7654f7f1bc6b64e1dcfb4560171b515 -size 32422 +oid sha256:35efd714e958551460ff167b652c58c68176e1633781d06aa93a48761fe51499 +size 39300 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 6a63d2345854..8d807ad0b306 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:ec1d73d946e24d08d749c278d3cda9621eea1115acfa52386c431180d296fece -size 45876 +oid sha256:bd152606b5920a6ebba2b2f1079befc5899574edb963a1860cb395ff1a84b7e9 +size 56670 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 8c1a6c638661..3d341091f6fc 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:aed3f26c8cac91642cc817ae818af041c65e99a431a282df19a422d60ea17550 -size 32523 +oid sha256:6fbb553360f9f7795beb66faeb82de35747aa74ba643546fa46def5baf6e7ab4 +size 39532 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 80800c0b7fbc..df088788485a 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:bf84be129fede8d764fb08fa42e91a16fdcbf01e12e37d3015622f6578414494 -size 29679 +oid sha256:ba89b1e4b0f0dc5b1d6617db004fbf24a40ba128dd8cd114c6452a3d3ac871ef +size 35690 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 772ea3f4144d..50c4ad35f60c 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:8271da3932118774bfacba7c8554eaf092134f5fb33549615aa80969b138ce6d -size 19949 +oid sha256:0364c5ab1f350045a7f13ab044269a63c14d4810dfa319cad8ebea206617d7d6 +size 23581 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index d2773ec8358f..7fcb46ecc860 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:f4a948d603ecdc3fd9fcecfa02d46733a2544c1ba922181e38654dbb33da4067 -size 32248 +oid sha256:f0e2a4920ad1c264c4bb1e202d8dd34ecb767090a05b27720067852590f3318c +size 38828 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 51607e214c17..3a28bfba1fb8 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:fe25a2ec024371c671964e2d71dd5124b84679e0d76317027dc3be63bd482c30 -size 22521 +oid sha256:3af8bbfac490c3d48b6cc57e2f5e2c83be7fef71ab96ff8f9ac913367617847c +size 26550 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst index 0322772b3f8f..1bdd1b618612 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:5b42cf1c8386008a37fefa89c33b5a0d2adeeac7e25f26f5c7f9e222a382114c -size 21446 +oid sha256:d83ff413aeec86d88f68517f31d6974ef85f22eff26195fe3707134cc7acf84f +size 25529 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f.cubin.tar.zst index f31293f5204c..2b57c480db83 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:997496af8f942d0b6909591a02e379154e2ca39613e13099306bbb9100b1559f -size 21899 +oid sha256:8801663a0a836d029a2b0ea707a708f7e66feeb1e3e37b5dd2351da4f32e139a +size 26042 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst index e89fbdaac9ed..7e31457f28d7 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:d14d3ca8df629bb0514f5ed8ecaa9723acd79a689cf17a52dad36736b683afea -size 24993 +oid sha256:9ad568456f231a62fde252e1ed31c901db7a2b42df19cc973fc8c721daf3fce7 +size 29930 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f.cubin.tar.zst index e7e849cbcc4b..54b1227c4c42 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:026771cc0d451929d2b70b1ba4c13ee8903baa3078ff5614c9b38ca6c2c7cd2f -size 25547 +oid sha256:2d2b73df5eabe6ad9779e6ddc9c48bced55995186b51c2e0428d18d0cf91f0cb +size 30605 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 2790f24ca810..4b5fb4116393 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:38a109e30500737d30af7e4fd757788e9852fbc8d829a8d7998ab111fdd4a495 -size 53973 +oid sha256:aadf4893b959b3b03e6ae47e03fc21855e91f2cf06aaf73d334fe2f01add7027 +size 67795 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 9e45a5b5a4bc..ce168ebf1b8c 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:8ea25f2405c3f8f783d4a87d7fe3f6b0898d241c8d255eb3743228515e7f0d63 -size 42431 +oid sha256:06c783af43f82dcbad5ca1fb3be566c3af5fffefeb68372691040267698f44cf +size 52652 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128u2_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128u2_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index a21d683d1dcb..defcd0653a24 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128u2_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128u2_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:9d9bb5c832d5cc516256d95b2c4684101e6ab6e9c58d8be296104c0281d551e1 -size 56322 +oid sha256:691c919465e1aa1959b40d016250d4ea01a0e5654d1908df69dfe61ce61b9964 +size 70733 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128u2_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128u2_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index a7b3cb5123dd..af57694e3dd6 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128u2_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x128u2_s5_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:7d0dd9abf026e17e9ce2b1e556984d2072ad96bbd870284449fc9fcc6a29816f -size 44674 +oid sha256:17e73130388a0d699a6d93a9ec689000311aeb664f70527624332652fc6758a7 +size 55539 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 537cefe38c94..da0e68082b27 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:f5b1a5c5ef3a5f4a70857f5ac9f7e22fe161ae8809b6a02b60297e85e8c74ed1 -size 64709 +oid sha256:767594667586c2730819df9b3b1bf728731f2e5c26d1dad561f6397188803ecc +size 81429 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 06b35c0f03d3..cbfb43abf031 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:8dc52e7951ff15f4c01688b38d68328271c7f6ebedd4ba94fd13d2949973b973 -size 60962 +oid sha256:f582f8282d1342d70857c72fead16d23925e3a7c61bb8c2005044c78eaacd271 +size 76266 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index f1e705f63b17..c8436adcba2d 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:916433ce57d132e0f7ef653785ada8e3c7500757ce0ac07e01afaf11d5644466 -size 52245 +oid sha256:894b5465fd3cbdeda1d82c2283a262a3a3c6de22e860737938627ef93efb14d4 +size 66185 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 8c5503f337cd..7482624bdbb9 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:c3858b3342133dd75188bf6da1be186b0c6c7bdc41b1aae9d0e67b94d245ec50 -size 49370 +oid sha256:840ebc83520a557c16ecec861f88304cb10f7136f2d00c2188b2379c7849099a +size 62159 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index e804a9fd8180..66343d4346d4 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:f40543e28a859f5584c79bf4325f90ddd9cc353e33b1567c956eb182cd6ae5fe -size 67490 +oid sha256:adc2a197747f0969aa9a5f645cbaa3d16f183b640377089a1bdb7886f509dbf7 +size 84671 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 47172fef1d1c..7b02536de77f 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:ca9aa1db805b2b0462e05469117e635859c9653ecc7492716f2db9d0c7e4fc57 -size 64347 +oid sha256:5761326c612daac435f5897d24aef2f97e6baf95d5deee9be6901c81091d902c +size 80653 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 97dfda31fadc..a5f1f4c345ac 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:bad086eaa06bedf208d09b0e1a88aebcfe4a66f4239123de7b79a06ed126d4a9 -size 55630 +oid sha256:a8e3ddd56b6426e2bdd6c9deb5a5a32d9b15784296ea6a224e289f16b45795e8 +size 70276 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 9febeab5cbf3..dd3821700788 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x16x256u2_s3_et128x16_m128x16x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:6ee598bf3dfcee7e19cad90ee6fd801c89e927b0d54f4ea96646c818853f2963 -size 52618 +oid sha256:e315823a32d4d61f572dee97b723a50503be1b7acf0813da7c81f8a68cf41add +size 66165 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index fd86f2f1cb4a..edd51b9d739e 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:5cc3f14754eca4eef3a1c0aa14fff0d1427921019279dc1078d7785c9ca58ecb -size 57475 +oid sha256:790411a6904ee255cd0476771a4283bc384606b2fc5835a51194307e8a4efe3e +size 72275 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index c520484fbef9..2bbf86b0fa03 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:87a22518852ae0cabb9cb17d2380cf78fcd7ebf120402dc39b20dd56640d82c5 -size 46171 +oid sha256:d338e6fff09bcf2a94a73b6191a6610c52040f2fbf830b14e37a5c98a95d5621 +size 57441 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128u2_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128u2_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 5ba49acb113a..6110dbbc0bea 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128u2_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128u2_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:841c875191fb197f4ee86b0a852f4468da53ef71a48b3223fcafd7ce608d2298 -size 59624 +oid sha256:6409583ac9694ddf10db0b091d2c2d753e5a87107968107564e6c942d2c9a44b +size 74991 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128u2_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128u2_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 41438e03ea8a..c715a23e452f 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128u2_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x128u2_s5_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:79318e908298355b00b30e935f8a4ae4007ce51bd36d8f530e2eb5db1e6fe896 -size 48056 +oid sha256:19a5ebbfa886e7a48f17bca1203ca1e11835bd038794c6d45caa9119d6d0a16e +size 60593 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 01b90851c83d..0e432273ca05 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:7d13cd0b4cd9c66adee966b945199ea368313ac733409602b31ebd1b82084114 -size 68086 +oid sha256:3b106d417bd25e719a956086b6e46aff5fa2d435aea9cdfe14640991fc4376a0 +size 85958 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 059ac2f56527..81a54f5573a3 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:a73162e76371df11043bdc101ae8aab1a6830fea0bc2b24a8c08375a9fed6cb8 -size 61511 +oid sha256:ea5297e6cc91c38c340c7360cafac17432cb91eba1f1cb7fdee873df486a753e +size 76991 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 52743489684c..cef5ebbf95b5 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:95f1de5cfad240025228e97b2b4f353faec31bbf6b0786c746051a1d24bec233 -size 56258 +oid sha256:4c679fb5d1f3c4129a8921f1357e0b5e373e9c7c228b7c3c85c7c7d516c52efb +size 71141 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index d9c2ce7716cf..45901b285686 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:2e092e73e83172890fa5b270ad265d1e838af3955c861c99e0d101e17fe01740 -size 49847 +oid sha256:575c7d44c62bef13535651882734ba632d07e7c7dc8cd7cb936f30ddec426e8b +size 62906 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 26198fa33ba5..bf1c6e6f4bdd 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:d3671ae4dbf1802f129f4a50b07cb67eb4de2076b1864612b678c4e5c1b9297d -size 71158 +oid sha256:afeabc25b762d04d1af8188405301c35674fde9af7bdfd8726920e2f8e6cf339 +size 89697 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 18194555438f..6b4ebb36c6e1 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:24701f8eceb93385456be0ad263d05976aface893ca78c2fb836661d6f9d608e -size 64498 +oid sha256:05033f48e4507e97235042f7fd2cd913ed7645c43fa480089c851b22dcdbfc1d +size 81271 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 1c215f0923b7..45dc897d19eb 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:b35b79a7b03e5056a7989c59fccef519a1f361a1bc466fc009483cb705fd3e89 -size 59204 +oid sha256:69f9d223c796422809f805c0771b70a167abdc0f5d14ef408349de75f0cce553 +size 74756 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index cec4e375b55b..63850c43a408 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x32x256u2_s3_et128x32_m128x32x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:536365af3a34fdb1f6dcac50e7d94710a7b3cd2d4ea82d7d10989cbafde8d48e -size 53002 +oid sha256:43c7bc0a6ac671c494da24433fab699be5aa26d8460368eb85290d5140dbcd38 +size 67007 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 9d443aa0b1f6..a5dc61bffdae 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:366baeff1e3e7eb6225bd080b378798de2abdd0fc899a35a0438fa119972cd46 -size 64932 +oid sha256:259d3e963045e15b9f4628f6115a57199ce3aa98cd1a53b06f52fc8214c9a271 +size 81596 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 4014b424564c..e48b98c9f56b 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:1d8f01fe33604fcb615bfa6ff894043f371b2da375a16462be2bf3afdedbb30c -size 52615 +oid sha256:aa84a80c53f56d51d9bb58d079b4065fa11db58b3b39aa063dbe52e1008da256 +size 67069 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128u2_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128u2_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 2596a875d246..8f99b4152687 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128u2_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128u2_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:bb0987acf9b3953899e7b80999de1f961f017caa0c8f6e5c1c55db3e234be3fe -size 67176 +oid sha256:3c29b4e9d12b9aebca071fb405d4230a61507a8bbe9b6cba695ba0009e416189 +size 84335 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128u2_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128u2_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 02411272b8b4..b1a524e3a533 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128u2_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x128u2_s5_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:739c2f700d47fe541405dbdb710354fb2ef05530cc2af43b76fc97d5522d0294 -size 54813 +oid sha256:f5fbe2de90dfd39a76139ebb0a7d5dd0655929eee6672f794d5b76f6fe9cfdad +size 70032 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 93911c38be69..18ec7eb701ef 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:757c5778f5c49f887ef1b7237af873afe50665c742ca1fb18b4aae9380859e4e -size 76291 +oid sha256:6acd2bd4ead7bba8e8ef770b2324671e3f9b3da1324d7cd8c16d99b12ec89ea6 +size 98206 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 531a63226e7f..302a040e4bd5 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:088ebe0ea42cae11d53fcec46fdd463c908460e18ccb861e917a18b09d337b7f -size 62467 +oid sha256:4ad014b8a7e7fcc250e11e65cf5006de0ae2f9d6e66f633a1349e814afa9af92 +size 78336 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index a55b74a6186b..cc0056469441 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:1fd8e476f82e53a9232446704c4504abbf21e0612640d8312bb32cf6e43e6a18 -size 62765 +oid sha256:7207c1daa9255667ff26989cdd710a7fa599220325d78be5740466e8d7f08c00 +size 80123 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 393171853f34..3caf298b76f2 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:eb41fd7b63244b0f627f18e1dc7aa4a74560e7fd75544d95ba6031929dae4d0b -size 51037 +oid sha256:de47be2ddf96597d3c0924de40362f27fc9de4e087e9a942fa6f6835e3a1fb25 +size 64351 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 84e6a5f29626..5e8a30a61423 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:7c10b77c674519b1dd630db49165b62b14cadf80598809e222df8fa1b5b409e4 -size 79317 +oid sha256:e3a06bf672205762afb0c88dff7255d081eeb3b4d147361c732ca70f286b99e4 +size 102307 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index b97810dfced5..66676c92b44c 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:8bb7e4c6bfd8cac525d31a4a70d61d745a5f719985bebfd6061c71ae4728ba4b -size 65691 +oid sha256:e4b921c4c5a27bf2b3b643429e79cb8fa2fdd5b7c2be8e940b67bcce4469aaff +size 82752 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 2ac51ad562f6..2772a20f9227 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:974c3e05907b4912912efa7ad2f4d95650c8bd82b58f5f88b6002ca2a15bf95e -size 65658 +oid sha256:2b5e59ddf983c65e00a73299e349917745794ff7abbb526d8536e1c9ae6d3674 +size 84157 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 1f1cb2a7272c..e742c4d331da 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x64x256u2_s3_et128x64_m128x64x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:a8c2cf222a19b77fc16fc7d7414a6808b5ee6cd4618d996f8f6bb764a8f1b622 -size 54271 +oid sha256:5ab7870b1ce6c8c1e606c4a924cc0e6633afb5cdf7c51ab7660f8a3697780b51 +size 68219 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 5f29144c199a..86c42c2acef4 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:fa0bdd1a9d8d280e8359d4842d0d43e233fcf0ede601d636edd4e580d4801f8a -size 51672 +oid sha256:4d675dc79ae8907642d2a98cd41e7b4f5febc31bded892b4b71bede1301003ae +size 64415 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index f9b2cb0e69f0..6cd178c1837d 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:a50395a433fe4066e76565dca8c504edaed6b51d1ed66f473d5818a105ebbff6 -size 40410 +oid sha256:48dcdbad3131aeb05e68b10f69a24a8b8e8b4cd1baa78564a262d6084b14ca8f +size 49871 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128u2_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128u2_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index b5cab297b1ff..8fc52e7fff49 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128u2_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128u2_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:edfa6cde89f4aeb0fc14ee8378dabc7b3f2be4b8e84f2ce070081af9b9acd937 -size 53893 +oid sha256:1829a81b57a65ada67d9284e7f6b9e4636b7bf9afda61d0823d65cd27cae3cea +size 67557 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128u2_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128u2_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index b05560a20ebd..62e96f0438b8 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128u2_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x128u2_s5_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:3c479c49208a0980d1cc462482d09e405f65106149cbea21feefce9e99bb2a55 -size 42692 +oid sha256:755e841b0183bde9e6a1a311addad16ec62f93b11df7d800f5a860b34f0f43f5 +size 52869 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index a40ee3a1bda4..f4efdfcc5356 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:5b382905f1baa207fbfcff500b383dfd6ac3cf811ea900848fb59ada123149d5 -size 62689 +oid sha256:b8afa3cfd9cb26653e1603289ef549bc1744f51322972895c5ec93c3aab2ae8f +size 78567 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 2d88618d67e7..10a422d909e9 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:fe059870526410c984cf08cfba4bb846c0e56b7478d11f387054c19c606f9e90 -size 60733 +oid sha256:a5116c1bc5db3a918b2fc7ff5e043769d299d7d95c1e534d2ff0b827aa5aa7e8 +size 76146 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 020182010547..602bd7fd5bdf 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:b722eb732a728574bf21eb09900c36dada893eb68195bc1be88c5d48bc2c132e -size 50170 +oid sha256:a29072eee16668f357589ba29ecc1c57ce1fa8dcba9aba8c7365786b9855781f +size 63580 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index b0417b1f7b80..9a64f4c7b84b 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:38eb34b7b268793cf74502378fbfa0ecc227975fe9761362fe2650ca8d976f6c -size 49024 +oid sha256:edbe1a1593d4b7ee912930e7d883ccd19308483597b238e8c47d88ab3ddc4645 +size 61500 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index b8c187d8da1d..91de0526b102 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:567a414db47360bbc223930aaf38c76d75084b62af1fa7ac6e206b6a89375ec2 -size 65843 +oid sha256:0550cd15b012ff518b9de8a14b7ba468eaf6efa1429fb58f3f4d3f9a03b2adc4 +size 82133 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 01e39a4cd682..55465693da8c 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:d99970efb80872837c531752d64c0feb8f598407f0089a057a780e8696fc40ee -size 63866 +oid sha256:1ed99ca65af4eb27c6d09b63eb1c14d97e76e0eb1a26fd20a698f15e3b259e5d +size 79992 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index eca0735a1f6b..276fe31b3f3b 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:623c25d8aa719905fb6da10fcb14ca2350dd8680c5b070db75aed4a9db4d2e65 -size 53586 +oid sha256:b83b25e11884bdb79dc83b2ae8c22969c447a9912881177f4f8fa79cd7490dd1 +size 67836 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index bb3568656839..5024e203da92 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1Bfloat16_castBfloat16_Fp32_Ab32_t128x8x256u2_s3_et128x8_m128x8x16_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:35430cad4f2c6b27609342c4399da8538f90003cb45b410513b8642d98ca8489 -size 52079 +oid sha256:d50f8aece62b0d4881a5a97aa8c610615652656f4dd21f506e41445c98a770ab +size 65716 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index be94db5a7a04..7e09723e451b 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:395060a72aec3cfa1989d722635b7313b6765c3abebf13e56a4577e815ac7a87 -size 40831 +oid sha256:5a2098d00ad0ef2f4475791267b20b327f8d97d95d0eb8a4967e25b4ec45171a +size 50025 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index febe187c4a9b..ea9e6529e025 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:9e668c626b7e015f2265480377d2d01f318bcd750ced64641c612158d4de0369 -size 28476 +oid sha256:e420279a8fcc30a7a0fa2f44668bcfde09d1bc0a9f9c9bb5091f301d6448e382 +size 34152 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index bb734f2bbbda..fc8fd75d8988 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:39b6bde83584102c7a03df11ab4c96149157b2746e1e0669e2bf62521f355ab2 -size 44700 +oid sha256:f8300efe205ddfbefd825535663eea98608b44f2e4bd8372982f6b227898e902 +size 54973 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index a3988706e522..b5cd3f84f07d 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:7853d2c2b19c5e151dc9b717e9c3d04fb06934b1831480a27f8b6f3d65e2434a -size 31655 +oid sha256:c08d596000c7764980e55172d33cbdc969ebe07352b53082bd75f84747b6bb14 +size 38184 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index c1b9d2c92e2b..c3c9c555fe29 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:891b5b10157e6bfc162e1f48277fc7d0eeca195f4083f692c5692bc018f45fb3 -size 42241 +oid sha256:5341a173dcd861df867da05ea89d59d772942911fa539c89239124ab7e22a5f8 +size 51697 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 6c7e8befc186..26d4e3485e82 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:408ac96f306065260cf7eab845578063952277d83d2bd91764c02ee6ea2dbc50 -size 29728 +oid sha256:704d598bfa698a4bd33311b9548aa1ee563e24b2d0d34547d074c0742afade8e +size 35809 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 865e11a381b7..ae0938ca5d52 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:661a1f27a7a7900abdd7d6d013a8f06306c478f80fa6e0947e1ff58987aa5ff9 -size 46339 +oid sha256:45b530a96e168e0c19550328babd59f529015fe765c3e3f6c7b8437163e1f831 +size 56898 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 848b53a8fd26..8d3751a32ca1 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:bd8f289fea162fb617f6db9ac6813498167d55f2db146c33d27dc6c6d7821a5c -size 32916 +oid sha256:157407a7dd4d5ab14f7ea84ce955a096bdad6da9b517b0870d3785e4df3d771f +size 40100 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 9ec475d0387c..5446eca370fa 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:3b37cfbfdc4f822a3eeac49d473a7377627e506e838b81c94f74e5a3a6201ce5 -size 43997 +oid sha256:e349229bf5ea04a0e2d3376b1c19efad190514bdccda74d37e15c690af67225f +size 54426 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 16de5298030e..a9bb9a838c72 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:59de805ac0cb7a83d61d6251251a494b91fddfe6eb5ce90af5df11d036199de1 -size 31697 +oid sha256:b0d0d06dbe9524a63369fa120f9b3515f360524eda7341420370a93f1fe534be +size 38580 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index ba5e5cf64bac..064853b463c1 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:b5fd494674dda44cda6d87db830c7278153df51a91f060972723b9f063530d6a -size 47790 +oid sha256:a43cfc02b3df61be2c0612515754ed132b425be49103fc69a99f433878742385 +size 59075 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index d9ddf65c2d87..8a17beb7e944 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:c69cd7c13ec62f41695c4e7d5726b0ec46d05e8dd11d2df3cb9c7749abc4dde2 -size 35150 +oid sha256:d671a1146d937a8d2d691367c795f8ce91e25aa625e75b9d54bf44efc000bd3d +size 42900 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 5bb123c19498..7627173c09b7 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:241b044a5c6e211597898c7092d83ea34e4ec58640ff264531aa612197f8b211 -size 40301 +oid sha256:36e74f0615ea4fe1691f172a69c2d49ffc202225a8acbc7eac471da0fc2a5197 +size 49080 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index a1e364661f54..6dad62d07111 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:fc42034e4cdf3d54353ebcfd76958cb76bb1836a01083b6e0d011161c4fdbcf0 -size 27790 +oid sha256:875222ca7304da4e8b3aa2ace21b899a4938a05b67e84f56238406615e2998fa +size 33287 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index da65a735d6e9..ab986bd11e52 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:fa4b6877cc529a1f12844ab90d5dbd248d8581dffbdde71aacffe94593acf307 -size 44238 +oid sha256:5c7c6d2cce09c875420a40fc9e8074c613575bbfe96ea79c536113d0f4ed0c90 +size 54256 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index edaac9829dfc..336c46ddef9b 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:976d8b120b24b4dda63be9c84bc2138e55385445ddc80dc8a9b0384d669fabbf -size 31147 +oid sha256:7ff5dbc626351b84975dc5ae9545e429df27cdec1044778ce25397f963aa6d36 +size 37517 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index d5a2904672b3..2982762ebc75 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:a5f912eed04128f2bc48d3ac8e8767ad15adf8b92ef129bbb2f25df36d274be1 -size 42761 +oid sha256:423d959aac5db74aa18c5c54a96c1821bc5a9c9f086cd71173fc82e7df9ffca8 +size 52495 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 70b572e91a97..eaf53555d340 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:390b5bb219a3b588075fcacc43480e0310f50b311f8d1a522c8741431e5ff185 -size 30234 +oid sha256:a3d958a1108cb3d43121f990e48bd156179c7609ba6281470e516e91f076eaf9 +size 36487 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 9b5a48a3b00f..b74040932c00 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:95a6cd5bf8d1be460a5217713a45e0a0505a70f712b78d53d004763c439f8841 -size 47523 +oid sha256:2f0a187921b2d1728755dceb903b67d965ae619b50fc299cace5654cd5ea3963 +size 58552 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 14a707a1e02b..3946bf20dd51 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:1c19744d2c611fee02e37d3c73b5a547a5d64818de8cdbab36162be7894b0daa -size 34647 +oid sha256:9c57f08fe4ae07886048dc10a7ae8c9fb0f4c133eee28f710943763d76832110 +size 42318 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x128_s7_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x128_s7_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 9b3dc42693be..2957ed8e3a17 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x128_s7_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x128_s7_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:fea813cad69b55c95fc19839311ce92140b3bd69cb85b0ae95f5f6cead927c70 -size 50894 +oid sha256:69673dd3ffbacf03cbd6ab90b84acbcb30005ce6c8be60bed20a427858f41655 +size 63340 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x128u2_s7_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x128u2_s7_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index bf36fd52466d..1a75b840d5e8 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x128u2_s7_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x128u2_s7_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:b37fe33f49b8cfc743e4e7a198f4605012be7498ba543208dc43d39a12d7c14c -size 53834 +oid sha256:1ed72347c96942819a10f07ed1401fbd11eb1a3712c518cc40036d29cc5c687e +size 66936 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x256_s4_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x256_s4_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 680d224c636e..4339a17a6cab 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x256_s4_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x256_s4_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:d61bfd817a9abe6976970a7f092095aa54e85f35b3e0cfe6303fd45f9b686d51 -size 52187 +oid sha256:6aaaebef4c118c37f77f4cfde880759e8403ee5c598cd42cedddaf488ac6c551 +size 64789 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x256u2_s4_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x256u2_s4_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index c627988c38ec..e94160c8c85c 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x256u2_s4_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x128x256u2_s4_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:79ffea016500d152e885d8e2f9b993a61f9f561efdeaee48a22e20a99d704bfc -size 55428 +oid sha256:6a302bfbc1cdfeb17798b52261048912d052fc8e308e9e0609b718bff98850a5 +size 69070 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index a20ef845c532..63a382839b08 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:ba8dd7c8e3344b3390defcf699d5d92f46011cebbd720fc5dc86417d794bb045 -size 44835 +oid sha256:6c92aa38df9b863b42fc798df1d71b634b88b09b60075a598ae2b8b2b91d112b +size 54524 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index a6a0e1eca075..39898ce4fbee 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:3d8bba8a49e349ba6b359678a43c8ed08cf925c30c44bce12cca2b9bcc291333 -size 31087 +oid sha256:b42ed20796b94d244a951c9d28457506abc0aed2937eaca7829a7d255ac06aa4 +size 37878 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index e7beaa710445..08129778279b 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:baf08ec7ae6eedeb496899548406cc24cbc53b926de918f03f028e99a8bbc905 -size 50478 +oid sha256:2fba14b1f84bf8268382a72a6c1ed71cf9da34753c332e468078d0c6acf0bb64 +size 62161 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index b8b7a4338009..26f824bcc45f 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:c0ca2dcda8bea968f0cccf95ef49bc408e7bc408510ef7dd29fff65b21a0dcce -size 36769 +oid sha256:f5e7fdc6af2060fb44d758f0c06fd5a4f7457ee158805d194b2e607187b96dae +size 44729 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index fc56b6f3459f..102d6e47cc46 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:aed32a6a90a2300d7f32ab532ec5429381c7e35209c51a5a7b42c9a073bee9f8 -size 48403 +oid sha256:a00fbf7cddab74cecf70486ef9e5abf153e82b0358a75d2f08dd0652575ba720 +size 60034 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index daf71e9c746d..9d7f018fd8b3 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:355a352eddc3d403b5e4c003b0470859a687d18edc98aab8c0d5ddba42e35a62 -size 34829 +oid sha256:717393e8798b0e801c52320f793dd863a1e9bacaeb0682029710061e0b2d0f1d +size 42818 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 4ee30e5860d4..7ecc131648b3 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:3c5d096ca7f5a1fe36525abb817aa5c6cee8a2184c74e6a4eb2703731ae48682 -size 53957 +oid sha256:5dced4d2ac63fd9976f992327052b36b721787de0f87d8b0bed0466c2e285666 +size 66937 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index fcbfe9b9e33a..e8778961aecc 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:2af92cbb6f875ac2a4363beb0dbd91f9dcad0f6d54808052436abc3080b600d6 -size 40649 +oid sha256:289bb8a9c9bea52015433e52a7d132e073b191415bb78413e1210ed53d25467a +size 50018 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x256x128_s5x5x5x5x1x1_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x256x128_s5x5x5x5x1x1_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 513969743676..b759146a09e2 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x256x128_s5x5x5x5x1x1_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x256x128_s5x5x5x5x1x1_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:2abe18d435f44fb0e3c0fe6a3b519a6193cc79d3662813629546175ef31cf24e -size 51314 +oid sha256:7de7cf38ac3b97dff31b9e0206d7917a3280e7f71088d76e717f1bd5464a03b3 +size 64366 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 35ebd5007149..911f047e3f72 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:0515435c041b7c63600963c470f7cefc3ba4e303b44045ad3cb5577c91354ca0 -size 45410 +oid sha256:427fad793524dda7501759641eb84cbd2e279e4eb9f5ed8da6265ea827e8710c +size 55473 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 5ddc8dae09a4..2d9cca044f8e 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:0a18c69a26ce13b825c8a30e27b92641819b93e7e05d8a1321576a1e0f037dda -size 31758 +oid sha256:06003cfc61b63978912da3d85274b99c6d6f692b29a53d40961f379ba3c4531f +size 38719 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 694ae1be166a..ff42e4e2f311 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:87e3fdc154d2d4acb002aa388d605ab5e3ef43605a7d629eb8608a8a4bf86d11 -size 51384 +oid sha256:71b671c49cfd7eb0601f256bb239834016c925c65dfa7225ee502e6d0ee1fe16 +size 63195 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 32513140f78a..2f5636a0e451 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:58f81ce21d8c0718675128e79a93248afda4b05353487430e5ab9717db53d8a9 -size 37277 +oid sha256:e0409098a75c19c5bec729d24b69b86349cfac0e026d64712e15240c1db824af +size 45407 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 383be14debf6..981df80a27fa 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:e2e330f6f8d858f31aef7889552740d30251454ace735dd4ea882cb7d1e2035d -size 49162 +oid sha256:78c05a19a1ccab8689f6a27697d2c384e668867fc42e4e786a951cf87de705a1 +size 60962 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index f16c1837c469..9be7de2a1794 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:7002dea65df860f0f8240fe1a0b19c23886c499b0f8e9e21432311f2eabaaf87 -size 35473 +oid sha256:0fec7aaaa9691422966f09d659e27528e3b8bb93672ab0c14e2c840b2404ceca +size 43675 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 51e83d3c7fde..dddc3a6833cf 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:d7d3fc2d9146cae73fda67628892d145bce36e61d3a3f73774cf549975016ea2 -size 54718 +oid sha256:f9e320fe4b81d7dcc923eda88108c5dba026f68445b3eb0516c716b1ad70fa16 +size 67742 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 8b453bb7ea28..ffd3a5f44aac 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:ff0bc5b015ae62faf3ebe39a34e4e34b3e03cd6792af34bc6d9300434eb6623c -size 41252 +oid sha256:c780efdd905e4da4e92ed47738a3760ad39796642e5ef439cadae7731211a3c7 +size 50887 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x128_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x128_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 688ee13428da..893e6eadd0e3 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x128_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x128_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:5b7a836a90d49127e7c4a3d97c2080670e3fe5c562b9acf0c3df6acc6b16ef14 -size 50576 +oid sha256:298d6df2c9523917dd8d29512c13447e362c0d2f5c8649daca161241a1735d78 +size 62576 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x128u2_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x128u2_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index b55a7ebcde91..ac615101dc00 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x128u2_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x128u2_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:a6b373d497f37a216d17fe337672505e0c3a9365df751945420039a939a66fd3 -size 54171 +oid sha256:576d397be02d4d0bc4b11cdb65666ecc3f08c7d7a0e5dcefe53a5870ff5b7f65 +size 66991 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index a57e517a80b8..c39b84bdfbda 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:c1b4ac607f89fc86cfdc72777a9d4b51bb5c95cd8119fd013837449fdfb2bb7d -size 51809 +oid sha256:c43882f97646be0806eb04153fb5595d94859d89e48a0e2f5ce5a810a10e2db1 +size 64299 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index b3d4e28fc0b2..4c25b0506ad4 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:cc9345f138f9c875a3d88fece86bf2db7a489148d755980324becf993557ecad -size 55976 +oid sha256:6ba7860e1b563b3787a7c912fd9df4164f6bd3eb027acb84e1f1b07d8b9f78db +size 69439 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 41fd5d3308d7..1c870465063d 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:3e0a7b785855018274f08fc37edc955e49a1eedff60121071f0978a2495691e6 -size 44530 +oid sha256:6e777e11e3801673a69fea9a2b00811ff8f975bd57a251b1e40f57082a7e2b61 +size 54420 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 6b9135332ada..365c66966c13 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:f1e5c8535970f88a20971fb9d88e264d12eca80847d5cbd776f66cdef780cc66 -size 30658 +oid sha256:3da1abc83dd99bce7ba82377f896a561dc84054a5ff2f0b5035ba09d8d1c3d1c +size 37132 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 0a96ee6a23bb..5672868b0563 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:e8329ec289b75188985b338ba0d585d5f68df54006fc79ca39d54bcf7970a447 -size 55134 +oid sha256:532e0870b8c73240de38439498950fb1103a0e90a23b04a619114b5c3d00672e +size 67981 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 49841f8969bc..f464d2927d39 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:de0a15f244e496247ae728ae0d9d9dcffa390c671400c938f8a98b47735c23fb -size 38477 +oid sha256:cc5e37ea9eb1c630a0a4a0b1d50440f49e97f54979e2564c140ecf4540c201f6 +size 47076 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index bc4a211066b0..9d5729f6ecb5 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:b443e172491c31964f44af0dd482793d246bf0f5b411d3b8db547ccc752b884e -size 48153 +oid sha256:ca9028ecad38be9907d76d8f54161727039657552a2cf7bd10b5d2ebb0321232 +size 59601 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index ecd78198198a..16390470074b 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:17d3658e600b8dda123debd3040032777af9adb372a8ae05018b41aba6fd10ce -size 34400 +oid sha256:bd8116c34fa427fd6c5cbb681f6bb8449fccda118a3f4e2a8b06c033fbb76040 +size 42255 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index a788af63762f..7de32222e116 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:6dcf05081369138c6dcccb226e9353682c5013107801286b312a1ed559f40d5c -size 58882 +oid sha256:2dda83def2e3c5e9072dc8991d598c45c0069c46418b866c42aced2ea1f58527 +size 73989 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 3748542c0309..2dabe93f0756 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:cf5ea24ed9492b7c575d94d89af5cace1084a0efeea8bba3f9f0782c53e4f72d -size 42432 +oid sha256:a4b48c57f7282e8ee7d54acdeefe8b7a28c8aef00134f0c28bfa87c963e0af6b +size 52319 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index a4186237ab2d..9447e26e389f 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:f3c975dc4edb6c7228a7806ce57dbcb4347f63636dcacfa09e5583169274e592 -size 46723 +oid sha256:96baa1bc3aaac260b3291b46b0a822e5694e7be0f13cf93c0d1aea5540e12e08 +size 57556 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 6271692106f8..c9ec89335149 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:8fa10f690ba45b963fee95289bfbfa3c0ddd491566d6b6876dc6f55967ec267d -size 32763 +oid sha256:867e717f85173618fc27c667fed954f342983b1082836feccb1b47b0f9ef6740 +size 39990 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 6e8363969d28..070c3c642d5a 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:42b67c6c4176c67c1d7fe5648726cb5ea48fc19bf1416c6e1023ccfdf3fb376c -size 57201 +oid sha256:6a7d3303d3751c096b73b765ddd08e3ee513b71b532dde2969155d77a9b8d3b2 +size 71136 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 0429028c74e2..ddb911b3e22e 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:882f2f74ecb5594938808e52b9732a5c2cae11189211b065b5c198ac4e7d9aa8 -size 40489 +oid sha256:55bf25ed8f19334afa7c439d1230b1db2e9ada5638ccc8d08a0e8a2d82b58145 +size 49872 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 747e0ddc9178..9c794330cbe9 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:63b5c148a0dd4a0c844054ffe3b7cf79e7433c5aa4d33888fca66488599c4122 -size 51002 +oid sha256:4f67cb512ebd3fe965c92da2decd00dde155c8e2c6e1ce80a1fee3f054dc1dd8 +size 63834 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 71812e7bb1e7..7e162f243159 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:96fd011c5f3dd099bb57fa5f20ccc84517ee6670a17e4de1aa8adee30ef3dc1a -size 37919 +oid sha256:fb7116daf2ad667d33f453991352055f86994ec7b4e6b82cc1767429b576a1db +size 46913 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 6509c0a93bb7..53c426d662e7 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:5b2286efcd357caf75c6a4bda34578b0e398dcfc04faa10b9d07e5e0d8709bdc -size 61624 +oid sha256:ac7e4f73a8d44f5ebcd725a90e04884f8de21c8003852a69ab659b75a5b897da +size 78048 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 2ef45ed95b34..e1bfad269dcf 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Bfloat16_MxE2m1MxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:35778eae9f6f0483a3c5b1118fa07a874b9add349573cfa086bf25b2a5de4e5c -size 45676 +oid sha256:1e6d53c28dd197ccebfb6f3a69fd15498d500821194555b4edc4156a9be2b5d0 +size 56476 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst index 281e21910005..8210c4e675f3 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:9c1e84348c6f9ee2a28963f6d8594ce2b39a10db8c56e05a6a4ab0ac16b12bdc -size 111471 +oid sha256:9649366b4974ae2b92be36a974a247baa646fcf6b3e7eda65204fc9805af5ba6 +size 145444 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst new file mode 100644 index 000000000000..e71f1112b1ea --- /dev/null +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1379ec3ad686f8dcf016b621c77c8333c961aab5b693b38ec2ce103d7a2b53dd +size 152447 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index bc0276e2744f..67f03a0a6fcf 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:96f5315fa91d6c08d2b3d6ad090b05152ee2862943c8da02e221dcf61585fbf4 -size 108826 +oid sha256:91eaf3b15dff6571b7d960c97c943a8554a8ceaad85aa4a45c864e155e52725f +size 141090 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 7d394200cb6e..d7ccc6c3d925 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:ce2dce278fec28b7886afe7700d7ff1120d2ea07b31ecd4e577fb70ab570d653 -size 95501 +oid sha256:799ca1916487a497d982bcb9d7c0a5c769e4e5a9164f8c6e6ee15fcaa24a911b +size 122247 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 562111c4bc69..b4ff74936fc5 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:8207d2390ecf8c8e7ea4ab88ccf11ad8acc092391905610b95dd2991836529e4 -size 103640 +oid sha256:2e4f0ca8ef81ba94fbfcf71711020d1065a2cefc698f1cc19f20fc627bbb7075 +size 132260 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_geGlu_lsfaW2_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_geGlu_lsfaW2_dynB_sm100f.cubin.tar.zst index 7d0f6b861331..57397238ea6b 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_geGlu_lsfaW2_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_geGlu_lsfaW2_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:1dd906900f0ad6adad467374259e3eca29b36731357a031aca9c07c10c4a059a -size 96426 +oid sha256:82a78944e28ce5d8afa9300cc8ce4da2ed075f6d268e805559daef24ddda1e0d +size 124436 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_geGlu_lsfaW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_geGlu_lsfaW4_dynB_sm100f.cubin.tar.zst index 60272485712d..5e9e55543e1d 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_geGlu_lsfaW4_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_geGlu_lsfaW4_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:dba02473557c925285ab967a706e41eea95908fd4a9375d1f04e35a07ae3d49c -size 92399 +oid sha256:0f1b2d7009962cadcb13006b9e149b60833e7e9dfdea7e320f3cc91c72e4b7e8 +size 119560 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_siTuGlu_lsfaW2_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_siTuGlu_lsfaW2_dynB_sm100f.cubin.tar.zst new file mode 100644 index 000000000000..e0097a0b6f96 --- /dev/null +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_siTuGlu_lsfaW2_dynB_sm100f.cubin.tar.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:17ea41de02dd7257d5836edc5d2406df1c869ff5771b01092c2b4bfbc97477cc +size 129738 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_siTuGlu_lsfaW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_siTuGlu_lsfaW4_dynB_sm100f.cubin.tar.zst new file mode 100644 index 000000000000..15d9382dc824 --- /dev/null +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_siTuGlu_lsfaW4_dynB_sm100f.cubin.tar.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:413bbd3f10097934845ad08d7104513bf301f69a59f56f8f4f36255fb3e939d7 +size 123872 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_lsfaW2_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_lsfaW2_dynB_sm100f.cubin.tar.zst index 6d15adee28df..4f7954f58c22 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_lsfaW2_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_lsfaW2_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:363c654c45396f43919c8a24b19e131ebfce0b8016468a67a912dd17d4ffc051 -size 93543 +oid sha256:c41f89d1e833f5ba9bfeb6e8a02ec7af52e73b1ae17e2d6293fda1bf442accd3 +size 121658 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_lsfaW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_lsfaW4_dynB_sm100f.cubin.tar.zst index 40f5d9f5006c..0cbadb5bb60d 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_lsfaW4_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_lsfaW4_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:83afcc51e94d8c0c8ce6a9ea27d75fe91fadc627733b8ee8ba73fa75e3345467 -size 89160 +oid sha256:685170aa60cd716803a4d9184db0df5fbd906f372405d39bb8ffe9c191e32dee +size 115884 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_relu2_bM_tma_ldgstsSf_rgTma_clmp_lsfaW2_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_relu2_bM_tma_ldgstsSf_rgTma_clmp_lsfaW2_dynB_sm100f.cubin.tar.zst index cf4ca0143d84..7f32e66d6613 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_relu2_bM_tma_ldgstsSf_rgTma_clmp_lsfaW2_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_relu2_bM_tma_ldgstsSf_rgTma_clmp_lsfaW2_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:b56d1c54ec2a4848f11e759b6607bd78a1e583b68b68f3770e2faf763a143ab6 -size 83333 +oid sha256:22a95f7a30a253f43acb864779c811ebc653a60123e6d867dd4b3de38bf086cf +size 105144 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_relu2_bM_tma_ldgstsSf_rgTma_clmp_lsfaW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_relu2_bM_tma_ldgstsSf_rgTma_clmp_lsfaW4_dynB_sm100f.cubin.tar.zst index 1a79b89bdbff..fc88724867a5 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_relu2_bM_tma_ldgstsSf_rgTma_clmp_lsfaW4_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_relu2_bM_tma_ldgstsSf_rgTma_clmp_lsfaW4_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:9251c860109d2fd0b49c3fcb3e5f41c8e1f9107b7ed3385355ca19a66d33c198 -size 78118 +oid sha256:7d304630aec42e065a787a8ceffbc841dc435f79697f76565bdddabcc7e9da8b +size 97759 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_silu_bM_tma_ldgstsSf_rgTma_clmp_lsfaW2_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_silu_bM_tma_ldgstsSf_rgTma_clmp_lsfaW2_dynB_sm100f.cubin.tar.zst index 95820b5c54c2..d877c4db40fb 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_silu_bM_tma_ldgstsSf_rgTma_clmp_lsfaW2_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_silu_bM_tma_ldgstsSf_rgTma_clmp_lsfaW2_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:7a8c27e3ec3068de70d6eec6eb02e6c8b90e517c2112a89fa066b54527dd880f -size 89182 +oid sha256:84c0bbfa8871bef051093dc8c0fdb24952e64cab6633995519dd64c1635722ff +size 113361 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_silu_bM_tma_ldgstsSf_rgTma_clmp_lsfaW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_silu_bM_tma_ldgstsSf_rgTma_clmp_lsfaW4_dynB_sm100f.cubin.tar.zst index 6509ad84c676..5fe557005dda 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_silu_bM_tma_ldgstsSf_rgTma_clmp_lsfaW4_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256_s6_et128x64_m256x128x64_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_silu_bM_tma_ldgstsSf_rgTma_clmp_lsfaW4_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:badb6c892aca1df4e71626be60a6679a08ae5eb4755edd214dfd1b9ba27ebd62 -size 84358 +oid sha256:180373f1f1d9d86a569ded781632c16b5cc22b385edcdec7ae30fabefcaa66df +size 106630 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst index 557ab95c9a16..e02711df831d 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:8468b994b07758e2909a9796c8dc158a5466e3420ccc566ba67afe51b15c1b87 -size 114323 +oid sha256:a2cbf5322b3ee4aec7d458a98e46f7bab58d244e663accee745e96c15745a382 +size 149466 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst new file mode 100644 index 000000000000..6dc07e5c642d --- /dev/null +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fa667590cb88a4118f1a4ec0d4bf1e4507659a30b24130b5525b8a46a7dc838d +size 156250 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 04dd9f8b95ac..42327d253be9 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:f24b7fa19d0cae51518a606768817cff8642a0003224b0e6713c2398e58fac64 -size 111652 +oid sha256:2858f948a9ae56a23b93546564a11d341d93bc1e6cec22ea29a4b69e538eb9af +size 145185 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index abc580313648..7fba60359c13 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:d5000187b05b14fef8d5058338d80c15b58d8cffc1baf43e5d662a0d576a0f1d -size 98625 +oid sha256:7156e179a26401e4f3932fc51af483b5ae73ddea199bc7db181361c8e8a61678 +size 126001 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 6b4a1a7f146b..f628c15dad1d 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x256u2_s6_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:a9242e418cd93970b4ef101cacc7e57151d411eccc7e85a6b8a92afadcfbe6de -size 106325 +oid sha256:8be456932c42879763a0db0365af7a59970b0f729921bfb1a750442fc237540b +size 136067 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_geGlu_lbW4_lsfbW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_geGlu_lbW4_lsfbW4_dynB_sm100f.cubin.tar.zst index 7b81eb96d4e2..0d76640ca49a 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_geGlu_lbW4_lsfbW4_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_geGlu_lbW4_lsfbW4_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:05aed38c5b67ce666317dcbbc5bb3bb88af3fab0611c373e3c8169c86f2f204a -size 93333 +oid sha256:63cabcafd75fa9ef74ce4a7c3cc4335d16d2f6aa648d47dd7605235550f9eaf9 +size 121646 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_siTuGlu_lbW4_lsfbW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_siTuGlu_lbW4_lsfbW4_dynB_sm100f.cubin.tar.zst new file mode 100644 index 000000000000..9fb5188a1cea --- /dev/null +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_siTuGlu_lbW4_lsfbW4_dynB_sm100f.cubin.tar.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da04fbe4b5062647ae49e7367a60e1a33107dcff56c52fae0ada5a493a7cb97b +size 125618 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_lsfbW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_lsfbW4_dynB_sm100f.cubin.tar.zst index efd1220efef3..a3b591a4e898 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_lsfbW4_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_lsfbW4_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:73ab339a2a3d70ec43aedc97a138395347949f08cf2b5f92ca666b00a1ebefc8 -size 89984 +oid sha256:5c431fa4c49efe0fc0175f84b45fc4b1beb03df1c8a97117dcf55a6c54b9f430 +size 117832 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_lsfbW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_lsfbW4_dynB_sm100f.cubin.tar.zst index 320858fb22ad..454912e86f3d 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_lsfbW4_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_lsfbW4_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:cc00cb22c2802dd064b5d9b1e710beb45b57273222ba06e569e67bf95097e896 -size 78040 +oid sha256:6f338965ad82eb821b0d3a89dbe84a8be1d47b952f9390564bd834d3afde4e91 +size 99477 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_lsfbW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_lsfbW4_dynB_sm100f.cubin.tar.zst index 98c5fa444b70..c3a6cec15443 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_lsfbW4_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x128x512_s3x3x3x3x1x3_et128x32_m256x128x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_lsfbW4_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:af54f4a4394270e8c1a3c7d868d39f2c8e2b40b5474c8cac87f36e284f63018d -size 85191 +oid sha256:72cdb3aca8086f379b0c39029f8cf88dba6d293edf8d9b794ac03ebd94bbe2a2 +size 109155 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst index 65c54e144719..5cf16d4e7d8a 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:2eed0e58c0bc5c28f72ae08134d0e5426c9ef5fba53fbc64580dd5d30487e62c -size 54359 +oid sha256:c8cc15ed443578965b3b1e2f91e5df75d93c3e7948c59f71bff780409186de58 +size 67450 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst new file mode 100644 index 000000000000..5c9b08e28f68 --- /dev/null +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:56bb0c28a7c8094226ea987f1ad8f733099b7d71d3d7e7b353b362981c9cf07b +size 68134 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 60d418a8d71a..ada4c25b5ab6 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:702d7797ae7034bef6ea9a226acf9aa61cdff8cef2aa5f816c42e15439ffac70 -size 53994 +oid sha256:0f815b3c65bb9232d3c8ee5432341894ce150bf31e37c740e169e3ff8b1804bd +size 66970 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 5157a5e4f351..b843ff375fa5 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:26e661acdccebabd6568386b8aa4de42fc92a2fee512f09a98b1f477ce776378 -size 51187 +oid sha256:c6ffc22be2082d8ea34ff1f5d2f86285a2b7f90c2f85b5a31e579bf839150194 +size 63277 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 7bc9dbc436f7..0aa1387c1188 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:77303cae94f2e41a57d6f4a95a9c36f7207d74b60ff37d815c1865bda5c1a09c -size 52215 +oid sha256:06bc2a6759558d9d3ad86903cf7246a96f28988e78f6f174286d68304a49facc +size 64671 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst index f73b2c47a9b7..048a143bf72c 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:91fea010f3770f97399b4bad9a35a3710b0a9bdc807cd4ddd688c88dc961c9d2 -size 41289 +oid sha256:817935cd0181c88ecdefe67ef241fb199fa2dcffd919895baeff3e637f30275d +size 50805 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst new file mode 100644 index 000000000000..bc3ff815b7ee --- /dev/null +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d052a74b30821e0dae06c7b620ced02743e88bc84347a2cced680f0ef889fa51 +size 51863 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 8c1a9948a41b..bbbf37f941c1 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:ad9d51412476ac4bd1aad382efb6259cd29467849f24d00949580a0b1d1d980b -size 40880 +oid sha256:0e7bfc2d4a80f0d42834cd2dce137a885335181757b1d462947d5726de65b520 +size 50339 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 37483f8c17d3..5d7ee6ccb1e4 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:043837aae51e6b8c942727e214a70547279287c9a1ef5f9f69b826494334947d -size 38197 +oid sha256:2b60c79eb7c07d433d11f1a1c17a8d257497522d5ae08ea39221daa11be09860 +size 46636 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index b6a31f24d890..f30518b3d7c8 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:b7f8e5e80ad6eb6641bcc09df0f7c50d7d23765da0e54af916d07b81da604dcb -size 39334 +oid sha256:4c23cb6dd50246d25327d5fc22157dcf8ba5b66bda90ea3fabfd5728749f771c +size 48064 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst index 5bac464d6859..38af338be080 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:8556a0539ab50c1fa34273ed8922045acb8020fcb804a6b866fda6b030ca86ff -size 57621 +oid sha256:b0826647465d107525d5fd7d855cd26e6ddaf2c099ee181d97ca8b839fe1cd7c +size 71814 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst new file mode 100644 index 000000000000..8eb55bcf1088 --- /dev/null +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3baf6358c1889e50a0d0acf425bc864e85d4246acd29b63c116d35ff97af997f +size 72485 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 67159547ff61..230b6149ff31 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:4a26c28d7f107563ddd65e8ee2bac688a2b9c627c75ac4de17838e00fd683afa -size 57106 +oid sha256:2f7e787f889973adde36fb9e4a7a27413242c4c51b2b97e5e0693a02c15e80d8 +size 71320 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index d96323c31285..25a35e465bd2 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:cb9f2cd7d7f2b15b9affd06e3b4d2adc20c2fafac2899b7e31e9cd94d0c8f320 -size 54431 +oid sha256:c5100f2209b5482d6867941504d3e434f5270b8a13541e5e634c49b4a1544a97 +size 67729 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 9c0c7a066493..593ed94824a6 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:9a69b5ad35655d85d325c54f0c96aa57360b26b6fa55825bfe07d6dc39e840d7 -size 55427 +oid sha256:912d2213069a25025eeb2bceee8b9c859eaa900366f1f3632873b323b7ed4661 +size 69111 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst index f9952e3de879..ba0e9b026d3f 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:b62a539470bf77f848cfd228203e94e3e67000e05261a50b189d2810c095eb20 -size 44462 +oid sha256:c325fc42710bac1e6669460d568801acbaddb366bab2f11ef0cde1f0af0d5cbd +size 55255 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst new file mode 100644 index 000000000000..d1538e648577 --- /dev/null +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:26f541eb5ffc27a82234ff6b1fa4c8c48602f5fcaebc6aba72c9a72574370be6 +size 56282 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 422f6a74aed8..80632d43ed6c 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:77164295046342c608d114faf50384a168cddcd0bf672c6f07d4f7fe30799710 -size 44134 +oid sha256:808832afd20fc7ffc4b8bdb3fa21d363941540a4cbb506544492c67d2d60ffc3 +size 54819 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 5284959541b8..3f4ae6a83eb8 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:e920e69c5fc2ebb62f23b91d6cb1b5c1da0c9c630916ed8bdae5c6da83cdf10a -size 41416 +oid sha256:a0ecd50d1c83bf220a4284433142c91bc8e4c24d8ffbc2584a59eb746939ab7b +size 51061 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 005cde433e5d..fda389793192 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x256u2_s9_et128x16_m128x16x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:9cea69caacbdc1c7f0477eb7f60fbb1084630e882689ef76c0583aabc2b29fa8 -size 42567 +oid sha256:a66f2fa009a0860ff35bbbdaeb36324e36a404c88a5f058d085cc5f068459f7f +size 52580 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst index 7850a4e97e1c..4f53600f65be 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:1642d9cd2333862bddd517fdbe714b4b12787a4fcee8309d80c4f6261d0b67d7 -size 64037 +oid sha256:b9556b14283eb3fa0c6f5c58e8fb172703f06f689e6edecbd41177c257f21205 +size 82093 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst new file mode 100644 index 000000000000..afea9916eaba --- /dev/null +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cb047860f45d4a75db374fa33d56bc00d83593c582336dc9bdacea195229b77d +size 82748 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index dad14fc006d9..e5486813218b 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:02b3ffc7c758c0c1c6bbf7aa32242dc44546746e42db319bdb0c6227bd16ef07 -size 62129 +oid sha256:504f6f1e41973a2554800646b668282a33354d39108699d408504c1019a553b6 +size 77335 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 55a286992b14..9585b7285e29 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:03376e7ef1a15c613c1e5ebb9f58756d2e238016c543ac688984d0b6b845a844 -size 63070 +oid sha256:6b09c658d31f522931afe7bff0a485bda2be34d31d8a477b6b50aa8ba8b3fb00 +size 78660 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst index 149aa9ca90cf..45b1e206368f 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:2db44b07e577b21330ca47d085d8c88975a713a91004f60452d971946810c5d9 -size 63700 +oid sha256:e2964b16ef2c13d99f31798a7d4a6baa9707b3bedd43955f1b9792637b8302ed +size 81976 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst new file mode 100644 index 000000000000..6708438c833e --- /dev/null +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b0bc84131c249583d5c531acaae07577fa6033f7889efb707c3c28fe119ecfa +size 82237 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index f376c7062429..3f962ae321ff 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:c661109d34f70c7dbb67e697eaf3adc10c1249d7f01af08fbc7f9a6e1a6d8e98 -size 62068 +oid sha256:3da13671f7b9a3066b8ad95888f97bf74258957a0a6556672dbe14e06f78285d +size 77705 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 5727efa16ac3..13fb59dc2851 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s4_et128x16_m128x16x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:98cdf94604f0f907a9635ccb79f695951bf84230a15a48b19531c25d0303475b -size 63102 +oid sha256:4a10173b57a3f25530fa65c5a568d2bed247604e93c1b727f74360bb7751bee8 +size 78973 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst index d839ad70e583..a6a147ced9c0 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:f1cea2e25e1d102d8660326575a7baed0c5edd6fcc5594d9d95d693b7d9711ba -size 59120 +oid sha256:851ccb692f8a8b76492b4084b0db1393dbd14a4b60b7d3616a2e3ea77fbe63eb +size 74051 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst new file mode 100644 index 000000000000..05fe9a0ccd5a --- /dev/null +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f9f90cce6c648426aac4305956c0d9a120afdf7cd102cdf5bee657467dc23067 +size 74848 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index c6db01bd05b7..a63a83d0ee33 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:25485d7748e927143246535bb17c4368fcb0bcad504332db95d7065b5602f5e2 -size 58972 +oid sha256:e32fff5b5870cb44f9f417e106f4be8128b95becd83e121c9d5c15b90ceef8bc +size 73751 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index d68e68798c70..0ed5e6ebebcb 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:6e1a76faaafcc4d24de8d9eeb54e6e1efcf7c7a5c809ed054ba9f2555dada366 -size 55917 +oid sha256:8c170a911238c20bbe19800a99441a508c855b55ca73dfb7e90f2a2007c629e2 +size 69637 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index d83964932d13..d7bdcd73b601 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:5e94459b1cde6fefa8bd6fa6ee127f97c574f849217fe97b3f35d884a12e010b -size 57249 +oid sha256:cb0f617565a20d6740b30e9bba5634464602d05f6375cd09eb86a492e45d22bf +size 71202 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst index 6716b06b7963..60e4069ea556 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:ca2aab24a25765d86eb4babf6ba3bab03541460ce66b057c214001373267367e -size 45343 +oid sha256:6a67905fbfd0024dc98f76d16adda12017a9d3fdc089de90bd7af513a2d3ac95 +size 56190 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst new file mode 100644 index 000000000000..4c396a78f10c --- /dev/null +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cf30fb716c713cc43966cdfbb4f53c784924455def5a457057f90e1bfa695258 +size 56971 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 98aae6cf44b4..c7b92cdabb88 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:a5c2f6b1d9a60df2dd4c59c8571c1cd0a31bcf7153a6aee05e0c30d1f939c5c0 -size 45021 +oid sha256:8f049c058cc4666ad0ab62eb95ae7d3c130f76c83c76f16b42a6788abf45d99d +size 55740 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index dc341313f40b..cb3098eadba4 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:cd24769260f7b9072e5f6967844cd8312a83ac1cde026cc0c2fc1faad18e5545 -size 42167 +oid sha256:7efa8db589e1b6efdf89e1be72380b71d5d5b0b071d432e0dc53a058197a449c +size 52146 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 3296b1016b46..b99bec8335aa 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:7d1f04bb10de18e37329e8718855c1fa26414d351b2a26050ccde460db4a6ff2 -size 43243 +oid sha256:a3532014b119e60f79f7b44fb61cb3de7e7e1d5590ce45566aded0eca529b759 +size 53459 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst index 25ac071c868a..9d661930e779 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:fafb14c3df10022ef0d04c15fc49a324dffeaf9e3ec48d44d3ab7a8f4f1f7926 -size 62932 +oid sha256:145c365c94a9491e68b8e21e0730bdd5fda263a723f76a7d92b9b7961e0abbb0 +size 78442 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst new file mode 100644 index 000000000000..e1fdfaaa3f95 --- /dev/null +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:65951a46783a4680b38296ac2914d625748332ba4e7927c1eff0f923dfa1cf84 +size 78896 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 84de040ce45e..1282456db4aa 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:47ebc145318da8749a4118acabfda5b50955b1419de7cbcd5b449ca701a5f4de -size 62637 +oid sha256:8c4c99b79f9639eb65bc4db40131f4e9b854ccbc25c365dfbf095bf8fdf7c141 +size 77918 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 5a22328ad072..c86c6dfbf8ca 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:148b33d97e8d2e058f6fdb451266bc3eadd52d45621a4a003a962359706b2e11 -size 59715 +oid sha256:d527cf3c086c4ac736c6af052efdc59bb6f03aefaeaa4035d54614696754cfec +size 74504 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index e98d3f40fee3..279411a51095 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:fad489206b96b0293ee21faef5932884419d33848edd1dc8eb547174616310c0 -size 60868 +oid sha256:9ff9b9177e7a49636f370983fd32d8b72de075b65c15e43ea63abcf6bde7819c +size 75967 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst index 8ca61904127b..ed2adfe95756 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:bdbe48eb3b2bacacbb24e6b32263fd1fd369261c15192e31b2909d453dfff15e -size 48947 +oid sha256:70fc808563e8738b4a8f7d8b0e2c6be2166aff600123cf386a02cb25693d299e +size 61048 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst new file mode 100644 index 000000000000..46217c0ab977 --- /dev/null +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a1e0aad28807f4512a11b54c73be726571264d5fa3cc73e770689a44a8aee3bf +size 62067 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index b80a7a2e04d7..6d28d95ce14e 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:f9cdb0dd6171cc01ba47a44d44838ff8501e8850672e082742313ea1fa23939b -size 48691 +oid sha256:34e8317fe8ffd0ea3576f3cc46f64434db4cca687873d77169b92381902a2ecb +size 60696 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 84ebfa6e1c7a..38863692cf14 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:0e40da2bd95845b3e40cb43b460c7f09fe2f9860a27a2c11aecc9d857f1fdbde -size 46374 +oid sha256:7b26827f276961f54e0235b00ead6c2dd661994fb752bb189a9c312276cca3fd +size 57285 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index a0f7e4696492..4e10ed4b44f9 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x16x512u2_s5_et128x16_m256x16x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:c957af9366fbd2adbb808606875220a19cf63be13c430831f59c6a8251c983c9 -size 47366 +oid sha256:1e9ba88cbd94916dd5255b0579d3da072a74193c5642461541f7f76491db44ea +size 58618 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_geGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_geGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst index 0e672879de31..ad6df3a394a1 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_geGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_geGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:da65e5e00e4995b05c986a6fedea51dbc3e10c723049900badfb961a024d87aa -size 93665 +oid sha256:782605039d5975d1e4d6ca82491a426a17fff3d52c268160c4091dc368ada381 +size 123081 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst new file mode 100644 index 000000000000..e9447a2f68cc --- /dev/null +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b1dc8f61614b79f2478a772466cb7dace2c5f866e52a79f53e0d676db4e0855 +size 130812 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst index 233fdc9fd5e1..8c49792d4f05 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:46365f3d66e10281fe3f9d69515e861cb676477f01731ebb742960de2f26d653 -size 91137 +oid sha256:fc59d54af4746457c1e589f67fe91ef89c7049a789d06d82d63eba85da779583 +size 120065 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_geGlu_lbW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_geGlu_lbW4_dynB_sm100f.cubin.tar.zst index 22f39770b433..97d3a302c8e8 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_geGlu_lbW4_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_geGlu_lbW4_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:990ba7af323ca426f1e1b1ba0133d4d95ffc23b35702b2abacc9490eea50cbee -size 134471 +oid sha256:bbee3b2cc137b45865a897e9f10e464f4dc6be118a71c91f3a100f7420feccad +size 180572 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_siTuGlu_lbW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_siTuGlu_lbW4_dynB_sm100f.cubin.tar.zst new file mode 100644 index 000000000000..9353070a90fa --- /dev/null +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_siTuGlu_lbW4_dynB_sm100f.cubin.tar.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d5dcd546433cd64f9605c0aa60ad920aa437408ca332401b2b3b9e9a49443e05 +size 192858 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_dynB_sm100f.cubin.tar.zst index 523842e88031..9ad33c49a3ca 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:53db89ff6e8ba3e36fd8c4fe735fea3c597001e05b9f554b927ec67cd93cf095 -size 128778 +oid sha256:f41e2cb9d3d3ad83aeb62337124ea8a3d060f9fd9e3cdad83deb5e18c29ea634 +size 172751 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_relu2_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_relu2_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst index 2c8051e23cb6..29138179c5a6 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_relu2_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_relu2_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:05b0cb7316e62afad3d31815516828a931ac1b32343cb10b3ba77216ec187c27 -size 80216 +oid sha256:0166af14263c59e033d277c420380c12e33571ffe18828e9cb2bba6be9a4ebce +size 101951 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_relu2_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_relu2_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst index 5fd7de763dae..362d22015745 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_relu2_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_relu2_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:8011a3e911204efc061707ba6597e36c03093ca98b7eadccf41dd1e38993e170 -size 106027 +oid sha256:19df5aa60ab5e8df55cb2a85d00b8ef79a4346e7fa7b2aab44da6b94c758acd8 +size 136859 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_silu_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_silu_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst index c7efcd70a24d..b901c1905c57 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_silu_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_silu_eW8_fCp_tmOv_bN_tma_ldgSf_rgTma_clmp_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:2c81e77c303d9cffac6c766a0a451b459a455eb3665cd4256b651e3bb208aabe -size 87659 +oid sha256:92f8a8c5ef4bc302949b08f924db23996c946e930129cfe5ea9144544a5eb161 +size 113215 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_silu_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_silu_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst index 66b9c53c2d8b..df7c43d25963 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_silu_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s4x4x4x4x1x4_et128x64_m256x256x64_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_silu_fCp_tmOv_bN_ldgsts_ldgstsSf_rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:dde9f15e747be92be953bee3df30340c138d754949124f7964e14045e885c3d0 -size 117659 +oid sha256:30f71cb3bd614ef3f0da552bd5e3113e5bc827416dd951bb5a110909eebb8d4d +size 153430 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_geGlu_laW8_lsfaW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_geGlu_laW8_lsfaW4_dynB_sm100f.cubin.tar.zst index 16d863d307a7..1078a4058b74 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_geGlu_laW8_lsfaW4_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_geGlu_laW8_lsfaW4_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:2a6b2f083361835ea52c2c7280fe28008539a8dfcc9dfa3957930f7104b74d6f -size 110712 +oid sha256:8b1cfa94fd60f2a72874c67fb2a4ed39ef95f47b26b6c16d5077e4c004946169 +size 148731 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_siTuGlu_laW8_lsfaW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_siTuGlu_laW8_lsfaW4_dynB_sm100f.cubin.tar.zst new file mode 100644 index 000000000000..ce10de79372e --- /dev/null +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_siTuGlu_laW8_lsfaW4_dynB_sm100f.cubin.tar.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c30750bf43558e1eaa7b26377e437e12f82c2d371e00e00881182ba12a77416d +size 158775 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_swiGlu_laW8_lsfaW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_swiGlu_laW8_lsfaW4_dynB_sm100f.cubin.tar.zst index 151ad3dd102e..9e8aeb951bd8 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_swiGlu_laW8_lsfaW4_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_swiGlu_laW8_lsfaW4_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:0e7459b834708ccaf0a93c2547ac3501ed0698173e6f6cfc76bdfc6ea0f3b980 -size 105304 +oid sha256:c456b2ebea4bec0c88409e52e135c41ab2e709b4626cb7dfbb6b143b45a5495f +size 142265 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_relu2_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_laW8_lsfaW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_relu2_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_laW8_lsfaW4_dynB_sm100f.cubin.tar.zst index f7584778c1ca..48b55bdd2f41 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_relu2_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_laW8_lsfaW4_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_relu2_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_laW8_lsfaW4_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:8a1a524b52ac4fcc878acabda15a0555ea4a022e0cebee6cb0e866a927084b6a -size 89793 +oid sha256:36d5ef77cfd08ac66860500eb4584485c3a02a22bf7bd38c0d6beb85c57f8f21 +size 114608 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_silu_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_laW8_lsfaW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_silu_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_laW8_lsfaW4_dynB_sm100f.cubin.tar.zst index 87a3c838a4a0..5751d8dc3f6b 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_silu_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_laW8_lsfaW4_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x256x256_s5x5x5x5x5x1_et128x64_m256x256x64_c2x1x1_rN_TN_schPdx3_biasFp32N_silu_fCp_tmOv_bM_tma_ldgSf_rgTma_clmp_laW8_lsfaW4_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:3a9bded9ad53ab7e107049f7a1f117cbdae683df8d4e22f3e986b5b96bd3489c -size 102860 +oid sha256:05cece6da98349f80e9446ff8b7250ea48383f9300752c8d5c658b6ef986b8cc +size 132954 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst index 3675b4765aa1..e50b9b00a55a 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:9e5a4d81efd518227add83b2c2340a92b921bb467713754aeb7b86d8ee4c3a09 -size 61169 +oid sha256:8f6bf755203edecdeab23433631c57f5536d27b62ab46c238ed1b5c69aeb92a8 +size 76801 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst new file mode 100644 index 000000000000..72a1c49c7130 --- /dev/null +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:015a69bfd4d26dae106ca09d730f18b7ffbf108b7b4fd1c29211e88203c2559d +size 78362 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 93b592ae2c7b..43436841ffe7 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:b5ac76ba9aa719a356476121cbc40006ef8d952249463b6cc9b147603571b462 -size 60565 +oid sha256:deb1eec7804920dbc566eadfb85f7007489349bb0a0fd9ddfbc35df8e76f28f1 +size 76223 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 6066d633551c..1a0c897eacec 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:954469d4f61ff4766884b55efbebef6b68ff16b9f041cc7f337da72b27bd3c65 -size 56498 +oid sha256:144c215a27e4a970ae898826ec583255691377b4df0b66ee82c8b935c81d3374 +size 70713 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 8b4487252e14..59dd21594480 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:0bda9a30fbe4a9fea2720298e665ef3fb49a51e7542c4803f19c6b4c50cb56c5 -size 58945 +oid sha256:593c8029a609c47388529550a590cb17f7f9ff168fd1dc2fb4782e15934ee05e +size 73496 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst index 175ecf06e703..e7786ca27b8f 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:48e1560a56064831bf51a282ef90ec687e510f1e6e0d51026a0f434fb4086c5f -size 49478 +oid sha256:e663fe358d78f46fbaeaed44b2598a6553a82d80a1b06f5f461cb7cdd2a9cf17 +size 61977 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst new file mode 100644 index 000000000000..43b01d14c372 --- /dev/null +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:320200a8e9b965c32978c5930676aa84961a6e5d6d059f9b3c9ef5d11f728752 +size 63697 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 2c71938ba367..2424e231b68d 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:36c39ce84a045a39cdc2a246ab36c4995893d6729ef94ad3f6903b342303a51c -size 48915 +oid sha256:88925a073cb379695247c069bea96aa3893a1da6c4c80e08816053ec75bd6e82 +size 61539 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 4a463f463271..fbe1f0dd14a8 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:8026417bc38164bdb6a8a7ba327a78a997aa85e07d2551b2d5092d49f093bca8 -size 44736 +oid sha256:34e51d26bce45de73bfa406a70771dbf05cb628d5cb793dedc1256373f98e74b +size 56148 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 9f9f141eb413..7ac7465b5e9e 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:2729cb70ace2e156f6fb66ac951964a88d0105bdd20bb07cbd1561de690ebcb7 -size 46872 +oid sha256:715ff1796827e351fa97ae57220295e51ee42e863deecb86551253c17084ff8a +size 58897 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst index 86663b8404a9..f9d1ee0a1cf8 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:9f2d9faf4636631c0003e2e6be934c62abb0558a672ed0d444bf8ae5142a5ce2 -size 64343 +oid sha256:82c4f4605c938b7e1d5abb23a72402cdb40a6505c4b62bb8ff47d39acc49e46f +size 80578 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst new file mode 100644 index 000000000000..e29df5179b8f --- /dev/null +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ef704c407c5513c0f3c16cc9a2c2f0507c9f0a8a6e1f20f447d3413261dd2bc4 +size 81851 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index a929fe28f430..bc74fb156a98 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:88f261a8d297da016cd0e534f5d05ab9a2d7008739e02d49eff9cbddd7745620 -size 63766 +oid sha256:852f0319082f319c0e8d9b3b9ab4c81ad21b0dc84164d08b5a8df731a11a6fdb +size 80012 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 87a9893a3026..c2ebcc885e12 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:3d24f8055a1fc6935526aa8512bffb295ea7e5a21a92bac2d20c9ad52baf4724 -size 59608 +oid sha256:43c6e7874a2200233cd30e401ba1cde23c9180c455b8477fcb7cfb73a20150b4 +size 74678 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index f6dfbb7c4696..cf61af66d561 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:a1ab4dede285bd57b08576095dab9cd5a0050a7ad90043d06d7a9a566bb11ac6 -size 61869 +oid sha256:c723b132323f9da79ba5be61330cd56793da22ef3370562f4f1a4945cb1ae249 +size 77243 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst index fdb7ca155b4b..17a5ff20bcf1 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:85f772257b77b00eb18e4bf1580ecbbc990a138e13df72af9fc4f2094bf58c6d -size 52636 +oid sha256:6af1b3e467fe1acb5015d16a34b2e640d285a2d3fc040a505e538d9eff595b9d +size 66135 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst new file mode 100644 index 000000000000..6459a8807517 --- /dev/null +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7b03b4d07e9046ddd029f03270e6f65a20b0cb6f66e182f1ee9c1b6c91cb53e7 +size 68016 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 74f49192a2ea..2a1343019bcf 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:c63f170023b97ede9d2d7bcf51bf9333ad9ccfc2f7aade8100ff3873c1c8ae2d -size 52005 +oid sha256:963a41aef30a4b041cb7e7019c0f5fb657f72d0e4dc2de8392c4eab0ad55d9a8 +size 65717 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 28d13449d408..e1ca5083db02 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:9d4c33f2b40a59868f34e912ac7216a683990340b934e7a91dd6103f047bd20b -size 47949 +oid sha256:4fc3165ed28eb67b78781c7b23745aa9037115f3f6273f115c0b02d2d07c012b +size 59833 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 055bc040b1d5..15d97240ea0e 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x256u2_s9_et128x32_m128x32x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:cd56f5d8c168ee39bf3350c8cf15df45c5ba37659267b5e47ddfb6786dde9660 -size 49970 +oid sha256:e7e93cb8bd7d9db672d7f4cc4ed30243f2c6c25eb26a7428c9bbe3e387127adb +size 62652 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst index 736775540422..60443dcf5361 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:a1014222bbd6f864b1556d39c343a94152d1d9080ac4b6971b5b1ce058a24443 -size 71685 +oid sha256:2a3dec885550a6e6a63721a4fc1270021bb15c82a24a4024d15552601a342ed4 +size 91618 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst new file mode 100644 index 000000000000..dcc252c5b658 --- /dev/null +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eda096f0d6852e5ebc75c1827f1df8069be9b2b85fcd69f1a11b6772d466e193 +size 92312 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 1d936beef160..4f96a51edc5a 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:dffb3e4bfe6a1fb80b176bd612bfb5de5b848fff0048a024ccbbb25cd12c3154 -size 68519 +oid sha256:10b4c3328d2018d7b3488bd7ab910ad52ae14ccad121c20c186d3926bf45efc5 +size 84105 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 802d07f7bba0..ad5235a1e7a7 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s3_et128x32_m128x32x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:e713266b25ab2292a777b1d85ed847c332816776d73c2a98ace8dbeb3173f03d -size 70726 +oid sha256:585eea8e66294aca15c983673c8b7f13d7e02e79f9953894c48424b2f58298cd +size 87368 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst index e332e8e417c9..1c5b1b5d2745 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:664aba0835d4caae473c394dd9bb7606a3f69b74e93a79ebf1a95ecfeb01a50a -size 64215 +oid sha256:7d9dcfbb62ab0f59a4d9c39fd7c1f26def4124cba7fb8a9724d1c67f6e531be9 +size 80090 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst new file mode 100644 index 000000000000..4201f616bd39 --- /dev/null +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a4a005d42fb8fb18c0e5ce340e6ee1f9f7ccd61f0f49b26ecf120750d61dfc07 +size 81886 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 2c6066775f0c..f2783b345c39 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:c032e52da88794423bb62bf3577cd8aea700e44e3dcdf599f849cd7bc49d767e -size 63598 +oid sha256:3f135f14bed43d9b3bf822ae94586a6034e5896da61f87a4e9b4e1744d9edb04 +size 79748 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index b8b7a69e1e9b..b9f6c679bc0f 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:1ee2e3ac8d2768abd5528ae1a2c6af98e180725a33930e53ce23db8bf5a82d5e -size 60306 +oid sha256:3457a0105e486af253d996b6a8070b0296ae742b386c73c30b3e6d9afcc07b3a +size 74846 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 9d4be011ffa5..02de925a92b0 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:0f1e9ea187e6c74ac6907178d6c62a14e061cd63ce524a629b6299c9c466ac9a -size 62407 +oid sha256:169d5974c4178beeb7d486469d64a4dd545e0b63858d14dcbbb5ba6c48956f9b +size 77644 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst index 04f8ec232c53..399560bddbe3 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:e3003e9f8294faaa840b364dc409e18d363a0aa3a83cce6069307d8007594355 -size 50607 +oid sha256:9e4b15ac6f5c43846087388b40f20becf2339823b2142673fff2848f5e4f3a11 +size 63464 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst new file mode 100644 index 000000000000..5ea2cf7a631d --- /dev/null +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a6610ebd88839042c452a2cda0619eb145210dec772b393579dd84e50ac08220 +size 65176 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 96f307ef4764..50dde8d5710c 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:17fa26c666c773d255f57688e758ad8a91585191f35456a324e2d7de10a20820 -size 49767 +oid sha256:71f024d5c06ecaf4691d7979b836c65ccd9a6c00e6e5202601a37d6671e3fcb8 +size 62667 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 0131bd4df40c..1fa53e5a9c53 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:d20f76890a84164adc3b5738f3697a946c25b75b6cc0494c83b428269bb93de3 -size 45869 +oid sha256:5e867a42ee799135642630f2835248d2ef9baaa5fca550ebbb9f695b7ba46973 +size 57054 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index ded888ff5a64..c187bbe0a1be 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:4b1cf4ea1347eb5ee0e9eb6a59f802e75100f602fc93279d635c9c58ae59ed6e -size 47635 +oid sha256:c6464cd479c00faa3d7e22d994ca8ed0df7d92917d2b1a4a02d6f293aca105fb +size 59263 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst index de8b4a274988..4214407317ea 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:45d2dde7ac84a4c799fa98f91873311cbd01f2e11f20a9d40ae819650aa52aa1 -size 68158 +oid sha256:f48e8791ef421392893478cb9cc1fdd561826c564276462bfd79004200eb4fa9 +size 85128 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst new file mode 100644 index 000000000000..6d2bb4b5450d --- /dev/null +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:93aedcb1f0fdd021f9735f80f9bdf801a9ee1fb1ce13b64ea86bba098172e8db +size 86799 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index ec436330332a..8f135ee2601b 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:1f24ce4572dbd87b321e847eabfc9c5560c7481995224c7f4fc326ab0041c146 -size 67353 +oid sha256:9b21e8f12e48459c90a0978bbac07d9a1e23a2135d3b9008ebf2dbbd4b5ca326 +size 84148 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 1c7987fe3aec..082b977f9323 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:6b3a4844b1fe7c7f3362a5ade0404a9d7d0c820022c25b33fbb6c24b83dfc0ef -size 63592 +oid sha256:950c4587c03aea86b72559992b92e93ab6b8cdd3320bdb10d0723257d92cf393 +size 78902 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index c3785bd5e5b7..977294a42d52 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:2568ed48c4904fc7d44a3778f206cf4898a0c01e95dcb3c19158f4fd699a6da4 -size 65728 +oid sha256:90160eb594a1977b927fcbdd457b52f90d9920eae889c01c5262e8a4d32b649a +size 81620 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst index 188785eac742..367d2a773b08 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:61030bfc78995f0c22270e70a65a13e915a4e0c7ecbcc3b7970a5856f7a4d6c0 -size 53969 +oid sha256:2b197c7d84608945c03fe97565eacdadec8c8e319b9f7639f68274e794ada97c +size 68608 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst new file mode 100644 index 000000000000..858242aaf460 --- /dev/null +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f2ab9b52b1b24d9cd26fee00b40e3281be680ff9692bb362210adb60bd6b2aef +size 70591 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index bd34cf6f6a4b..11e61c30197f 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:335dccc8cfc75fafff5f1616962bc1a02f8d457981243a65e81274296689665e -size 53525 +oid sha256:7b73007fa4d9fc73a9ad8d110093191d2dcbc493fa26b14407b16010f20c1d04 +size 67915 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index f1d0a2c11c08..b78e435ff0af 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:db18120f0658f9c0ab2dd83b90e0549eb3f2229daa2022d5804e19851a3f7e92 -size 49608 +oid sha256:0e8b722a56f73b67f63d5db3cde482cd06be69949c8df8749ba5ecbf30b7535e +size 62254 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index bbc627cea315..5de4e456eeb6 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x32x512u2_s5_et128x32_m256x32x64_c2x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:069de3cb6642ba5add3f7dacadc800aae61f56d81ddd952941de44c2266f584a -size 51672 +oid sha256:89cd9e6180b90743d3ffb273a4eacb2a406360ee8715188a82aba2d1f1fa1d3c +size 64806 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst index 4c783d671c57..45b9da725eed 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:a0f0f935232533bc0055384f67b0fb0e479153b931599a5c9e4e8b9410d94bc2 -size 85935 +oid sha256:3fc55cfc8fa289612fbaa43f07b0879957b05a2da258305f38cbf47b23c5598b +size 108996 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst new file mode 100644 index 000000000000..d9a813d05db3 --- /dev/null +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f15b6106081da112e1826e000675edd371e8d9a62abae390892e94dc0c02f5a0 +size 111986 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 478a8f1fd3ec..e0465006366e 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:86c119a9d40a2bdccee471d0e240ea5cd6ec1a55d63e2f00ea743893fb0c7ba6 -size 84083 +oid sha256:cb32cf8a1a439e8c37bfc38ec34502cd7d035c12c228d177f74ee0d6470eea3c +size 106658 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 52f3f9ea3b80..258ab0a665ee 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:becfff1b7bb6bf27f58ddf06f6f707ef0c91f7aa385dfe49d169f2858df1aa1d -size 76795 +oid sha256:8fc3fce615f1339068e533ca4c1241c1876095ad1ac9306592f3973caece3024 +size 96220 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 9c0ac12f5dfd..51959f752037 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:478034e4964e7c38a4665c693eb4d891ad23d8448d5acef63e3ea62ecd469e75 -size 81269 +oid sha256:5dade061520f72086655023d18eac6dcbb4e54415ea5a7d54112d8d211ee56f6 +size 101226 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst index a68262c427d2..abddccc5dac5 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:a3ac3235c6aa111301e66e2acb0117cabc376a76aef05e98493a220f9557784a -size 88838 +oid sha256:0f6004f3128ab0232b1b4d617b20f3fdec3ed1c0c22b8d86749687e373884241 +size 113285 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst new file mode 100644 index 000000000000..e9c3a39ca209 --- /dev/null +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:47130693824f628a60ae0a4304e8eeaaac703a2dd160be19e254f7e730c79ed7 +size 115289 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index d89342071ca9..c81bb787a243 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:9bfc7eb21b19b551b408e5cb45302b55e1965c312416add1e14fda2950fce8c5 -size 86837 +oid sha256:a3064e4f3ef21dd1cffabdaacff352150e9c750ca9640d3d0dce92ca4151a494 +size 111484 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 02a2eb185681..47fc05fc81ae 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:c85bfc1f04793849e787c3ffb16f46e616c3d8c5c31553b1d8792a3e1ae52acb -size 80314 +oid sha256:41a797c2691ef867204b1f96c5e37c9560d2d44bb9c2be6cce0cbb503e2e38f9 +size 100617 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index a826fdd9aebc..950bbe43f82d 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x256u2_s6_et128x64_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:b7a88ba8ef84ffb35031620ce2e7be6bb3aaadc080af4e70e0fb323b2343b14a -size 84302 +oid sha256:f59a3311c0f3b45227710ef253a6a33c3e2dc133186e15fa2f31dd76f07ec598 +size 106135 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst index 46bcd65a9285..9f83a1836b99 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:a96cee6c22e801bcc5c72ac583a6521593077405d1c3c2721a50ce16c8b1a81b -size 76514 +oid sha256:31d6c0898c7e7ef3dd460b36db5b117839a7d14303f86bf71fcd5e610f4c13bc +size 96540 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst new file mode 100644 index 000000000000..b77ce2544149 --- /dev/null +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b9cd4579dfb3d8785ac17156477bea26b59f39c12b384998ce8e7aae2b3aaa9f +size 100088 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index c17ac0c8de51..856e26b59745 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:9bd2dd91aace8f48c8d5df19bb4e2dcc4bb7f42e25c64de45229b14f8c4e578a -size 74719 +oid sha256:4691ec5e542451944148a79fc55b426a0be11f2c6fa7120b7bba03cede97a6c7 +size 94478 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 4cd5c77d34ee..2df5088b6285 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:540d3285d924f4fc9370cff8a54bd6cf6bf9fe819781b8a33237046774cb5727 -size 68539 +oid sha256:96d11f7256fa9bdd33ed6254e47b989aa44aec2b74da1229ba24b42a8abad9a8 +size 85609 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index fcdda5010722..ab91d6b6d6aa 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:02dedee4d33068106ca7ce53a58496bbc1efca53a4dd5185241ac3935741af8b -size 72309 +oid sha256:fa8b7a986b98294cb0a29fc4f43a5cc54e809053d5f72e15b798f5ab8726e8ea +size 90407 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst index 862daccba02d..c58f2138ff3b 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:9c0bc75dd7ff3a35535b89e5b37c6b90c717e6b9bb39cb19e0858e3a7039dcdc -size 80630 +oid sha256:57a68eeab0b2460570c2483203c150f09fb52ba2c4222abb725523befc7853d5 +size 102088 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst new file mode 100644 index 000000000000..e54958ae1af1 --- /dev/null +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dcb0c9eab47b7035fddab36cfa9d0118b7aff9816882d11353bdfb2f9195cba2 +size 105485 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 63c1977fd99c..3bc18d8a21ea 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:07c9cdbca6a10d265ca5452d376d28dc9f8e217d4c884868873458c30d4659d7 -size 78889 +oid sha256:7123d189d7c43f3ec8b0decb528ebc3dcea8e1fd4794d512fe497adeafdfd83c +size 100122 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index c21a72cfaaff..cac0b3f8d6bd 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:4e14d5c6dc0e0aa11dafb4744ef3388a0673b8ab9270bd6746572a5a2d998f60 -size 72381 +oid sha256:60e5915f9f1319b8838e508e324190bc1d82b887ad27cbfaf8a7fe4337ad8cf0 +size 90758 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 16a3b92770c5..4548bad4caf4 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x64x512u2_s4_et128x32_m256x64x64_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:4510ffa96738510f8f872b286dbac61b089fb2fa7e5bcebe500ac0c2fe47e58d -size 76842 +oid sha256:b5bcb1a5392235c4686d2ea845fc239d93da9b4ad388b9b6b5253d9a87e4ab52 +size 96279 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst index a0acc359da17..dc4123c1cbee 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:af42056e657b12bd9f0ff9e24689eecb5bd6679e4bb8d698fe5d3097b798c2da -size 49191 +oid sha256:76fe5853d27026c446868144840906aa33e0512c0978421d4a9c2b24213d5ab2 +size 60817 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst new file mode 100644 index 000000000000..45a5231d51c1 --- /dev/null +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:47cf1fac259369e7687be1ca8d4e46d8991441a6bab1f56c20588c5a85e0926a +size 61141 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 07161dd8fea0..3db326190e16 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:8f435d7e95d1cff067feb343845e98d5121a1e7438108359bb25b0b483bc5ed0 -size 49011 +oid sha256:a10b366afea42d418ed00ac653d077029087d5db0c9d29deda60e5e9269a1c47 +size 60636 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index ccb6246008ba..00e744f6c4c1 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:699482a2083ac68d64195bf98fe6a8ee66138696ab5295f93e55cfd40dc78b1e -size 47332 +oid sha256:5937e85c7986a316d2c0967253c9d150af361be24ac57a298c7e76cc27645ff5 +size 58009 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 5039743811bc..62fade9603e7 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:687dba635f513a440a9a4819a5ebbe0de1dc2d4f6e7c436f7d995c1d62d17a66 -size 47797 +oid sha256:3d275dd1f87d727ee6264d3c4abadc1cc32a1f7ec96f761b973915440b5ba7bf +size 58578 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst index 7535872ec06e..6e08243d2133 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:056ce0ffea9d07df839370971ae2d956dbfd1d14ba22026481f15460898f1e3a -size 36964 +oid sha256:b81e4b0f80e63aaed2e63a5b54cb4bfe7c216257dcf34b6cd2c26fa476c190f5 +size 44972 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst new file mode 100644 index 000000000000..f93c066baf97 --- /dev/null +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1a6700ed3dcfe6fc8ad54e5d008218e93ac3a44b4415ecc2e1971e5d45eaf891 +size 45390 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 21de025c5aa0..35b22e7613a9 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:f353a22c058b475307777b81797edb50fa36887b719ec5755c74c4325b07633e -size 36799 +oid sha256:a4b0c52c375bc267a173b78da3c60df03950a2161ed0b6576bcd46406766a7f5 +size 44686 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 879288300e3a..7ece9d5bf3b2 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:d345711712ef258436b5d56e99ecbfc55b75f42674e0cf20fab6a8b785c8c774 -size 34892 +oid sha256:b3942db42bbae89fd3c5bca411d1c9f92bdfda559a6091a64b814904d8b152f5 +size 42353 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 68c1e4c84657..dd03409e3d24 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:4c586b663a446a408d6a5aba653486c4a611215108c07e3582831df6092294e5 -size 35472 +oid sha256:8047409fa0fabd949b35f5a13356eb4642e1bfb951b7600ab9c1d59957bacaaa +size 43064 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst index 2a9e0ff7d3bc..6dcbc83ab65e 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:0d5c3a31753089b4034056444d316711d12c5569a4f34b71ec14a0f3d1a44ff6 -size 52382 +oid sha256:3e37f053503053a295773735c15fbdb5454ade61afc0ada038a49584e99e97b3 +size 64933 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst new file mode 100644 index 000000000000..df6907b88ed2 --- /dev/null +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c1d97232ef6ff036fd5437ddac04b895c91197cb32c3ede919c532513dbea0ca +size 65330 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 4aa531223241..e6a7afb7724e 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:8f0cc0db7a58dd1fea99496f023028886a73dcca569e0ca85c7cde1a1b3d6d28 -size 52246 +oid sha256:b1e823abebfbdce2ad4fe8a7046950088fab607c720bae2b43d75bada54cc4e6 +size 64765 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 2879dadc1005..aa85c9e56be1 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:f49229090a7f6213620c6f89683d9c9b590377cbf202be8daeadd9c51cd9f445 -size 50117 +oid sha256:5875f68ef6862d91e75137708d40509d27163451aa593daa2aa2d4248b5423f9 +size 61921 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 3963fd530879..eac1869936ac 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:b286ff8d054f96613a6c0fbcbf4a66c46d16444e322636fb56360efec0e664ca -size 50525 +oid sha256:6d61d397bf3e807702d92bcfd40291c9e99444881fdff3635ee3c64b420db359 +size 62522 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst index ce3bd56bcecb..361abebbbb4b 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:45cecce0fb44a7fdb04af0c2c4ed496938a94b9e8a27ac61883184cd0ba3b06a -size 40205 +oid sha256:92a478edd2fdb569a3b9d9d311465f847e62b1302d67e9158d545529d497f7e1 +size 49104 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst new file mode 100644 index 000000000000..88f7c32ab3f6 --- /dev/null +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:25da641bfaff9330669541b1c177b9b746a3ff162c2fe48ab6daa165cfeb0127 +size 49527 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 0a3abc84291c..3c5d2cf6e277 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:4e3ecf99290c71bb68642833e4cd8253a15ad1830cd7ad687573111f7e353e87 -size 39980 +oid sha256:a689f94124f52d87040d546a84a32c4def1c974554a36f6ac67a585ee378fa6d +size 48831 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 54e4053001c2..a2941cc848b5 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:b7c47664cedaca04407cc2485ac1089c16fa30f1002dbd0b0e39b4f6c222ed37 -size 38176 +oid sha256:0258377a84d2b71ea130289a6553a6ed277f1da9fa02f8beb00143a3124858ba +size 46533 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index ccbab5b299d7..d00ff3391ea4 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x256u2_s9_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_ldgstsSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:e08faa28161765af6ff40a18ef5d96fe67789f6ac796927ec2aa96a4992b724f -size 38685 +oid sha256:ae366c4371a413b6eff937282e44bb8e34b3530da2c407caa9eb586ff38ef32d +size 47259 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f.cubin.tar.zst index 6b8f63b57ecc..611ae9c9977d 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:cd72cf8811543f216def8464fef4671834f910c2d19107371ef752cb5c9b4e64 -size 33758 +oid sha256:1d93e95e1de12662e0498cc725926d5d00afe534f9c05d39063bb2ade3c0901e +size 41104 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst index e7d66b01dd4a..4130c317801e 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:6bc84657ba5c428288261c5bb4a5ea7ec54233429619ec5fd0da3d947ccce50c -size 50318 +oid sha256:c24695a4f1d94e55513f13b077b96735df084630dc64c34ae928a04bdfbd7600 +size 62416 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst new file mode 100644 index 000000000000..65cedd770a7b --- /dev/null +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:527dce85226b69a8f29860396933a4389030cf78fb5c27b0566f0d5f087ab0c9 +size 62900 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 63a9ea66ef5c..60e02f10c608 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:c898867089dce6ed8f46b74c198c2d7140e8251e7d0e634ab39d212720d53aac -size 50209 +oid sha256:abd034f8b19ceb5151e1ad5dd54df06777e16377b9d5361e64569e6b42e91927 +size 62186 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index ec777cf0eced..fccff603496c 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:5b3abd5810cebdbfc43c9981b30b346d5fa7346c87fb9f29c31ade9a6c02e965 -size 48341 +oid sha256:5e4937bac59d3568cf30a81d820e041b688c8148230606e3abf7e5f89ad364ee +size 59672 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 211a51199891..9e011650fde5 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:7b9a23f5d15d25a9c6b06b079861161b58324fe242aba757db7e2986ff75f133 -size 48984 +oid sha256:d82ce66148a2de9c32e4c6623508792853765852ae2a6acd62c9f10c7c8a3e7a +size 60520 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst index 806ff9675b7a..225dcedc037a 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:a87569751000e3b48acb9dc3eb841ccc983566397d5df9a8149a99a716219db7 -size 36370 +oid sha256:fddbd656226ae98b46e495b66ae2d4e3f26ab238bb8295bd1caed3a75c69eda6 +size 44603 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst new file mode 100644 index 000000000000..cc1255f24b9f --- /dev/null +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0f974f3568985130d2c8aaa4842ffe38c2082ce9aa167d26259b4608fd8e80b8 +size 45146 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 56a88783756d..732bcafb92f1 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:3972d351ed5b2902c52808ca03ed2cc04ad8f2297db14e88e797d28c8fc5fe50 -size 36198 +oid sha256:89edc7fe35d04b0d6fb589369691e097b310113b67bd9ff9f7ade4af9debd0a4 +size 44313 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index b853a7eb67bf..beb7e49cd9ed 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:24a5c7266d3de71ee205ed06f64a2b54c1e76b67e03cfe0fe1970ff86c6266a7 -size 34369 +oid sha256:7dd074515960db6e989a70bbe8f62ca04d41e491766cb6bcb88c2a2f6ffcffed +size 41961 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 8378d44bfd10..91ae8c5c44b4 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:d070402244c6e977c9f4d39ad88022269d64fd4752f1cfa015d20f027dc353a6 -size 34811 +oid sha256:14e8c5576173980cbd66cdfa2a079ddbe31a805c1458a84b530f43c60dc03b96 +size 42615 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst index a7902514a24d..5833a1c77197 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:798ee39590cdb47a7284ca2985508465c0ec26cc7d51f48fdda322faccf9c7ce -size 60019 +oid sha256:c904e6d67c15dfff081c2ad216bc074468fdc7ac9125c62d28b1d837dfe7c711 +size 76776 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst new file mode 100644 index 000000000000..0046dd6bb621 --- /dev/null +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:865b03f8d6aef170c0367bfb5b5e1fcb06c40bdd12d627e7565099debd61de09 +size 76940 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 65636d5638d6..de11dc9ebcd2 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:7f94276268a7d8ff38ed4b866e9e15253aa364729f350f41cfc85e4fa2c236b6 -size 58423 +oid sha256:90543ac4418d38f95dfae0ea04a84c18f1018f79dbeaef2f748806a90fd6c63b +size 73762 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 97c60b08dea9..6315d0165a69 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:84546ab399954f12767edbb71a2c5b2e4298dd4888e97951c9ac1cbf9274199d -size 59024 +oid sha256:5bd86d6b978848b40b221c04c1eab002c875294b430756d417ddd5f142558c9d +size 74372 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst index 87ed9bf0a254..e7ba0ff9acbb 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:deb1d1ee96a3abdcec677f42c1b3908b14073012af0329a0e1b69ae4ee2741cb -size 60476 +oid sha256:45653c25102888987722e3402660fe3aaf64846c8cb52f043430768d3d442c2e +size 77418 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst new file mode 100644 index 000000000000..bf1215c25390 --- /dev/null +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a848396bdde6c380f091d754cdebdf7df4f56a94c7b4f39866eaf551e73376a9 +size 77758 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index b9334909cbce..5d26d133becf 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:209d5df8dec6505d283dbefe1add3d260ccff0d2a6641bf6267d5c011c55ca7b -size 59004 +oid sha256:112eb0c92bfb9302dc32b291ee730eabd1b945449617b65be51aa7e2e53d33c1 +size 74533 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 9d2b3743f683..1502367f283d 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x3_rM_splitK3_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:25e2876590101e075706cdf56eba028eedd3498e88710d94fd6775f9cabb1612 -size 59494 +oid sha256:55e81d6cc478925f886892baee5cfdbd19deac89f528610f14572cccbfac026a +size 75107 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst index f73929a67fcc..cfcd5dac2503 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:acf4da3443d77d76602122b319bb4b5a2c4f72f9773b8c87b75643b8a7599abf -size 60375 +oid sha256:6747f9a2fd4230273207e36e8dfb7d5dd77c40b0bcc4e340c253b7f155314213 +size 77126 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst new file mode 100644 index 000000000000..10f49ad0b21b --- /dev/null +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:18233bc64a5b1421e61b8b36953640f49878953a6b58a1cd1647bba70667e4ca +size 77822 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 44240ef03bf3..41e6951e8f4e 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:7c92691f9ea75e83362e0aa8598a108a515641bbd825f50ed50876b729d1fbdc -size 58925 +oid sha256:37b3c4382fe9230e96af673af533d62d4511d640fe6054d196e3dbe207cf077e +size 74573 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 023412709e38..88e5e8bd8339 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512_s5_et128x8_m128x8x64_c1x1x4_rM_splitK4_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:eeae2640698d9aa285f44cade6dcc8821d601c5b8a8b06524f30e250d9a3db32 -size 59380 +oid sha256:ca68494e918d46f4583275c154aee8dc449b22c26b02d44036ebee80760905c9 +size 75153 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f.cubin.tar.zst index 8f03f680065e..6ecf3fbe5396 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:ed261d1a734d5874f8d80bbc0aa76a054d60a53671488d001372c92df33c80ca -size 37597 +oid sha256:68c732872acf11b485361a7e6565010ab8ca05095e1fb4fb4964b0021ac474d4 +size 46296 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst index 164c762d6df1..ff4b1fade7f2 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:ed4edc9b094aa779cc2190943da517dfdb0c85b39b84357f2c7b694fc723ff5b -size 53927 +oid sha256:94ea5cc6b2128a6551b94c83f31c050d096edfd5b315c342b215196ca956231a +size 67420 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst new file mode 100644 index 000000000000..3ecd8615e715 --- /dev/null +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:580bcac7383472ea617527b520b6956d74ce3183b74d9cb0dfb22eb514fc5075 +size 67801 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 67f6a7889eb8..a64326afa589 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:1000f12c9162763964d35914690886f471098dd4ffd86e15694a31c095ea36ea -size 53858 +oid sha256:006e2cb8421ae1fdb2d2133cf4d786789e3bbb7220205884c1f7f2fcb052ae93 +size 67292 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 2a6d22adcadd..8eb6e1b72b47 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:5edb7a6b42901a09ba2b6d5330eb6a306aa30b1943d9cd85d4a00de90f2386c9 -size 52079 +oid sha256:f76bf1f2178f68caf1a2cb62b0c7fa833363a9b003e8822c2f22713272b4b408 +size 64738 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index f9dcd09db231..0c47e3cbdd54 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:4e7dcbb4d052326bf717e3eb4bffb5983fd7daa06dbe92b0dfbfac8fc6d73a4b -size 52750 +oid sha256:e9a35453497e39131950987f1cc1a615d0fc239e51c051fd85fa0a9a3f972a9f +size 65589 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst index 3f0ddf4ff50f..51a2cff2ffcf 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_geGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:401c95ada0df004bc26c157876f0e41fc3fa2935721affb0584532bb3098af24 -size 40461 +oid sha256:6bc2ef8fad32d6ceb2fdb124540727a276e1a2c882976b516b21aece5c0da3dc +size 50170 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst new file mode 100644 index 000000000000..29c75d8bab08 --- /dev/null +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:43633db882b8cfb22e36d724601321ff297446f30722b6afa117f3f82e8fa583 +size 50713 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 06e3457b1ff5..a377c13bb8da 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:51df065ceeb70da317b9a116938e5b6bb28c38eae13262eccbdf95a214485662 -size 40336 +oid sha256:e5a176f68cd5fbbad31d1455f9a5f88cc6b3c830cbd9450b4a5176a8b97ea019 +size 49972 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index edc8fbf520d4..6f78f17cb7b6 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:4c859cfdb53413a1d187c6220eac9551aaf48975c808cf8c69dc7e73eb9d9301 -size 38493 +oid sha256:542a03dc642ade4f6aa200e0d369dac5f22cc071987b1c90befd3439cdd324df +size 47593 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 19434e275abe..2466018f9e35 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E2m1_E2m1E2m1_Fp32_Ab16_Bb16_Cb16_t128x8x512u2_s5_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:d7b4f92e3e317219d7e1817c95b7179892e35a6f9bf288d276399b54b059fabe -size 39070 +oid sha256:6499e1d062942f002e1b9b08914eee704cd1c626d96bfd9a98da61542a7b1f50 +size 48266 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst index c89b2918e39f..fdcff480a896 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:180e67c9eec1adf492cea3415de51a3cbf16c408e0d31009005c9616fde406d6 -size 49389 +oid sha256:ea1a910c33e569beeff4f55ad1eec011172100ac9b3791ca312b84f9a60825a0 +size 61346 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst index f88494e59613..3d76aef1f7b5 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:d513239c2d6b89e39864fcd022b0989738dfcaf0474cb0af08155c874185dffe -size 49455 +oid sha256:0eeb57bfc78102eebcf76d71fc0ec34d4a23b5e53dcfa7ac19e5840717547d5c +size 61254 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst index d54d61c12196..52e42b68114d 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:16d626b50e20cfe7e8043f1949b8c6264ba9790d68abd7204c20af42852e4211 -size 37319 +oid sha256:0ba1d0aa37a63436e7abfdb160302728bb18fb1e6260ce1fe38dc2ef10ae8a0c +size 45609 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst index bb2ad4e2b6e0..08b365e18cc0 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:5ef2ac617ab35fb1ead2aaaee9e844ebe48ec8fab5fc6e4f3428153fff2c8c0e -size 37367 +oid sha256:d012db2b154ddc3712fd45b61b206182c51bc3031cff502fc020ae97c541bdf6 +size 45674 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst index 23e20941a476..a2bf22a72c5f 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:407737f94847f1a1fdd015b3c91444113f488c939ebe33a38a734e49c790045e -size 51562 +oid sha256:f9a9a6543534e91230cada9cd7487bd59d3bc3eccc90634bc4ecf1816b6325a5 +size 64313 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst index d14842f0592f..e4a989e6ae77 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:bbed0a81bb6ff83cea2fa830b32bda37014e2859cc1a099b90869117bb714ba2 -size 51557 +oid sha256:cc4972c2587fd9aa2bf8fd72774791e55bdc3685ee2b1aaebe592241d99a2e25 +size 64414 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst index a704891c5af7..efac896bf96e 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:efa3807d7ec3f896cf8db42981595d25e70033d3d9849eac6838dbd31e00b3cf -size 39651 +oid sha256:2d8cbdbb9d32f8d877399847be954ee7cd5eaeeefd6fbcfa9346b6d350373457 +size 48898 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst index 63733106eae7..699eb48c0c77 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:f853e35f02169226329d2eaa9b54b98a0365244e1af979314fb3726a65c28678 -size 39742 +oid sha256:1621f7fe62317018ffe2965487f85da211d4b32045e9728c2be35cb9d373d59b +size 49036 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst index 73b8caa06b5e..4c941371843f 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:ce0d21704f7cdbf2c503e9cd3fd4b746118dc47c54eba8322da17327eccd0720 -size 52716 +oid sha256:db2ccbdfa18e8e51d6438b86f9220009ad2b94233e932c9592c37d0c8ba70aa4 +size 66095 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst index 2fe6494d42fd..459e8f2f2195 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:cbb094534713fcdd677bdebed76854e7a4a19b487f207f4f867144c30380118d -size 52829 +oid sha256:8732b128b5937772641fc7da19176241ebc944e83821273868abefdbb9df5c24 +size 66282 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst index 02ef665b99ad..3fabd195d6ee 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:8945f50dde192b103bc951e31e09f1b1874aa1ec11edb675efc93c63549e34b0 -size 41255 +oid sha256:bb002fc24ea5421a82cce842baf9906560e90dd5e699830ff42c25d8a12e5d60 +size 51166 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst index 86b45104d9fe..18e437cef67f 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:cd8ca81bc4d6d99b4b56ef741f54fbc927e4927bf29d1b55cd781a63856ddd93 -size 41347 +oid sha256:16278b7603f72f0542aa900123fef1f9ea96d47020668a2f41c466ca7541ac8b +size 51203 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst index f34823dfbb32..6f0f6c478083 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:242d9445c4e8a82d16864abb9161c564012ac726cddb35d0cac5db7aec806437 -size 54998 +oid sha256:04a434cb6c51537a78f40c2509316c7a40ef851a0e3fc49ee330bb3c214e8841 +size 68814 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst index 06344000a915..2aef8bf0f0bc 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:f2f14dd33ae65dfef9f5be6063f6d5e8223552f7d1d7180ffd18aca2ff7e578f -size 55183 +oid sha256:a346aeed9d08c52ce3fe2934f1fd97d26099ffc22a86e80c757e878e14b28272 +size 68991 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst index f03a0575809d..aa64d645ab01 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:7ebdd3c835ef8520084abacab02bf66933e72dcc30d4cfe19cb5906bfa0662f5 -size 43524 +oid sha256:344cd549596b4b7330d85b05173f97dba0f49836d6d7b91b0eea19eae7f5b472 +size 54355 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst index 3f12aad1a837..2ef52ca8845d 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:b06166a75f10208bb50f224e975007cfc7d7b6d347826293726d67276bbe62bd -size 43511 +oid sha256:005b20b0b84d450ec52a3a0460c08eb4449aa888763b36ea134a451ae75c2368 +size 54465 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst index b9373b938f5b..498c76adacf8 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:768cfe41273747363e9a93097e3f2ba4f818244f902cd67ea822169e6b831597 -size 59493 +oid sha256:0c10962a3c1e8171ff55a4142666cc24767ca91488c9f6bfc6df044d2f7574f7 +size 76608 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst index 3f7cb1c5e385..58fd34fcb520 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:dffeda3cb6d36c1629eedb91778b92d12b8488e6d08dde8d516279a0e1ab48fd -size 59651 +oid sha256:3070924e76b9d7445a3ec22275d68f02a5f3c4936764951a4493cc96c93ba933 +size 76765 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst index c17350982403..d38a346a40ff 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:ba33885ce1a6169bcbaa4a27b92c108dff23ff5942837d9aa455187420410369 -size 47965 +oid sha256:70abf5037124b2d440cec73ca6d8bb303ddcc36525f217ff6a2c8d4fb5294cbc +size 60756 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst index 4b32f5264e66..bda8a93a1ad5 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:32019f8c70914c0c3315cdc890c0f357900d1d132f5aff4176171253b06b322d -size 48457 +oid sha256:18b6a022168036d34bbfabbe383c64ef070e5f559c0ab00e691e13e79288f537 +size 61557 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst index 4023cf0c1519..c1bc6ca9a658 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:ae49ae07ed87bc37d31a533c5c234a90a65f01b77212e3d31da9f0f1ee52f1a2 -size 61725 +oid sha256:ee03c0a537c7783c417d73563171990ab2fba71fcc1cac03e8977943da5f3336 +size 79474 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst index a110fdb26c14..339b5b50c41c 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:22bdb882c06a32e41ce3a38174b5d201766fb15381dc2780875d14c53fc6f98c -size 61955 +oid sha256:af67d845c4e5ded37446b612c707704c136f5c321549c566efccd32e045b3199 +size 79753 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst index ea6500f2e60b..419fa5afde1a 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:6e3515ef066764a1633c8c538bed605c38807ebfbbb931dcffc5ea7f04fa8e9b -size 50266 +oid sha256:2f95fffc895d86ef1e8a95be42e4103acfa5f270421eed7b128672f37a6375c7 +size 63719 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst index 0f1e3f72bdc2..40ee6d794a07 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:d823f1a4fa3b7fa3eb574600f3220e71cb100f5caa9813e19cdb66e02518308c -size 50551 +oid sha256:a786499ab7c31ae6fc05deb1d1d580bea9061624180f23aaf55f1c5db66d9dab +size 64380 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst index ddc54806bdcf..1e9e288e9a23 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:7f69adf7a9dbef6af75060d19b9f3cea4fcc2fcc1cf3c6b49a640adfda12d8f0 -size 47238 +oid sha256:887a019b740a38ea70a0bcbaf5526a1e886bd6ceff234fe5d91ffc567c93c400 +size 58239 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst index 56566bfd24a1..29ad49935bfa 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:6bedf80bf436402fc03ab09ddab704b7ddc98df3b0c7a1bd93eb0281fa547789 -size 47289 +oid sha256:86554d4166d88c95a5c8247198c3df9a0cf236c205545b78c15e2f0a362fe0c6 +size 58121 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst index dfbb4c83b212..5aa2f5689011 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:9a6017ec9f0894789a69cd5064dc685abb7b9a7e67e54adf516570a853c334fa -size 35116 +oid sha256:4db43f309c6f169bc831a24c4818433944e58608e0403e736410891703771192 +size 42941 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst index 78a1f73d260b..7ea5c4526cfa 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:dc0b6a5af75e235642a9d88db3df0a8e5be8c6a158e1a03859d702632c0a2b7f -size 35115 +oid sha256:fcb440bf4decb255d87bce5975b46b11690e463824847ae75bcffa408f294005 +size 42834 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst index 074617806042..6ed5f172c0bb 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:7c904cad72ffe64c6e54367c655d938b40727444e4defdfa2cea0e0249b06f66 -size 49344 +oid sha256:a22273da40dec172a913863d581d654ba244c7bea800b4bc5f10e54a7cd6f831 +size 61088 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst index 33878902e3c7..7dc6dcc805ef 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:676aa59aebd7263f0bc0e607ad2f6e198bb48f3df844a68c8891ecc9aa807cdd -size 49374 +oid sha256:b91c6f2bb9296fe5e6ae0b1fd189edde70febd335cdd4ea1f32aea1ebefe3a6e +size 61069 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst index 5d0aebae759c..1528451531ea 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:83f5b08dede053296530f3eca5ff269adcc4471bfaa71cb14edbb41da57b5c5c -size 37578 +oid sha256:c7c5e1da8703d6ca5241f2d69d7b061362c0a39ec10e552d595cdbd578a32475 +size 46105 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst index f06b2b521d2d..6d871279b2c3 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:6e089c001a1b966e56e2017d809dc1840c70c9e064cd52d9aaa522ff42b67716 -size 37520 +oid sha256:747283318882ce6195f71776941a60bde000a47f899aa32166c52a1b78a62c07 +size 46100 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst index 5158ec9ffeca..6aab3a19388d 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:91501eba5ea59c9bde0547dbc70f77b1a426980f363530a8c301777cdb0893e4 -size 57501 +oid sha256:798d2e8fa8454551988b78f5ed63b550dde7f8f72bfe3dbd16ee9c3e58e56d90 +size 72265 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst index 6e181bc7f8a6..a9b0d685a707 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:40cdacb25ad963478a93d27f725c5ce74dd0ee65ad261bbd6c051633acffa7a2 -size 57626 +oid sha256:97e2ad159cf925e5daaa4b23c5e524b3f4027a1bbf57a4c41610587addc5de13 +size 72187 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst index b984f9bc3971..0e10baf5f201 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:06c2f1df5aaabc69df6a1924c6093fafe4f31630546d44fa967a15c7ce68f1c2 -size 44744 +oid sha256:954f3061b8b97c8f2505556037e0c63d013b385eb217335e830a6e024bd24319 +size 55747 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst index 7c551f150c82..f5b562043203 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:1c5053b8d456839ff28f559ccc9be03da7953b83edcb28913a5b3fdc8ad2dc5b -size 44726 +oid sha256:c3871de45dbbbb1a5bd9e3d365ef5737663b2272a8832f8b89058d220ce0609d +size 55586 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst index fcb093a8723e..e049ff14ae2e 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:e4a431cc1674aed1a89e5d98fe218ef8265213c122d567cf2e12bc1f4fe694cc -size 60272 +oid sha256:3bead3a74e4ebb4b8cf7f900502794f21b8bc55a9300672b436edefb21534059 +size 76180 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst index d309dba4f58a..c1db5b61000a 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:140f9d6d524a51e257cddba8f85b71891b2bac244ae775257d0a2b79ef05eb4f -size 60391 +oid sha256:971b99b6d02d1e9ea30a80f04ec38303365a430cc1b025d44412782f38b507f3 +size 76090 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst index 94b45f57d422..57d80f5a18d1 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:4417fbcd9dc536932987fa26d887e9680145994a6b7893b8d797805347d66912 -size 47829 +oid sha256:a12fa62e80dc200cc3ec271f65c8d0b7b168403685de792db1595785ea3c402e +size 59829 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst index 85d0e3f39e81..37009f72e125 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E2m1E4m3_castE4m3_Fp32_Ab32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm103a.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:1f8c6b14b60fc5b0b2ed8c374b1a06b4665fe7b0b58da26f898656402ff7223b -size 47752 +oid sha256:3ff0010d50c0d5ca43b1dfc7a54096530f6869ae44a468e8025894851be72460 +size 59740 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f.cubin.tar.zst index 50c9b97c3a7b..884ae59d13c0 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:c770cb2fc3b56d26fb645688cb67418a7c89482378cf6e0ef676c31558b428c0 -size 60847 +oid sha256:78dd1ec223fe73738874d47f1cbd8d4c4f5c79f887e315244d62a7d3d07c5c24 +size 79214 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst index 33b60200f297..fc99a4fe82b4 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:320093858ecec5d28462d32a1f4893527ba847a783af540dda142007a14732b8 -size 48835 +oid sha256:c312ef3adcac4aadd953750c91f029478515eacfe767450ae13086486f68fefc +size 61059 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst index 0321ae7c9f14..df66d5c23748 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:231b01bf30b88b4864c3945721b5ba10c958a5ccf0602fdb662253d9553c87e8 -size 56720 +oid sha256:26f2a288a44bcbd8d8ad860cf13a8b463d9823833705d9f5c7bf5c32f105375b +size 71748 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128u2_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128u2_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f.cubin.tar.zst index f04403da2533..c8b33be3200c 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128u2_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128u2_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:72c1b3a269363769155ed7f56293793abd6e98ca278a66707d64603c742cbe63 -size 62975 +oid sha256:85120fc604ad52a55ac833b42b05da609ab8e9c8dad2faef1fdacf7a3e8e3fbb +size 81531 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128u2_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128u2_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst index e1538fe18941..89dcab3f539b 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128u2_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128u2_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:ee73d3a65fceb7b9f311cde6559491a025d0526d831b6ae7953df9541eace06b -size 50672 +oid sha256:0699743f01a18894279be217ea1eab5f050a928d03c1186dd5146c5fe418a7d6 +size 63567 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128u2_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128u2_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst index e60bb6e6dd72..37931763a989 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128u2_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x128x128u2_s8_et128x64_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:c863fda20e6d2fa96cfabf8401b93c1ec969f99afb2e82682f4d12db44b29f06 -size 58662 +oid sha256:c2e1b20f7111ae833572705341e93eb673b942b1e1be5093e40db4a97a42502e +size 74278 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 004c42dcd632..c14171cfaf0a 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:820ed3e29b317d93dfd7e966548c301c59864cb7e76a761cbc8bf7f37fe25379 -size 34896 +oid sha256:ce780e31ac964bf3c7e8db10ab11508fc02c64e5afbdde470814125390d4759a +size 42334 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index bed77f9df085..dcc57457ff59 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:33e1cf3492ffc663c16711a36ca235ed3b30f67eae4841e7f80e710638ef1668 -size 32139 +oid sha256:9a5fccc3ff9ae4e60f65d80de01b72d77bfa4980fe9ba44ec5abdc598d179b6c +size 38810 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index e062484c62c9..5c6e0209d62b 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:4d254c32dd877d6ebf67b02c670bb5cfba1dab22d3aa1c76bd09c369e461a9fc -size 33197 +oid sha256:83486d28541aac366559023f200dee4a66f757c4a26a5af4afa79593bf506ab5 +size 40279 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 58d44ad597af..98f7d3d2de4e 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:9000ff2e48af5f4e459cf6c364f6d3f9f4a9e68778ba417c1e23d5eb163877d4 -size 24835 +oid sha256:1de3165baee6c86a3073262242893a4a2f398665620860e168fb9f05be349ede +size 30002 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index b9629ad24ec8..81410baa00fb 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:0d7c99fbe2edde9d185098b820ef94bc57441741392c6be019c60982d5800b56 -size 22487 +oid sha256:202ae79c9d78e7b7cee2369ca344309471e13e1957a75c2397534882aa5dc450 +size 26881 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 4f0b30c25f40..35f790c8dea7 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:913687777e0643044d6d60fa30388f765d1400823209da92fbee6a84a9a03b95 -size 23553 +oid sha256:b3fffbfcd8decee0fe6da040b10beec34f757e7773a77b45ddaad55febaeeb18 +size 28272 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 6b9bf78d591a..7e2226f36559 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:f2b6fd56eacb023d06283f4547ed8194375afacb8f17fb0d7564ccdbcdcdaf50 -size 37294 +oid sha256:5d3ebbb3815d9abc8aaaebf07409910a7c13dbae18154b93d35ec6680999d1c1 +size 45619 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 91948572a09d..5639b9d26bad 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:0ffc602d74c26bfbe784fb20c359957507e5dcdd0d301a12001ba401eda781ac -size 34678 +oid sha256:4ac45d140b8a3c27e347a03ae264494c55c5a3a6bfb8b6874b155203dfd6acd1 +size 42249 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 23e58ad1c49f..0243ea4a4d46 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:8764858e8791f6c79d3fff5dd929c8d5b29d02ba9030a4a8f3c5ef3f74857a45 -size 35886 +oid sha256:e27bd0c195643d4b1f73ad332a5261cd2677056f30bd0211285d4efd9bde9786 +size 43685 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 63e7a7d88faf..6303ab1e3830 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:c2ef6ab896481cec1696bd4c70b3a740029cdcb465feb5d08019d366fa2a8a0d -size 27335 +oid sha256:9f68a1c3773f7ae75112f3b6181cec33e48d9b007672c917ab92ef0c2fa6743d +size 33048 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index a2367553e4b0..b3040b3d6460 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:9c8b278b4fe8622de149eb79ebd444c4ffd10bde39ea5f005683c5e693c44553 -size 24869 +oid sha256:81b9104f67ee6c2b7e3c177c5ba00987e78cb0d1b10b9d3a214cb104ca93b26e +size 29916 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index c18be0abeb0b..0d2005561024 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x16x256u2_s6_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:1a74585c7f24c40ba5dfd8d2348aa6b57ca06d261d558c1328a455075434062d -size 26072 +oid sha256:3ba4f034984085dc8c4b94650cf2a0e052880f5933c9bf4058f6f3ca7d0a62ed +size 31448 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f.cubin.tar.zst index 4cd6f7fe295f..650c3cc94527 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:a78c710a449525cc425c2ee9027fd014f31ccdc29bbccaacf63bb85f4de23e67 -size 68766 +oid sha256:be279513caf728904d9c9aef9a523d0004b346767de9a0cf104e640247ec32bc +size 89688 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst index 9786175d673f..10adceb22bdd 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:21a15b8804498600077eb99f1171fe22baa03e830e78b0ed3b5580a77d37bf1d -size 53223 +oid sha256:56b41540e941bb3d8284b6c21cd634ff98f34c2ef1511a0fc2875facb2c59fb4 +size 68117 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst index 5fe9c58627e1..8eb3726e071f 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:baecadab3f6925ca9e316fa2e798adfaef402d6dcc8929fd6419504ec933f1b2 -size 60820 +oid sha256:b76d6f3d4347576ee6bd905da8f37a0841aa4abfd7bee54d6b037b93b4403735 +size 78117 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f.cubin.tar.zst index b9b3f1c8d48e..1f6c8918131d 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:54512b47f6a1c476a3f90447bc2a5f60065c3b1dedca605c8a5db9125ba7decb -size 70593 +oid sha256:85ae80f68e83ec2b961a227072b6f3b97b63889e083da2f080f6becf737551aa +size 92283 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst index ebb452820aae..ae32450238cb 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:e646973edd99605c03815bf232565be0e1dd3ebe28151fd18ead3ef5a888dca8 -size 55128 +oid sha256:a71892410e03ebe0dac25e809ad002af479b3e741cd2d22635aae123a052beed +size 70298 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst index f7404e3efcb7..6ac1663b5372 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x192x128u2_s7_et128x32_m256x192x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:3e428c480307e29cfae09c171749cebf86814b6ee0d6b5d5eff1355faf2d6e81 -size 63003 +oid sha256:cfd74c0cc4a18350ffbf7529183ecd869b1f0087a62fb835be1e335d5b729f7d +size 80418 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f.cubin.tar.zst index 0dfe13cca20d..aadf29a89804 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:a8a1ee6aae038f3b8564fd657499adc8dcd2228595daec772e4337c3ccc57d90 -size 60301 +oid sha256:02494ddf80afef26b673d6e52b1669a72b7dc9a470b16f0cf460ebe4baf89f08 +size 78211 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst index 55e61cd06896..0e1e08408da8 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:50bd7e45a6e30913b1648fb3962d6dea8b2c1525f3b16a557ffc4daf7fbc1dea -size 49252 +oid sha256:c172a7138e97046cf15971934fb31a468e6ac8a438a8880d568293084ca25281 +size 61938 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst index 89735a135fd9..761b959251d3 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:e1ebf34ffd6cb9d042d828cbb725f8d7d07c54e675feac805b7ee002237bd937 -size 58528 +oid sha256:2b9e5191c71ce41e387add0dc29186e52bfbe2850a7067bf060de403ab68530e +size 74160 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128u2_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128u2_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f.cubin.tar.zst index 6f1cfa925260..1ece45883559 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128u2_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128u2_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_eW8_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:fc974d06f541c475f3512b65bed9f7de4fe3b7b60c66d8db514dd30462bb7059 -size 62052 +oid sha256:cf57335f834946bb9507faf4007e936d903ca50a6a5b66ca0da550d0dc05f773 +size 80538 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128u2_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128u2_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst index 0e7c3bbc7050..15b7f5042ca7 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128u2_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128u2_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:11f109b6eb58124cc1838b4d8609718c247aa5b4995bba8f1c48147f98656968 -size 50971 +oid sha256:2a7aec62b81f2cc68c34dcfa30e643d2cae5071c7cf82c05789b50d4111d38a9 +size 64356 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128u2_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128u2_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst index 0494a6842201..6fb9b087c088 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128u2_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x256x128u2_s6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_eW8_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:722d9966822f2168d212a56ecb1437bc1cfb7bc0239a7d4d6edaddce37802fa9 -size 60346 +oid sha256:5ff1dc4de3d886119ae11673ed77606739cb23dc43ad0cd02de412780ad46574 +size 76551 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 0303f63d7609..e7dc4058b9c8 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:6ed4adbab5d81fccd8969d029f220528947ec26274dc24715649372e15b78be6 -size 38829 +oid sha256:fe7e1a3ed55b8b8255656ef5734f6ba7125fc587408558c0d244570d146c6eea +size 47871 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index ec72fb2ac90f..9d8647544098 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:abba265e5facdf4a0d90ae478b628abc1f8ad3550d6ed68cf9fb27496b2415ae -size 34759 +oid sha256:80696aef1d6d74955b898d73cc2dae366457c1c5f116320030bc145205e3015f +size 42414 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 8cfbac20263c..8962bdc968dd 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:467d14b3c8751bbadec5d3b684dd6cc1da30bf851ddcd0e4a0a17a7901ce9bd8 -size 36922 +oid sha256:55bc15bf93831c985dab1ec59cb7f0031651e8ef798663b5c8d0d551c98c91cb +size 45145 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index e4c5be1b04fa..2c64b5e41d60 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:d8de28892fadb41d62db25280d851cdafbcfd6bb5487aa2ced25f3a28a88a9bf -size 28917 +oid sha256:93642b15e38b9e9e045284f84eec3c31bf4d1b459bb8cf8163d63102c1111031 +size 35773 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index f5e69c4c0e6e..9cf0d8188d5f 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:e5c39972758b8bd549f5a486592930ad9b49b82087dcd8eaf7192b1ac2059905 -size 25230 +oid sha256:6a1022c4c9be3ab60078a17b7227f0085cad55fe846a11f07e3b9fd8b254d98a +size 30475 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 69d0dceacf97..43965fa197f9 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:d7ea5b89db1b6ba4733d794d10a7e8f1f2b9bd2c1315bdbfea7ae4fc8bc9b25b -size 27330 +oid sha256:494af95fad99a76808ce92ad2bf0406da0182f43aec33a6e669e75326721541c +size 33265 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 611a72708d0a..3b32f02a9d65 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:e610edbb21361049d76ee4aa6e9b592574d3e0abbc0399ba6232968bfe40fca3 -size 41154 +oid sha256:84f7f9e59c057c5afe093f4d655b94d450bf393f7252895713fab88cd782a728 +size 51286 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 6873edd1a222..fb4fb68a7434 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:c1bcfd03a025b2350ca5a8b1eb816e0a9848db010dca40b2a2e2a5d1878ad15f -size 37392 +oid sha256:da68202512e2d7b5fdea9cc5fceaa45c672819827b9a290943b24bb6fe9b785b +size 45800 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index eaa527374c9b..01f277b8bceb 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:50ff4dd8808247667d5b6c16bb165971a2a55bf0fc0b3a2177609b1ae2d733a5 -size 39563 +oid sha256:72ef496f15fc3702b091ba2f1dedabb3b16ed7b4318130ea51a0a58d627af158 +size 48694 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index f415e032d02e..3f06ec0cdc86 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:8efaf5c80fc10561d7aeb93ae169da0c1a38cbdd383ba563035889922e64cb30 -size 31303 +oid sha256:e0ebb9e2715bb368fb4b4e62569d44bf4ede292a1f52366488940b6b2877dd23 +size 38788 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 0c0413944602..8847f8f28f97 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:7742a27af5e2f0afe2133c32dd55f0ddd1b1c27e19fd5eb8e641bbe392845544 -size 27698 +oid sha256:90ff17fb9f4de3cece5d6e569a5e8f3db4f23bec1913be171a6d61e851f31f16 +size 33840 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 702f482cd2eb..38147c8b7fe3 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:0f71558135906f8fad77c90822cbd97d79a872943a7f790357864a4e0c71e6e5 -size 29880 +oid sha256:15849c363dfc758532856074a6557de167d719e9931b41c3c4b2f99d114bc136 +size 36533 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f.cubin.tar.zst index 950b8d35debd..1cce5ee1a3f7 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:1abe9a243da7f747000487cfc51eeedbc98826317a5d1deeb1f442601c689bbc -size 48910 +oid sha256:b71d986ccf623ddcf83373879b020ff450901fee1b58dca04a00519a8256cd3f +size 61507 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst index 28998846f129..61372888953e 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:6795c4e733cc21d057d0437544952df8e4128f758653e2edb2b5d22831bfdfe4 -size 41904 +oid sha256:7818f130cc1f817211d5208b1c0f3c348b0db03d8d949d5aad0a276163258667 +size 51749 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst index 9d85f1586741..2a595f08f1fb 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:7c33b604ffd199977e54fadff6c3807b335415bd3101762594c709555f64a409 -size 46291 +oid sha256:b6a2fe98689d494f97551bc0f06d28a126e5fea5fa1cbfbc08d628b99fa800f4 +size 57578 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256u2_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256u2_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f.cubin.tar.zst index c3af2cd186ea..190d9c2cdebb 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256u2_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256u2_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_lbW8_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:639110f343124ae493ca797c73074bd7bcc2f6ea4a69c816ff2b180681967caf -size 51163 +oid sha256:7cbd4f139918637bb5848e9f7ee4b2bb9f04d2d129ea12c5b866774401dc15eb +size 64938 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256u2_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256u2_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst index b2dc4f30c1b5..e2bb68d72267 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256u2_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256u2_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:9b4400fe2db9d6aa0185734077737c8e42af4c40ae5b1de8a3c3f095b1f4311e -size 44321 +oid sha256:57956ca73565319531fce87896e7c08489ac1c3f3405b92771d7193e07c12707 +size 54780 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256u2_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256u2_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst index d3c5cab77146..2d48d266886b 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256u2_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x64x256u2_s5_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_lbW8_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:443cf674001766bbf9ba5a3c4533716364fd8fdd255cac5c2faf37a343241bb7 -size 48787 +oid sha256:828c1ae598646e80495c401bd33287caa3b18f5df88227077e14a23f86669ee2 +size 60021 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index ef7f4bee4f1b..2a8cdbfff9ca 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:bfdd10653e425347f62044917f1ec6d49811c0b5a4f24ee592477c959c08e746 -size 33086 +oid sha256:e12d7388ee5370d3e7a65528585573666f1fff8f914d967344d0164805d19ed9 +size 39837 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 774ab7fe0c89..e2f33cee05e0 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:cf321ee3b4570ad41944daaf83f8716f808f8f90c93a5eb6e195587090d79713 -size 31309 +oid sha256:ed73b6d486e44b157283eb7f04622e465253a8bc959cc99530550d99f1e095a0 +size 37613 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index a415e25d67cc..5fe786d24714 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:913ad2f59866aa17bf0684c3d30553fa4f1bd963e1d6ebed4e4120816ee2c609 -size 31796 +oid sha256:bf8c52e91f0326e16f6b3528cf1170b00210452e48428a1d6c31a63452e48908 +size 38312 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index bf0c2a65ecac..77acee57b268 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:1990ee95c9fcf3d2af852e3a0cdd066f5f7f01368479db9fb837f6cb7517f714 -size 22396 +oid sha256:19bc20e62926db8cd849200b036555c1c5a0b301a39c1664d57dcf4f6f056d46 +size 26750 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 1fa581acf2f7..a0ca9320fad3 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:4156f6e9bcf3b54e658267cdcefe0818757dc65fee915f0e8984567f8a301d4f -size 20579 +oid sha256:0fdaa312d48d6389f2602b2c5e4cba10a47b6933ae32708a5bf71e40e6634c1a +size 24557 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index c167aa312844..145e2a5945c1 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:25e51f5fc4b9b76fb93edc9616868c1c5caa6af1f083fe4792070eba9ab6b1dc -size 21294 +oid sha256:19d095a9ae03f05000342b1bac8ecb4483f02dd700e44fa454671ba19e179ac7 +size 25448 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 900dabe30803..02b3c36462cc 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:536b0511225e6a83c957b78ad83a6735258327e91e6056a41ed4eb8e1ce762d6 -size 35460 +oid sha256:f72a6c2868ac8be70381a8ee257aab75dd0a1454bcc9e92bb751d0cad4ee8602 +size 42927 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 89df66772fdf..4a175f7ff9ad 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:528fe92d65e0cde11e109bd0441f0880025856c485d229df5c1ef97277878171 -size 33563 +oid sha256:4dd0b1851b2eec6b29ca935f2f34946d59a04199b7d03be9227ead0f51f4ddb6 +size 40522 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index e15a66478fea..0fcd61b812e0 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:96f197988eb8206661302186fe61c09b7ee69b3ae6a6216d9fa5c9be6b4ba1dd -size 34097 +oid sha256:15b6ccc7a1f179b7847236c7da2a2a5c3b08704239643546b4d3a2839524aeaa +size 41378 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index c4910370d72d..37e63bd3dd9e 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_tmaSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:a7ec882681476603569d329f28850fac201b376e75364b19bd6ad88cbd10b8de -size 24633 +oid sha256:127426854db7f9e6bd9e69d4441cded68cefc5a2760b26b2c09a21221e795544 +size 29681 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 68d8e72dbcf0..e692d0fc5643 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:5a2649b06be114b33ed6a9ee820ede50f1d4abde5456ce0fb8aa2072aa4f3174 -size 22835 +oid sha256:0d2d452d63371b99f637941af017991f2b6be7432fd9a330238f2ef20daec57f +size 27459 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst index 42c1e14594ae..1a37f37c6850 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_BtokBfloat16_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_silu_bN_tma_tmaSf_rgTma_clmp_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:ec3623c39477c7b4e6461c4578d2c088468dd677fc9179caf2cd85d3b7e5f80d -size 23530 +oid sha256:bda4c1c2568e015fa7fd853cc2ab47148b6e9100734f31625ad91a03de7a46f6 +size 28240 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst index 8b68f1fd5ff1..4009bf319815 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:9182d47ea47a4e75089cfccd7261b8d0e5741d04d618b63e3781f236b1d6e866 -size 92493 +oid sha256:acce2829f92b10f3f285beb13b29be02016dad3d6a608dafd99e8bad144ba1a6 +size 120663 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst index 35b8b6dfef33..5b9f8d0de97f 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x128x128_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:d739bc56b2e0a03d63f5f4c48b5f34528b3eba8501c21b32051a8cc403fe6367 -size 78865 +oid sha256:65db36e50eb6ab388ff4a7af6f9e4fe8ec471f6f619515fe231dacb0ca056557 +size 104067 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst index b7aa7ff23298..1a26bf841d12 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:593a961224774d9d511b8df883ff238ad2a01a1aa66afaed52de21baa87a21f3 -size 95786 +oid sha256:4b549efd2f6edae1cef989aaaa821ea771bf2fec35f5d23a800f1ed64007a4b0 +size 124017 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst index 531d86e6a711..98909c117a35 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x128x128u2_s5_et64x128_m64x128x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:1b413edd55642e60d116d498c3dd7aef1ae6d6544e1ff65b1e2351f1ebf41688 -size 81561 +oid sha256:138cfd66a72fc1baaed56eee22973b2e373b81a8ce8fb2261140e7b7a68aaca0 +size 107898 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst index 456d7f8ccb6a..1c9bf58f91bc 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:df53d101bb379263b8a163dfa44ffde272323cf54521e70bbf530f0c2ec8234f -size 50202 +oid sha256:56672d57e5b3e700c1432cc5d8f6cebb2c381cb73714e561d309c59a6d68999a +size 61976 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst index 1f77a385aefb..10cbed5faa46 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x16x128_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:da0c248419b7f827bcd5cc17f6af72952bd60cf671cb348ba0bb032af3b417e8 -size 38797 +oid sha256:f3b8adcd3a595a7fc4d2444f8ccde01ae023274118f1cfe515e8cbef6a506e08 +size 47617 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst index fbfd480ed7bc..54058f8dd71f 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:f8466b71a395969a139f0e357676b203c77e6ce924add551c61c55b6ad6f149c -size 53250 +oid sha256:e5ebe4495ec5218d5db59af9cbb370130406eee48181199aba1264cd69ef8431 +size 66451 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst index 25cc6b269450..75ed744ee006 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x16x128u2_s6_et64x16_m64x16x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:c8816b432404375ece2e6e482f3af51d269f19d8139c5adf701ee3fc0fd72029 -size 41714 +oid sha256:e498cd2829d186e612b759bba5a523c8f17e3067e67d5fe1e83823b00871cd93 +size 51194 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst index 46540d2a443a..9c9c1a8b05f6 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:4344c61cdb81793401e4b3f539d65e41170b99fed748f51b4654d73707fa7609 -size 57176 +oid sha256:5a039d5a1317b0eb112c457814da01e0ad4762bc11159c72cd374f702b794e2b +size 71696 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst index b2b0fb66ff4a..986ad330e246 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x32x128_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:231665dee8b6b917dfceac8a0eae6442bb310757cd1a0f47723ddc1fce09fa28 -size 46130 +oid sha256:d0a1ad871bd32a097540c87bb0348e31724e3bbdd890c4c24815a9f24a7252d0 +size 57694 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst index b5bbe09fcf62..20292bee5fe4 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:e12e43c19822f74955641ccd0cb0d399f70d37b82a491a2c277a8d0db009e02a -size 60419 +oid sha256:3b5befe8faa07f08ccd991540b72a3fb99cdf09e1383841d11495221ef4e0a41 +size 76452 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst index d4a982a40ef8..cf2a870d5d9d 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x32x128u2_s6_et64x32_m64x32x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:e5e0e66e2e7eee86d2cca8353f30679d121aed4d8a0faad68447dd1e16a141ad -size 48723 +oid sha256:bb43da00af1af15d098899b73c0753ee05a96070d1b3dbcd795b67b63af9e0bf +size 61078 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst index f394c589ad49..9b7154c75e54 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:ed7d142dd7128627339ae1f44602ea51b34bd62a6aaa9cccb7fe08ebe1da786a -size 68433 +oid sha256:26a0f11e20ec2584f2a4d42c63f4d3e794d9473a281b3f1c53d8507741b3091c +size 87095 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst index b3bb231683aa..c2c9e4e3d87d 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x64x128_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:522584de50b8a23440a45883d156eef0ce111adc9871561520c4b32358ad66ef -size 56473 +oid sha256:8e603a2b8400f3497128958eb9bcc3193ea4b063e9e4e354a871734924f27bde +size 72590 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst index 7a50b8f31893..079cef6a3c89 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:de6d5244a8a64fc6108ac315ad997e22dacf191a83e29009d0ec0483e8745492 -size 71669 +oid sha256:270e6b1789c0ac0b62ef0f13ae0bb65d50141bba372a628eb5825b1ae23eaab8 +size 91035 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst index 8351c7448466..eb388c781c39 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x64x128u2_s6_et64x64_m64x64x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW4_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:5ec3c7ac5b97d9d0ae1129179e37378fd30822dba230621f57d3f79cb4da6a60 -size 59195 +oid sha256:ab5633b63c5eedfa4105ff8d36094e845d6eb45a7847ee663f0313a8d712b9d5 +size 76685 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst index 43e7fed65f2f..0a376238f7e7 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:714b744e624ffdc102ffc1afdf61d25c7a1143ad15a97184f823d0b59575d1d5 -size 33080 +oid sha256:94d78f35b46622961d80f5cd3541e0b6b9ea56963c9afe01a9ed404e5d404559 +size 40398 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst index a5650ce3c140..58edf0416852 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:aa03fb09c366dfa0ecf5d492a032b69a10e3e8b99411809f5d99fdf37a67c492 -size 46440 +oid sha256:88cb055f88e92beddd72ddb6edf252c76b05531c508e1964c004f9e0a079c487 +size 57136 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst index f94ed165f2f1..9d5e59ba4a2e 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:f5e5122576af1624acb3e91440d035635beb84aa21fe12401c640b79ad6d3e82 -size 34375 +oid sha256:5435f560d761677337344740b1f3e16a5f7a5277c980cf618e0b5f4aa2e87e54 +size 41698 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128u2_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128u2_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst index 57930adab7a5..ca6d6b3440d9 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128u2_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128u2_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:67c3c6991f7a1269d38a10a8b5769366f5eee5d47903f3c95ff0a5e0af557d38 -size 35635 +oid sha256:ea94e4d118bf51b2289f2910b26e6157c67a9b4e94c45752324cb573eecf8887 +size 43557 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst index 5d54970b04eb..edaab75b4f71 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schPd4x2x2x3_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:71e742707b6ea8999a720e84166e581a8dd2281d392820f32a392832efe75fe6 -size 49786 +oid sha256:71a6f88070e058a45cb42519aaf1e6aacb3dcbbc79a3e15e1328423eb397fe97 +size 61382 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst index bcb46b0d6606..54f04093478e 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x8x128u2_s8_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_tma_tmaSf_rgTma_clmp_lbW2_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:d28ed6a96fa9c34365c0a9b6b08b5b21281e52cb035ba31267e85ed139f8b47e -size 36973 +oid sha256:e12b0bf83a80f117b6ffcd8f1222e9ea18da25f7e5766616f139ee9d9517157c +size 45074 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst index e5223bce3bd5..053fb9264590 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:024ce0efdbe7a7a50ba4129dd50a9289a8d8b6108ee950212d5f3a224418cac2 -size 21331 +oid sha256:274f737dab81fbdd50d0e842a8e68ab2ccf91cc36ab745312b1405f05e032adb +size 25261 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f.cubin.tar.zst index 3398cd3d32d8..6aaf089f3b42 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:d68b97071957dcd913d52d3abf1d598ecf729a05ac5f0ae5255b23023c71af46 -size 21809 +oid sha256:9ca288dc4b018b1cbda71a2c5992def9c6fc4d31482897d0460447cfde344bb4 +size 25778 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst index ff7fbe00455f..2892ddfb8c36 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:5d45f66d7c24d0735aea327c4774cff5349123e3154a459597c640191fe5565f -size 24866 +oid sha256:e5f781055ba100cf31822b57290eaffee6c030c88974e87342ef6f968b996686 +size 29772 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f.cubin.tar.zst index 9bb47739cf7e..e38ed380499e 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:e23b5bebf5ff620a8e626ba70fe59c95200933412974c3da4a68f9ae5f881ca5 -size 25418 +oid sha256:483dc68f773d6886713d85227e3b7de24c7f273392722d18171e82fa1a83f42a +size 30431 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 15e02c876f9b..dd30e9944d8a 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:5202253a3114b3ceae488e7d9707a02091e51d03935b14860ccac09d42a3cb99 -size 44654 +oid sha256:5bbbab4d400d6aa686eb4d0ff9baeb1125e5ec9a9c3708f271ff16b516ed1fb1 +size 54640 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index ed5f012c3283..d432559b6546 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:d51bc424121245c2d9ed34a4471274aad8a813c0ab8dfbe50f41f3391ad1bbbf -size 32326 +oid sha256:b935469804f18dbe2e596c31cd15614461b7c3e8bc89ed15c42335c8023faed7 +size 39139 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index cb2f1f6ff983..146f7608800d 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:3df4d045ec1fd91da6e291e6d8f5d3ffe20f1bccde1ff07d82ea14628df3d5e5 -size 48133 +oid sha256:a2d462f5dc0fb301e68f68298463e752032812ed4d14485a4f0a65a4b14427dd +size 58932 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 311262cfac62..1edaf5aa905b 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:ae7124206f9ccb4415ae195fceddfda791e4e5ad850b676d93acda528a18c7ff -size 36017 +oid sha256:05714de4f2b5fd162b34599f52526c68101af49723280460db3c2a05194474ea +size 43747 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 7d2b38271b39..890c433d24cc 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:b18159abaee52bdd20c3cb81b8a9d7683273744f4798c9a637408e1c89bc4346 -size 48503 +oid sha256:1baeab9295c0ed3a558e0c8e7049debbfd507a51ed8d0de463edd289721420ac +size 59704 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 2ea2ada53072..4af64bcf9d0c 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:ff065ff7e31b8c02faaaa1a0146f9d08f36c22ffab14a5894f6f14fa258a774d -size 36097 +oid sha256:53878d6482dd450241ceae5fa17870cd5ff7dd3cec48bdc63f49083bcde21e97 +size 44358 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index ce516e83a1c3..e4265fefe168 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:c64ceffa1a051d6b5a5802825e1cff1fbe60862ce24afbf94af87921d81762ce -size 51824 +oid sha256:a7ffdb1f03ae1a8cb3460c20c5fd97d8a325d5dbeccaac78cb036c4ed5a83502 +size 64412 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index dfcac141ce5d..0b2a495feb64 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:8d6351d78d7f634d53290946d780159d43db1d34fbcfed36a8436c82485f3e3c -size 39917 +oid sha256:f8e0683184ecb0cf2436d8cbbc60a0dabc9f95d2eb38ead39924a23dc9798560 +size 49152 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index d0fb7cfca89e..53fca582d5d4 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:d5db29f0eec69753ee7d360eb72207290bfcb25cadf05e558956185bedecfd6f -size 54746 +oid sha256:c1d975aba0dd86e47854f37685aa26038f577a331f97c9ccbccb62929841ddd5 +size 70353 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index d10dc4b9b576..f45d02a34d77 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:c852ea2538802cf5fa7b60230494bd9d58ef64a1483f4074f24905d58f525272 -size 43128 +oid sha256:ed86798fc66dcc6b24e1696b0e4e1404c6f6a654b38c3ce5e27956f5c0878ad9 +size 54545 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index f68a67609cb0..8efd613c6558 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:88fff557e069e451967a2f544a4d17ae6d0058ec8fd6b4e508f085dab72bbc6e -size 58428 +oid sha256:6ffec6357516bf766c38c7f5eb13bd6baefe085eea28653ffb8c232b49c0c78b +size 74756 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 39f0e653e20b..765513abfa46 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x64x256u2_s4_et128x64_m128x64x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:cd1f3bb4d74110d17ea815171976f1511987d5dcce2595e5367b663fbc9855bd -size 46752 +oid sha256:a14900cd333fef7c32cdd58367e3c3de12ff95eede4021fcf617cb8700791dfe +size 59115 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index a6bf24983aae..17076a86f3d7 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:032387376bebf2b09c4fa196167e67f7bc3ca9719c7aa24adfcf8858bad19a05 -size 42434 +oid sha256:19fbf81f31a1c766adb5a7815f56fa2292b66100784ac602cbcde2c10e812527 +size 51650 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index d7d414c36be3..77d5faa31b58 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:9b0f6f8fd5e26be131fb97c0323651a915e7866f892f4f81a2a32ec33f0879ad -size 30367 +oid sha256:4f11ac82a96abfceed845b4c1e3e199ba608a41054cc5de51d9802dc9cfe83b7 +size 36415 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 158cc7fc8e39..292bba7bf629 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:f06ad33869ab7c1fa9ec046c5b4ac867274b2fdfafc9b307655da1895088396a -size 46352 +oid sha256:05548e3f37a5fc43b6a87e3b2cd28141388ab06fd6d28ebfb254bea3aacd100a +size 56734 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 9ab2e0b341a2..dff32880a87b 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x256u2_s5_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:d7b4b362c14044dec145b5f4913bf5a51c4fbfaf64033e70c26fbf5f066ddd0c -size 33755 +oid sha256:af890ef49990699a47965925a215e333e98131625acbdbd742afc0359c23b775 +size 40777 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 405281081357..cdd870d8b245 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:19a04d7565ee21c41a5f0ac24a43648504319021320902aa1aaeaed0ebf9fe11 -size 44505 +oid sha256:bea7c0ef5139f8ecc745dd6f7df4567ffa9cc16931acd54c8227200add47b7c3 +size 54453 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index dc36a8223311..1d020f62c7da 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:eeb1da383d722e2fb476a7c5177c70b3fa99a41fbfa9ed78091ecabba3d51071 -size 32332 +oid sha256:ac374d6b149fd1a4e73e28b82487e204bb4871b731f3ba2f9515fe0109d47289 +size 39099 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 6e6a8137790e..c3bc8b2a8e45 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:6fb50f40ba94b353c75877a1f12a0603ee05737f0e34f6936a619ced991ed860 -size 48913 +oid sha256:10a2f9946f7399ae0aec37a5e99f01312c86366f099eb73dbd40450903fdd980 +size 60537 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 8ba598687f84..fc769865668d 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_E4m3_MxE2m1E4m3_castMxE4m3_Fp32_Ab32_Bb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:e0ed6c6f23a76597ce5c698e0704c41c94598cbeff6cd5fc6d1b959ac22fd051 -size 37445 +oid sha256:b46ea3f0cbb69428f96284efc75bdf683eedf3b81f932502794c6b75f5e07ebb +size 45610 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Fp16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Fp16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f.cubin.tar.zst index feac1ad373ee..1602d1d1c926 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Fp16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Fp16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:1292d91f20b0d4f8409af2ae9364efe3a7dde09210908c24c9f54d24ad72855e -size 31828 +oid sha256:22627bc0509a5d4de88d542a6d9d07d3aa4a54e074d06a1f9a07be989dc32c0f +size 38666 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Fp16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Fp16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f.cubin.tar.zst index 5c824412068d..f5dcaa0b5f0e 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Fp16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Fp16_E2m1E2m1_Fp32_Ab16_Bb16_t128x8x512u2_s4_et128x8_m128x8x64_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_rgTma_clmp_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:a4a851763eec47cc6ce9496e6eb0b83e58a3bbe914bb943ed908b653e68cb0f9 -size 35682 +oid sha256:9473f2df1d7abca3ee26e9d7d107bf83f5e46b8ebf669196ab912637e09692a0 +size 43836 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Fp16_E4m3E4m3_Fp32_t128x8x128_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Fp16_E4m3E4m3_Fp32_t128x8x128_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst index 86b8349e67a7..7132c507028e 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Fp16_E4m3E4m3_Fp32_t128x8x128_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Fp16_E4m3E4m3_Fp32_t128x8x128_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:a4dc00383510c496d2a6b7402062556d81e48415252f4bf0871bcefb2e6057c6 -size 30053 +oid sha256:920e39546cb577ba05f45ef8d16ab7fe422c90acdfb640ff88d53e73c435682b +size 36187 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Fp16_E4m3E4m3_Fp32_t128x8x128u2_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Fp16_E4m3E4m3_Fp32_t128x8x128u2_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst index df4c805e92b5..b22586296dbb 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Fp16_E4m3E4m3_Fp32_t128x8x128u2_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Fp16_E4m3E4m3_Fp32_t128x8x128u2_s4_et64x8_m64x8x32_c1x1x1_rM_TN_transOut_noShfl_dsFp8_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:6e7756ea6466848062163ee07a9fe49ec5d56e1c301f7c15da860e08c624c690 -size 32546 +oid sha256:43db3610aa110634140785a8d32f7060f9b04558c042bb575b05e6f492499c4d +size 39415 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Fp16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Fp16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst index d4a80bd67238..1cde280ff466 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Fp16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Fp16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:ded4f02645f73e78ae8213f15be2c9f66d9a8f5f83737e49985ac38f59c51f0a -size 21512 +oid sha256:407a79c828541f0eb67204d13e611057e6108fa0c5d07fc6a303b0a35009562a +size 25514 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Fp16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Fp16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f.cubin.tar.zst index dd8479bee8b9..141cf6c5ca1c 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Fp16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Fp16_E4m3E4m3_Fp32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:dd6675b226060968f69b188c1f9c2bf4b99b74a4a3774dfe57d9fe7bb8876b1a -size 22036 +oid sha256:70c07bbf93a2441eb6ed1dc7374576e800903df9c520f180ddf113a09db6f85f +size 26132 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Fp16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Fp16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst index 311001b22837..a2981de8005e 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Fp16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Fp16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_rgTma_clmp_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:52909062183c21f72f9fc21f519b4968bc849093d42518c682b71fa5ac9be1dc -size 25073 +oid sha256:ff1ae3ae7d68dae585ab10f0c4fec109e5ec26a78b52dd93fc67160735a3765e +size 30046 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Fp16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Fp16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f.cubin.tar.zst index dd0eed555764..26718c294e17 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Fp16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_Fp16_E4m3E4m3_Fp32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_relu2_bN_rgTma_clmp_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:8edb13ac7730d80a0cd8e480d65262bacca4b6c63a42f0c206a0a9d9f9bd7c69 -size 25672 +oid sha256:8ce96961d7207333376cd5920f32297c088601df9f42ef5967c722f6e6dbcee0 +size 30722 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst index bdbe3a3391ca..e4ea60fc26e4 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:e931e93b4ac79d1d338bef933f6db2bf37c21feca456b8ca0649bf95ec7bbd7d -size 85176 +oid sha256:d86359d51a7b930777ea3cd1f1ca26a38e969a52cee400939cca1d59e45e9783 +size 113279 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst index 059bfd7bfa91..752088986745 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:d4d99d08259e9e43a59500d7ed3c5e6cb3229d54a3379c23903db4e941bb274e -size 85660 +oid sha256:fd7df6a792ab2faf4724f2145d878ca480730fda67a3d1f5459d46ec35c7ec62 +size 112325 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7x7x7x7x1x7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_lsfbW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7x7x7x7x1x7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_lsfbW4_dynB_sm100f.cubin.tar.zst index 54dc2431cc8e..a9381f617acc 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7x7x7x7x1x7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_lsfbW4_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7x7x7x7x1x7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_lsfbW4_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:6da9fa7f2d8c47868a1deef35bd0f2a6023ed7ed53963da8926336c28d8b1487 -size 80034 +oid sha256:444271862f77adc19b7bdbbb01e669d935662f27405608f5c8e2fbc77f9807ef +size 105243 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7x7x7x7x1x7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_siTuGlu_lbW4_lsfbW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7x7x7x7x1x7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_siTuGlu_lbW4_lsfbW4_dynB_sm100f.cubin.tar.zst index 2bc532bce969..e9c9a889b88d 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7x7x7x7x1x7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_siTuGlu_lbW4_lsfbW4_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7x7x7x7x1x7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_siTuGlu_lbW4_lsfbW4_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:d8d965f61e04f9c5d7af9b7e371e07f6536607cacbea9668bc470fe345d797f9 -size 79283 +oid sha256:eac38a0dd6d026f36d05b6a3ba0720f258d3043472212d1075dc6a80e9dc0e44 +size 106340 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128u2_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128u2_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst index e3f57fae580f..4140f97a7571 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128u2_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128u2_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:beb87dfeceb5981bccdf75380d61484bdcdd7afcc44f7389177c3a5c8899d441 -size 90088 +oid sha256:56b5c175c6d47f3512b1d7fa306f10576794f6faacb4031531645492957ab90c +size 118342 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128u2_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128u2_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst index be9764bc7c90..73cfb16fa80b 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128u2_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x128u2_s7_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:89f837edbfe00b401c43ca170a06572e84cf42ad5e9e53547c0aa4e7674d4c1b -size 88375 +oid sha256:341d745267e450c24eff8eee7ec351c214d411715eea2e7a96f149af1c8be019 +size 115774 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst index 094fad66a4ba..929633b40a8a 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:242726ebd5597dde8d6cbefaf6a1a8ca32ff43b8719a3119fa925c778477de47 -size 90536 +oid sha256:e7de516aba040863ff96cdc3ecd580f2e9a5447b556ae414694c1bf892c5473f +size 119395 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst index 680e59d6a963..43ba18839be7 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:cd9c35ead5d44dcebe0f578e27ee224373d0c017c747e62fb7ab70116fc6200c -size 88875 +oid sha256:185915fa010a33266214770d3d8e3ed6e8474d2cc20afbf90d67b8902c8444e5 +size 116635 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4x4x4x4x1x4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_lsfbW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4x4x4x4x1x4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_lsfbW4_dynB_sm100f.cubin.tar.zst index 4494a75920e8..a518bcdad74f 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4x4x4x4x1x4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_lsfbW4_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4x4x4x4x1x4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_swiGlu_lbW4_lsfbW4_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:b5340ef04bf91adfcb7fc1c4d6c4457b09de4497dd248d67d08efec5a9376451 -size 83701 +oid sha256:2d3a57e6640fea3373e52a47833ef9d74249cab65eebde9f64c7ae2020cd2756 +size 110631 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4x4x4x4x1x4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_siTuGlu_lbW4_lsfbW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4x4x4x4x1x4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_siTuGlu_lbW4_lsfbW4_dynB_sm100f.cubin.tar.zst index f230e7c3f31d..7fb4546443ca 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4x4x4x4x1x4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_siTuGlu_lbW4_lsfbW4_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4x4x4x4x1x4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_fCp_bN_ldgsts_ldgstsSf_rgTma_clmp_siTuGlu_lbW4_lsfbW4_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:782ffef9bf6a287a4d03990d1cda9ee9dfaad79795fd0a0a5dad2326d3602f48 -size 84919 +oid sha256:ba7199fd03eb5d1b3f6cbecee05b5b048f710dc3970a072236a8d9c06b742174 +size 112949 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256u2_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256u2_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst index ded57de5cd11..e5a6e1e5c07f 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256u2_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256u2_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:c815a55cc2f6a496897f4a8a17dd336ef3cbeb47f8080ce11576f226f7599f5e -size 93778 +oid sha256:f6e956f49de21b4f79074ef515bb89f13cc5059b93ad0b928e4bf03e6b9989b1 +size 123417 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256u2_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256u2_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst index f416ae110df3..df09f9e2999e 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256u2_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x128x256u2_s4_et128x32_m256x128x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:68817a65ab6334c7a68a320a967d7c46abcfa65e37524479818d4dd93a666d74 -size 92912 +oid sha256:f80b4c6304543f12a229fa30e3872bb3b3a247ce4b5350c336cc2c3d0fdbd879 +size 120990 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst index a27fd5c634b7..25b37208cc9b 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:b41856c1585cc9b8b663e0d63460e22c672600c4be60d3cf6e89b7abab7e4c49 -size 52454 +oid sha256:c84963e2d1157d34207b1358df42960f98eb73d48eb87f2cfefe2b4c2dbb5920 +size 65476 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 29a1f4e69923..1c9616352cfb 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:b3cbd74b6e68eff8e06cc43703b8d40276051e456b71f8c93193cee9d0e57eae -size 53156 +oid sha256:d85592277250d1ec5bb5fc850a54f622d77adc4af0c0450054eb57a33e85d9d9 +size 65899 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst index 586fb527fd47..c623aa6225e4 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:50aaaa2d7c5781bc3d9afa6219d5b02c6d663c8756946669c3df623736421459 -size 39585 +oid sha256:4f83c418d2e448da7dacd6973c61a45319e70831bd094c2cfb42c27e8c9396bf +size 48486 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 5be1ac984a5e..d95631c181bc 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:86cdaa4528f071f7e3d3dac6950ea03f00cc16419c2557e9773f4207ba954cb1 -size 39818 +oid sha256:293f34fd81bba37f4be7b34dfeaf3fbcaa0ec7bfdc584ee6df4cd625fcf57285 +size 48806 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst index 99b947c7a636..177a1ce28694 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:54ac7183d0da1d2fbe08b85c35c130845ec9dfb7848b90457973b54f85c54e28 -size 59871 +oid sha256:aa327ddc951f1f544c10977df20903c8cfb8eee7a1ddaf287e78fcf7cac4fb01 +size 74624 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 1b7e5f050962..ef546a6adb6a 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:616cdb6b5e76e1a8355ea5fe513ab77bd06f391a0245bf4b7a4ad5966fa0a4a9 -size 60296 +oid sha256:26538dbdeba060d4179b3d0a3cdb85f4f0119ae0aac8dad91cb53adc8be26cc6 +size 74802 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst index ab76b75ee3b5..0f498c319f43 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:cb114a651ed49d79dda25fe46c441cb7614dc7a28b676f52b887b332a1911dfe -size 47253 +oid sha256:7b00d13c5a7d0fa420a47641613aec8ff13c32f5bae641fbe409cc17cece4fea +size 58507 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 9bef8565625c..a9a6224c5bc6 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:4fee5aa3f1b15a81bf543a96d5b984d7edc3a57449d9c88adc5a7ade8f931f81 -size 47608 +oid sha256:5519929dc6969dab6938c9b132b40d30b9762f56bc4af9f4c7b98b76524d3c02 +size 58511 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst index 0a9b1a67aabb..32944b785f63 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:c7d787f1c3daea801e8d16bcd54c9dd31a160a41412f9260dd61390d012a5274 -size 56302 +oid sha256:98638c4c52ce3b5e9bcdf1483d287a7b215c7afe574f2e4968e1368d437304bf +size 70443 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index b307b75846cd..8d9d273ef4dd 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:57636600d4328413e82ea4bc0dd15200eb5979305864ebbd3a6d3a02ff691275 -size 56404 +oid sha256:25707683b2455f9992f978ec382c13071372f92abb4f7d2ad36156b41841e667 +size 70405 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst index cec6f4b7834a..baa37924d3e2 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:c66d5ee51620bc36134e123f94e226afc5a732483883a9dec958d3a4de704c9f -size 43072 +oid sha256:550b84d37341abacb2df66e76600dd2f2b5e4ebe8893b7f4bff5420a4d1cd58d +size 53766 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index d77005edf9b4..23022fba40f2 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s5_et128x16_m128x16x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:41067543f7f95cccc8f103ea9f89d07d6720f9d8f03e678c317cd01ebab48e3a -size 43456 +oid sha256:e84d3597e9281d21eabbcee7140365ed26a78486fbe8a3a3b2a8a6c380b42bb8 +size 53957 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst index 75121cbfc256..d3706698c141 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:0b02198db0613bf83ccd6b16f06374921c4a499c798b517b2c6cac97d2368af8 -size 63345 +oid sha256:0878b4e0944a1534af9afc344ef778c3da44647deaec03e94993c9692597dba3 +size 79130 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index c25b897a6112..49579c2ecb78 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:4dfba0f0bf1906e4d1ae22ed29cda105ce5838f65b47f776046e9e28fd195441 -size 63684 +oid sha256:5e79a856ae9250668415b22c7ea7cc40b6e6455a7440ccc7928fd7d2798b7f89 +size 79197 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst index c15027dc421f..47bfe1dd4054 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:38475bdae7c610540f2390f275ebf98a1ff2704ac2e503c23deb5f730073a7f9 -size 50813 +oid sha256:6ee720f0a3db6784344d04209f950519e7f87e688a8256cfcb4b97a86220f355 +size 63321 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 7cb239ad5539..3c53e2dc7023 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x16x256u2_s6_et128x16_m256x16x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:894acd360252ab0da1f49ea82c71888d5b57bb712d24dcf3473d717876a7c5b6 -size 51189 +oid sha256:16421cf47092b32b4711c9c90d3082000c22f0fa4c52eb62a7e7a53fbe46ee82 +size 63360 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x128_s6x6x6x6x1x6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x128_s6x6x6x6x1x6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst index e4c7b58a4a19..a68a86e18b76 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x128_s6x6x6x6x1x6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x128_s6x6x6x6x1x6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:eb190fd8b4b12acada0c9eec4c83492302f7c6dfaa45953a98121258eafeb2b4 -size 123701 +oid sha256:6657507724cfd81ab627189746585c923ab90301e689528de54832adda9989bc +size 165322 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x128_s6x6x6x6x1x6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x128_s6x6x6x6x1x6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst index 9df0705d1ee4..5624a746aa3a 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x128_s6x6x6x6x1x6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x128_s6x6x6x6x1x6_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:1aa5dc980c575e4d0779310e2a0e173c93371d69a09be88f218816a826abe595 -size 131787 +oid sha256:38e0c94b8007a012a481702bc5b7f98d1feeabd9df272e52f1bbcfd00a0cd82e +size 178606 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x256_s3x3x3x3x1x3_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x256_s3x3x3x3x1x3_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst index ebbe83aeda4b..77d80eb92ad7 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x256_s3x3x3x3x1x3_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x256_s3x3x3x3x1x3_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_biasFp32M_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:09c22638a72e8b39a290c8f71be7d3620ebcfc8ed8e995b9a688cdd7ee15dc6f -size 125015 +oid sha256:32afee86036a428ed1c2727e8dd2b213ff1b704eca7cf3f42ceb26b63cf39698 +size 167971 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x256_s3x3x3x3x1x3_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x256_s3x3x3x3x1x3_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst index dd82cde1699c..0a037517fe81 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x256_s3x3x3x3x1x3_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x256x256_s3x3x3x3x1x3_et128x64_m256x256x32_c2x1x1_rM_TN_transOut_schPdx3_fCp_tmOv_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW8_lsfbW4_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:fee4c4c8e612def8ee2d910153e686c418b1ede9fb440a767ae9bee0ec1aa435 -size 136031 +oid sha256:253fc319ca01435768098950f3d69cc4412aa5a1b3b57dc9a89a4ad8ccb3fe11 +size 185186 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst index db45b13bb49f..20b17adaee9a 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:c5fd6ae8460fde86646fec8c922ea54384db4ec978c05c075dcb2c306f9ed6c2 -size 59459 +oid sha256:fc89d92d21af6f92f22fa9472eba2e6c68dbdb062451691e69713308c4f2a75b +size 75185 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index ee9cf3853e89..17b3b44c7cd7 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:25e6e7f16c5c87d394f31eda2be216baa40d5214ca4629093d78b3dfcdb79a65 -size 59153 +oid sha256:1c2e02693a97610157cd0c387f959c5280d7a73750bcf8b3a74010a0f7bda60d +size 74581 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst index 80ab0accd65b..2a4c6b6b93da 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:9f739b7cebd82174a09e1d720b445089ef13dec3c9bc4e4089bd602d0c68728e -size 46314 +oid sha256:e609d57de352f6cfffac9e0cb60445551f9c881b82180676e411161a2b0db0bf +size 58620 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 816d4b62a38d..42afb07d4fd0 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:295881343f19511e9ae87575b4faaef640be2a2051df57916131d10b09742bbb -size 46415 +oid sha256:cae4a9886cd0c6936cef894f92494f672f6ec6f73725166d885574fbeef08325 +size 58143 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst index 575405c0f6e3..5fb596f20ad9 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:abe1d46847218db224154e25baf92689a669f603582a992b060fa51007d0395b -size 66711 +oid sha256:86edfe474090086ce12b5ec1a03eee5ff18efda4cadffea85542906dbf2d6737 +size 83423 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 36545d0e12b3..0323764eb21d 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:c1e1a7f3a7b1ba29f4bcd779ef50e972fafc4e79dde31c0ab71192793fa7d838 -size 66774 +oid sha256:c9879c4e348487dfd6a69ec60b30d0735e71c4ae9cfe23c0e1ef5348612be490 +size 83214 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst index 2d6d6058b0be..e433cd2612fe 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:8794f3685691aa71597f4d77ea994146f459e034ad9f768fec5fb9d00610460f -size 53310 +oid sha256:b8c598b44164b81f85c4d079c51ae79654b2924b0737b3b6d5ce8ecf138542b7 +size 67286 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 892206e5650e..c0deef2b9ebc 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:bbfca9bc509492d5b6450463cd6270fe1913669176ee9eb3e29593822b70f2d8 -size 53391 +oid sha256:4a7dc6a2caf81da1bdf26d65da6aa9a197f5a5a48211bc7eaf54a03f8ca62357 +size 66711 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst index a91f169105cf..39b604105a8f 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:6abc3ffbba51418fa0f1e05e1d69dd406da39b98194db602306e758a602559f6 -size 63242 +oid sha256:a8d11137fc93b6cfe2d0174b8514cca7ed6b8be2b80bfd386f0237a2c92158aa +size 79609 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index bc2b9337382c..98422183030e 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:dedf4f8b9688cc57c247e746952439e6cd5bc9039135a971e09f303e3fcefa5c -size 63032 +oid sha256:a705c3fc999c07b37759a87b8d550881c5310be1c8073b8d57a48b721aadf9df +size 79220 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst index f90ced06d929..084867715865 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:03a2d1679def061efa0939ca78440127e76e7440a14ec128d07c0ee88cb0b598 -size 49831 +oid sha256:bda0d32c8f77e71881376bfc5535ab4ea266d9cf1d8adfdfbb9460980b32b4b6 +size 63378 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 27f24378bf14..28f0d11ab04b 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m128x32x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:c8da916f1f9b60a36778819ebe88e96d9325c3ea56468caa44990a79357690f9 -size 49956 +oid sha256:7c94937cac09df635d7994cc3faf5567363c360e17b74caed5f142035e6bf598 +size 62940 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst index b5b77d1eb446..56722b31c669 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:623daae40fdf3e4d8cba2d3c7ae4923149a8fc081694b7aeb17a59c4e8ade801 -size 70236 +oid sha256:1ab99b21df199d40e4ddf0d8f71f001c4d1fef933241721a4d3936a24317e8e2 +size 88020 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 32074e799798..3582d3493832 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:f9bf61a065d821de47a58b149959f2a6acec10e9641c851b8e55ff852753c943 -size 70436 +oid sha256:c7fbeb461b24446a3dcc19a54b99f6026db37b3d6c9bb05e2a6c5aaf6d342d99 +size 88149 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst index c0c6286599fa..4f037be31a50 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:95842461625f600a6c5b02248715b42b0ba76f4a921c2458bac63c80aee301b5 -size 56941 +oid sha256:e814bd4f3001698ba82d66b191604a9280d9694f4d2666682d03947b0fcb69af +size 72275 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 311be2095b25..4b2301edb35c 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x32x256u2_s5_et128x32_m256x32x32_c2x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:321ca0107654af2877d790f8fcebfa65f13c5a7c05e9d54d22d07c9e5c2e921b -size 56854 +oid sha256:8796858dfda4ade622ffdf388208dc10694340ffcc612bf149a129f8f6ddca09 +size 71648 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f.cubin.tar.zst index 8c94c5fbccac..6a3f006b0a13 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:6b507d39c5a9a63b5fa5740433334a001db5f5911dc153e8c3357004446cc0fd -size 77132 +oid sha256:9b1b9ed8f217278321bc38fe607fb3b86b89c2a6b51918a917cda252873f489c +size 97690 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f.cubin.tar.zst index 4cdbbdeb5b97..e4cc11ba7577 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:7cdcf8a46100ccfe8a9048eaa0bbdb91136c7688df15fbfff137e6ed303402dc -size 76284 +oid sha256:fefeaa137214b5e6a45515c046d5356297b7c082030219ab87bf2940b1cdd834 +size 95802 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128u2_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128u2_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f.cubin.tar.zst index 42539f6bb949..656bc4e88e57 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128u2_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128u2_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:30cc3638b3779520d5cdd37ec711e17e5c43c5d2401389dfa1b389f9db0bfcf9 -size 80054 +oid sha256:0c001c31989219e334053ad68b1c999694dbe5620382f8444b0effae3b95620c +size 101908 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128u2_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128u2_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f.cubin.tar.zst index ac63ca665c1a..02862462ac30 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128u2_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x128u2_s7_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:f0c730eeb3ee582bd9da74f5808736bc25c0a79a532763d6c2ffef6dd714a52e -size 79199 +oid sha256:b86adadde940edeb035ebac685ad5c2e3106360e6c03f61cc56149895ba03849 +size 99961 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f.cubin.tar.zst index 483195f4ddf5..bf62c5b3a20c 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:6058a8acbcd22714d5d3fb9897ee2df54b1ff6ee26157c7c297aa6ac9ec0cf15 -size 83225 +oid sha256:573a9b1f5e3a263297aad4607ebc739edb9363623a54875e57716e10e0aa16cc +size 106895 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f.cubin.tar.zst index c03b7960d0a6..d324d52bbfe5 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:3ba3adcb6a979cf34df4555bcf225a631b1054a030bab429fa3865fdeb0fbcf4 -size 82071 +oid sha256:2bde5cff235429d2e70f2c9ff99c8d6c0849e2ea3127c812f1daf5baf7d67dcd +size 104260 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256u2_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256u2_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f.cubin.tar.zst index 5c31f04cd51f..492cfa3a0ed0 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256u2_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256u2_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_lbW1_lsfbW1_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:5b0ef7f22bbc80f091816af10459b544b296171a4a32d16233c88f46146335d2 -size 85866 +oid sha256:312d493c6462327b0e565beef40c662830fcacb475363569ec151f8fff2ac796 +size 111542 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256u2_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256u2_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f.cubin.tar.zst index 33f0babba0b0..dbf1d7bc0f05 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256u2_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x64x256u2_s4_et128x64_m256x64x32_c2x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_lbW1_lsfbW1_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:c3a65c8ef4e38eda1b836039ca4c9d422f7ee33c8bea4d9328a81c332dc324e9 -size 85646 +oid sha256:a5ea28216b1b48ccf51a605ff05f5db7e1ba5efc3c7ed3da9fa997a0ed21f326 +size 109694 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst index ad3a52f33617..b7c9b030cb11 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:0d5469eba92f88827efcfd343d1e49ab04f4908bc9188afc0f88cf5d7f371e65 -size 48744 +oid sha256:facf68ed94bf161fe0e8b96bc36785ad97fb3829c36fec2238a61ee382fea309 +size 60122 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index dcb54f02e61e..4915fe05cf2b 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:5d71b1cdf799fe0a118a47dadc4c9bdc96ea6cb5c81a098b73a38f2848b32ebd -size 49373 +oid sha256:df999f23c625a9c09c04e1148985544b98458aeed4b5f6b291c6e3daa5975086 +size 60738 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst index cf2ccf6c2f8f..fc94dddd0e30 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:900651f9751e93e5fcd5e7a605bdac3cb6379990bcf67be91574cb733a7402dd -size 36707 +oid sha256:5c2b657e804a2c86f362bd116207a668c700727a59ed69c718928203d85fda76 +size 44829 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index eecc28ff7acf..e20e672034ee 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:c62bbc704a088838720063e9b8f70df9dc508f89a5a255d8077f9e5f161f3c50 -size 37198 +oid sha256:8cfa1c73a8cae6ea60e57c9eb1d4edf3c39f22c7c92b40bb30534eba032ac804 +size 45190 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst index ce8265479e2c..b11585f5ac75 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:8b856c952314567c4cf6bacdc7558937133dbb1cb16ac729c2487d47cd83a206 -size 58806 +oid sha256:e2ce72b79245a4ccf008fb9a21aa875d2f3a702bfacab060067429bee43dca9d +size 73947 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index de86458cb7c8..526ca401e47c 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:bd00b1466fb2ae39cda25f61c6f60d3039e962b53ecaf733c1c1014d6c101e75 -size 58980 +oid sha256:05ec18be17b9186641ce0de8331c7e4b26dd83ca62ba633d44650d9880d15873 +size 74986 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst index 824695f13b31..19a3f09f7f0d 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:b4377465a8e5cd1166d3c461965dbe8e1e8c7b1b24e03730abafe4335370f8ab -size 44439 +oid sha256:a8c98b2edc04dad2d801cd1df63f28d41fef8f6beede03ca3c25261fb1ea8a59 +size 54749 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index aa5648e9ad94..664fecd61be5 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:106137331cfa1d581443075a6cf2d9b18defac478bd13894f34786c86072658d -size 44469 +oid sha256:82930b0ed805c7073cc094fa7f1afe7180d797ef4da4028d7cdd593a9383cf35 +size 54879 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst index 0e4a4c6951a2..5ebdfca78eec 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:3c09bbf6c85ab45a3d76dbe49c6fad4a2ec286b84c34289f7c246bd34c634f68 -size 52693 +oid sha256:6771dfddcb5bc9e45ae6b95a269bc414f56a8d6507b5182fe2fa10258539b68d +size 65443 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 803b3d5d498a..52726a0cb40c 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:b053f7bf38854af13b56c5cd90431a48e7f229e7f62cc4d1930e730d3021ebb8 -size 52920 +oid sha256:1bdaa7d0139d24edec4e1a055b6c99e0aee1ddf01d4e3b477be6c1d0919da738 +size 65891 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst index 93c904cbe793..611c059bd86e 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:c08e34e0be92cc4b1e8c182fed71b422b4e2876e045066c3fd078457aba1e23d -size 40389 +oid sha256:4dad1ccb40cf86eb15e740db04d93a0cb1d2f7fee0cbdf7d2b0fa0a859a2907e +size 49775 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 5b32a9d8d110..a58103704adb 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:a8de8cc7648968fe5dadd4c72c7d991a5c07c36d027ff65ded9f8a90c89b636e -size 40861 +oid sha256:d809fd4c2a4752921bf3e8955370d52ee61334a19a831c8066ca10ad09e7d76c +size 50274 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst index a4963f062570..8d931a14ce44 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:f8961d9acf9fc0e41be4e4862e9bb63fbf5d38c4b9abf203b8a54ff534f665d6 -size 62417 +oid sha256:b9e2b3cd03e0131b9f88faebf5be2dc05af55dd59f3fad2bd0adb8518bb08e95 +size 78609 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index c6ecbf7cd8ed..93c5140943fd 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:d97945b71da1454db440fee03ecd9b343049c0e6146c9855f09aa5974a42561a -size 62667 +oid sha256:3f645792488130538b00672fb4b5db59795abab7ccc941d7e302319d8ef3decb +size 79476 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst index dcc154a2e753..78f17a3ab371 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:69f9598562f8fc57d456573cfd697ad341d41debf383bf3e01186ec1c92af5e0 -size 48193 +oid sha256:75406150bc4bae2ca1607a69bd0fbbdca922d160caf24eec9d1e71e971f39b6c +size 59443 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index e13370daab0e..29b38ff0a793 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x256u2_s6_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:4c9593eaa8a62e418237d089b57df053439e2e4a964331a7209a08a5e7c478ec -size 48486 +oid sha256:f3d3e5bc706b95c72cb0033eda05ab5a9cb1709d1fb6640f04be5a2add41b6f0 +size 59715 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst index c75dc505c3d8..d3f50ea978ba 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:a55f28e435cd9ac7e86cdcd09fee4dd41a9696856ba9cf631640843422f03ecb -size 50501 +oid sha256:2f76f58108d4aaa25e146c6b86bebb202623b7f65da199d87d3bba8f515acd9e +size 62848 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index da8f7f66c4dd..9d5efa20cb94 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:852f895e411bed93c9e75488b2795e7fb11c6f763dfef167b2f523b5a5117934 -size 51156 +oid sha256:2fdf791d9f98fcf6e5121f89b379b9736e92e24e22d9969fd84c6a72c3876d21 +size 63380 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst index 6d22a6b7310d..6a73daf141cd 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:edcd72ea9e4d96295880a3e4280dc1f785c634bd49d29867159855189dd8f904 -size 38773 +oid sha256:c3ff5342eeffcea4cfde87743a41e303e3fa03fbb963653c6798b3e117470058 +size 47603 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 96889e57f3b3..d0610c38ca73 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:1a62ba6aeb55ad77db14c00593ae727bb898bfc73b8e7b3074ae7484aaff0d17 -size 39203 +oid sha256:ace2dcec0b4f480e855db15fdfc645a961f9725b32aa27a0013888e863fe6b04 +size 48047 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst index a32439b31304..d6a9e5f4f03c 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:828c3c9d719a2da3bac10258b82db08decdde5ab2b38305239c8205c3abef7eb -size 60724 +oid sha256:41dd9310de2ab0f16628ca4616342134af591fae3ee20e47cf673f24eb279404 +size 76301 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 1099b4ca30ea..d82b94acb72f 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:f35268a0286868dee2af4f5575d65a856252b005197768a2174b8f4615f95388 -size 60498 +oid sha256:25c1a7a9fa36807695a052e0017e60263f9bb1ddcb8db99b6b5b2b474d945c6c +size 77283 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst index 32e7cd7708e7..332be2235bfb 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:3baa0f9c234044722f2a5d2839e45b40d5f9766a56a46cea8d4fa9ffefe6c1f6 -size 46399 +oid sha256:7b39a70c35a22cb52189df1359383f53dae9c3e9ff7e1201e3509b4a6bacfe6c +size 57540 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 7078b8c8838e..f91ff459af87 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:9eea75f3d939a2308064038a871262e5fb730b81e273027b19ad661de62a175e -size 46467 +oid sha256:0f31488793faff1f9adb78589685b890f5f5ab2b397a8f796d110a8975ae4527 +size 57561 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst index c4b4f3d42d68..1c0275477d06 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:90f67b1fb2d8010227a14de435a3174b0e9d1fd90697696af432145b49dc5ae0 -size 55331 +oid sha256:a94097172efa8696728e6b17b6865bd620204790ce14f79cb7cff153425afbb7 +size 69488 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 263b169d2667..bec07e10079c 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:4f6b4d0f32f9f54582d6265f801c538b9c01ae8502c0b3ba89e93823212e1982 -size 55391 +oid sha256:bcc318f490104e7b174e1af171ec201cacf367e5acd98a1590c1ab332cb21056 +size 69502 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst index 47403bfe7aee..3182662ab355 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:b83ac839fa9abed200650decb265bf38578720d8b4a23836769b9d4ad18957a3 -size 43876 +oid sha256:c96a65a867de71028793aecac0f5762a342120613ad31e34d3b3f254d71d8871 +size 54454 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 0b8450ca1559..a553b42de543 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x1_rM_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:7310cd6b59d78c21a5e76509f4b4bb3b800fbfa17993dbf981605a5722bba14b -size 44321 +oid sha256:fa047328d3d3350e43c8c4829bfc4ef0426f7a989dda3e3ec6709ce8bb2603bb +size 54841 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst index 318ffdd1f14d..0bf731e4ac9b 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:fb10d7dcc5e310c2abdd49c7855f47241b35d88472a3371b71b7ac4fd4786ae7 -size 65064 +oid sha256:20d6d69a31033a4bb3216ea207b49e684deb18d0260eb4ef0336f37b33db1bc2 +size 81515 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 45881ee5e925..6221af24358a 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schPd2x1x2x3_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:40fb4db421639030afa488e8e46a0ffa9eda7ac76413b34037a47150c6689bb2 -size 65195 +oid sha256:077a0f84e1689e480b72eb692f8ad40912937f06e27a95dd6c70e73559878ad6 +size 82210 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst index 84a1491da22d..2bc1ed5e7ac7 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_bN_tma_ldgstsSf_rgTma_clmp_siTuGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:fa1d9a7ab64332d275392ad2cccf403b9d214822ea6de6a2585bdfb372495842 -size 51199 +oid sha256:4c744d172d3b23f559d2b657e57054308ff24a0acdf9eebf75fc5cc3922e5d25 +size 63620 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst index 2773abd5ecf7..fc6ff57552ef 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE2m1MxE4m3_Fp32_Ab32_Bb32_Cb32_t128x8x512u2_s3_et128x8_m128x8x32_c1x1x2_rM_splitK2_TN_transOut_schedS_biasFp32M_bN_tma_ldgstsSf_rgTma_clmp_swiGlu_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:d9a364248261c47a50aa692f4c446932db317d9f1d94ef0030a2f1b0bb8bdd11 -size 51283 +oid sha256:16a12159b04a8239957329a1c6bc87ed77f05a03652ec6f815d49317e168360b +size 63625 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE4m3MxE2m1_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7_et128x64_m256x128x32_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_laW8_lsfaW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE4m3MxE2m1_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7_et128x64_m256x128x32_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_laW8_lsfaW4_dynB_sm100f.cubin.tar.zst index 954b5d8bfbb9..02198cfe226b 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE4m3MxE2m1_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7_et128x64_m256x128x32_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_laW8_lsfaW4_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE4m3MxE2m1_Fp32_Ab32_Bb32_Cb32_t128x128x128_s7_et128x64_m256x128x32_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_laW8_lsfaW4_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:776324bf080073f0059a2c6d7b45508db0e812471f56454ef59874a5b3aec773 -size 84260 +oid sha256:52e57bd6536043f38ca4ab399e77087f2dd6a571173a3e8bb167e56e06e71858 +size 107513 diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE4m3MxE2m1_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4_et128x64_m256x128x32_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_laW8_lsfaW4_dynB_sm100f.cubin.tar.zst b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE4m3MxE2m1_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4_et128x64_m256x128x32_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_laW8_lsfaW4_dynB_sm100f.cubin.tar.zst index b9b443633c1f..f9b4a12149ab 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE4m3MxE2m1_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4_et128x64_m256x128x32_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_laW8_lsfaW4_dynB_sm100f.cubin.tar.zst +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/cubins/Bmm_MxE4m3_MxE4m3MxE2m1_Fp32_Ab32_Bb32_Cb32_t128x128x256_s4_et128x64_m256x128x32_c2x1x1_rN_TN_schPd2x1x2x3_biasFp32N_bM_tma_ldgstsSf_rgTma_clmp_swiGlu_laW8_lsfaW4_dynB_sm100f.cubin.tar.zst @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:6b990868fd33bbf1c8756215e4ad7136aca0363eda19a56d0123523ed8fd8f6a -size 87396 +oid sha256:3e5f48bf8cad45b568a0dad5f0012cb303e6145325806ddbf6b3cafa748e32b8 +size 112252 diff --git a/tensorrt_llm/_torch/modules/fused_moe/fused_moe_trtllm_gen.py b/tensorrt_llm/_torch/modules/fused_moe/fused_moe_trtllm_gen.py index 4e3e93eed597..49bb019074df 100644 --- a/tensorrt_llm/_torch/modules/fused_moe/fused_moe_trtllm_gen.py +++ b/tensorrt_llm/_torch/modules/fused_moe/fused_moe_trtllm_gen.py @@ -129,6 +129,16 @@ class TRTLLMGenFusedMoE(MoE): ActivationType.Relu2, } + # Quantization algorithms with fused SiTu FC1 cubins in the TRTLLM-Gen + # batched-GEMM kernel drop. NVFP4 feeds the group-16 `Bmm_E2m1_E2m1E2m1_... + # _siTuGlu_*` kernels; W4A8_MXFP4_MXFP8 the group-32 + # `Bmm_MxE4m3_MxE2m1MxE4m3_..._siTuGlu_*` ones. There is no standalone + # SiTu activation kernel, so anything without a fused cubin is rejected. + _SITU_SUPPORTED_QUANT_ALGOS = { + QuantAlgo.NVFP4, + QuantAlgo.W4A8_MXFP4_MXFP8, + } + @classmethod def can_implement(cls, p: MoEProblem, d: MoEDeployment) -> MoEEligibility: """TRTLLM-Gen kernels: SM100/SM103, bfloat16 activations. @@ -373,21 +383,24 @@ def _validate_backend_local_activation(self) -> None: if get_sm_version() not in {100, 103}: raise ValueError("TRTLLM-Gen SiTu requires SM100 or SM103, got " f"SM{get_sm_version()}.") - if (self.quant_config is None - or self.quant_config.quant_algo != QuantAlgo.W4A8_MXFP4_MXFP8): - quant_algo = (None if self.quant_config is None else - self.quant_config.quant_algo) + quant_algo = (None if self.quant_config is None else + self.quant_config.quant_algo) + if quant_algo not in self._SITU_SUPPORTED_QUANT_ALGOS: + supported = ", ".join( + sorted(algo.name for algo in self._SITU_SUPPORTED_QUANT_ALGOS)) raise ValueError( - "TRTLLM-Gen SiTu requires W4A8_MXFP4_MXFP8 quantization, " + f"TRTLLM-Gen SiTu requires one of {supported} quantization, " f"got {quant_algo}.") if self.tp_size > 1: # Intra-expert MoE TP: w1/w3 column-shard and w2 row-shard along - # the intermediate dim (the stock MXFP4 quant-method loaders slice - # the group-32 packed bytes and scales per rank). Require the + # the intermediate dim (the stock MXFP4/NVFP4 quant-method loaders + # slice the packed bytes and scales per rank). Require the # per-rank shard to stay a whole multiple of the quant method's # weight alignment so per-shard scale groups and the padded # weight buffers line up without fractional groups. - alignment = W4A8MXFP4MXFP8TRTLLMGenFusedMoEMethod.weight_alignment + alignment = (NVFP4TRTLLMGenFusedMoEMethod.weight_alignment + if quant_algo == QuantAlgo.NVFP4 else + W4A8MXFP4MXFP8TRTLLMGenFusedMoEMethod.weight_alignment) if (self.intermediate_size % self.tp_size != 0 or self.intermediate_size_per_partition % alignment != 0): raise ValueError( @@ -557,13 +570,15 @@ def _check_configs(self): if not isinstance(self.op_backend, TRTLLMOpBackend): raise ValueError( "TRTLLM-Gen SiTu requires the native TRTLLM op backend.") - if not self.has_w4a8_mxfp4_mxfp8: - raise ValueError( - "TRTLLM-Gen SiTu requires the W4A8_MXFP4_MXFP8 path.") - if self.scaling_vector_size != 32: + if not (self.has_nvfp4 or self.has_w4a8_mxfp4_mxfp8): + raise ValueError("TRTLLM-Gen SiTu requires the NVFP4 or " + "W4A8_MXFP4_MXFP8 path.") + expected_scaling_vector_size = 16 if self.has_nvfp4 else 32 + if self.scaling_vector_size != expected_scaling_vector_size: raise ValueError( - "TRTLLM-Gen SiTu requires MXFP4 scaling vector size 32, " - f"got {self.scaling_vector_size}.") + "TRTLLM-Gen SiTu requires scaling vector size " + f"{expected_scaling_vector_size} for this quantization " + f"mode, got {self.scaling_vector_size}.") # For SiTu these hold the backend-local activation parameters # (populated by create_weights, which runs before this check). for name in ("swiglu_alpha", "swiglu_beta"): @@ -581,10 +596,19 @@ def _get_quant_method(self): if self.quant_config.layer_quant_mode.has_fp8_block_scales(): return DeepSeekFP8BlockScalesFusedMoEMethod() elif self.quant_config.layer_quant_mode.has_nvfp4(): - return NVFP4TRTLLMGenFusedMoEMethod( - ) if self.swiglu_alpha is not None or self.activation_type in [ - ActivationType.Relu2, ActivationType.Silu - ] else NVFP4TRTLLMGenFusedMoEBaseMethod() + # ``is_situ_activation`` (not ``swiglu_alpha is not None``): + # SiTu fills the swiglu_alpha/swiglu_beta slots from + # create_weights, i.e. after this runs, so keying off the + # tensor would make the selected method depend on *when* + # _get_quant_method is called. Like the SwiGLU-alpha and + # element-wise cases, SiTu needs the padded method's + # alignment handling. + needs_padded_method = ( + self.swiglu_alpha is not None or self.is_situ_activation + or self.activation_type + in [ActivationType.Relu2, ActivationType.Silu]) + return (NVFP4TRTLLMGenFusedMoEMethod() if needs_padded_method + else NVFP4TRTLLMGenFusedMoEBaseMethod()) elif self.quant_config.layer_quant_mode.has_w4a16_mxfp4(): return W4A16MXFP4TRTLLMGenFusedMoEMethod() elif self.quant_config.layer_quant_mode.has_w4a8_nvfp4_fp8(): @@ -616,10 +640,11 @@ def create_weights(self): # rejected by _validate_backend_local_activation) and feed the same # gemm1_alpha/gemm1_beta op slots. Safe with respect to the # `swiglu_alpha is not None` gates: create_moe.py checks the - # constructor kwargs (None for SiTu); _get_quant_method consults - # swiglu_alpha only on the nvfp4 branch (SiTu requires - # W4A8_MXFP4_MXFP8) and has already run above; _check_configs runs - # after this point and its swiglu gate admits w4a8_mxfp4_mxfp8. + # constructor kwargs (None for SiTu); _get_quant_method's nvfp4 branch + # keys off is_situ_activation rather than swiglu_alpha, so it returns + # the same method before and after this point; _check_configs runs + # after this point and its swiglu gate admits both nvfp4 and + # w4a8_mxfp4_mxfp8. if self.is_situ_activation: self.swiglu_alpha = nn.Parameter(torch.full( (self.expert_size_per_partition, ), diff --git a/tensorrt_llm/_torch/modules/fused_moe/moe_op_backend.py b/tensorrt_llm/_torch/modules/fused_moe/moe_op_backend.py index 3f37c694615d..06e6ef3827b6 100644 --- a/tensorrt_llm/_torch/modules/fused_moe/moe_op_backend.py +++ b/tensorrt_llm/_torch/modules/fused_moe/moe_op_backend.py @@ -335,15 +335,50 @@ def run_fp4_block_scale_moe( ): hidden_size = gemm1_weights.shape[-1] * 2 if gated_act_type == int(ActType_TrtllmGen.SiTu): - if hidden_states.dtype != torch.float8_e4m3fn or hidden_states_scale is None: + # Fused SiTu FC1 cubins exist for two input formats: NVFP4 + # (group-16 scales, packed-FP4 activations) and MXFP4 (group-32 + # scales, MXFP8 activations). Key off the gemm1 scale group size + # exactly as the op dispatch below does, so the diagnostics + # describe the path the request will actually take. + is_nvfp4 = ( + gemm1_weights_scale is not None + and gemm1_weights_scale.shape[-1] == hidden_size // 16 + ) + is_mxfp4 = ( + gemm1_weights_scale is not None + and gemm1_weights_scale.shape[-1] == hidden_size // 32 + ) + if is_nvfp4: + if hidden_states.dtype != torch.uint8 or hidden_states_scale is None: + raise ValueError( + "TRTLLM-Gen SiTu on NVFP4 expects packed-FP4 (uint8) " + "activations with block scales, got " + f"dtype={hidden_states.dtype}, " + f"has_scale={hidden_states_scale is not None}." + ) + # The NVFP4 op takes no valid_hidden_size/valid_intermediate_size; + # padding is absorbed by the padded weight buffers and the + # caller slices the output back down. + elif is_mxfp4: + if hidden_states.dtype != torch.float8_e4m3fn or hidden_states_scale is None: + raise ValueError( + "TRTLLM-Gen SiTu on MXFP4 expects dynamically quantized " + "MXFP8 activations with block scales, got " + f"dtype={hidden_states.dtype}, " + f"has_scale={hidden_states_scale is not None}." + ) + if valid_hidden_size is None or valid_intermediate_size is None: + raise ValueError( + "TRTLLM-Gen SiTu requires valid hidden and intermediate " + "sizes for padded MXFP4 weights." + ) + else: + got = None if gemm1_weights_scale is None else tuple(gemm1_weights_scale.shape) raise ValueError( - "TRTLLM-Gen SiTu expects dynamically quantized MXFP8 " - "activations with block scales, got " - f"dtype={hidden_states.dtype}, " - f"has_scale={hidden_states_scale is not None}." + "TRTLLM-Gen SiTu requires NVFP4 (group-16) or MXFP4 " + f"(group-32) gemm1 weight scales for hidden_size={hidden_size}, " + f"got gemm1_weights_scale shape {got}." ) - if gemm1_weights_scale is None or gemm1_weights_scale.shape[-1] != hidden_size // 32: - raise ValueError("TRTLLM-Gen SiTu requires MXFP4 weights with group-32 scales.") if gemm1_alpha is None or gemm1_beta is None: raise ValueError( "TRTLLM-Gen SiTu requires per-local-expert alpha and beta tensors." @@ -359,11 +394,6 @@ def run_fp4_block_scale_moe( ) if gemm1_clamp_limit is not None: raise ValueError("TRTLLM-Gen SiTu does not use gemm1_clamp_limit.") - if valid_hidden_size is None or valid_intermediate_size is None: - raise ValueError( - "TRTLLM-Gen SiTu requires valid hidden and intermediate " - "sizes for padded MXFP4 weights." - ) if hidden_states.dtype == torch.uint8 or hidden_states.dtype == torch.float8_e4m3fn: if ( diff --git a/tensorrt_llm/_torch/modules/fused_moe/quantization.py b/tensorrt_llm/_torch/modules/fused_moe/quantization.py index 07463e8550ae..b661790ead75 100644 --- a/tensorrt_llm/_torch/modules/fused_moe/quantization.py +++ b/tensorrt_llm/_torch/modules/fused_moe/quantization.py @@ -4886,6 +4886,36 @@ def _register_mega_shared_staging(self, module: torch.nn.Module): delattr(module, attr) +def _fc31_scale_c_omits_dequant(module: torch.nn.Module) -> bool: + """Whether the FC1 kernel's scaleC must be quantScaleC alone. + + Mirrors the trtllm-gen host-side rule in + ``kernels/BatchedGemm/BatchedGemmTestUtils.h`` (the + ``!isLinearInX0(mActType) && mFusedAct`` branch) and ``getKernelScaleC``'s + ``doesEltwiseAct`` argument: scaleC carries a ``dequantScaleAb`` factor + only when the activation is linear in x0 *and* fused into the GEMM + epilogue. Two cases must drop that factor, or the kernel applies it twice: + + * Relu2/Silu are element-wise, so dequantScaleAb reaches the kernel + through ptrScaleAct and is applied before the activation. + * SiTuGlu is not linear in x0 -- x0 feeds ``tanh(x0 / beta)`` -- so the + kernel folds dequantScaleAb into the tanh/sigmoid arguments via + scaleGate and applies scaleC to the activation *output* instead (see + the ``isLinearInX0`` branches in ``GemmGatedAct/GmemGatedAct.h``). + + SwiGlu and GeGlu are linear in x0 and do need the combined scale. + + Getting this wrong for SiTuGlu is not a small numeric error: the extra + dequantScaleAb factor (~6e-8) drives the FC1 output's per-block E4M3 + scale factors below their smallest subnormal, so the NVFP4 intermediate + quantizes to exactly zero and the whole MoE returns zeros. + """ + if getattr(module, "is_situ_activation", False): + return True + return getattr(module, "activation_type", + None) in (ActivationType.Relu2, ActivationType.Silu) + + class NVFP4TRTLLMGenFusedMoEBaseMethod(NVFP4FusedMoEMethod): weight_dtype = float4_sf_dtype block_scales_dtype = torch.float8_e4m3fn @@ -5271,10 +5301,8 @@ def _compute_fc31_scale_c(self, module: torch.nn.Module): # c_global_sf: fc2_input_scale # For gated activations (SwiGlu), scale_c_fc1 includes both input and weight scales # For non-gated activations (Relu2 or Silu), scale_c_fc1 is just the input scale - if hasattr(module, 'activation_type') and module.activation_type in [ - ActivationType.Relu2, ActivationType.Silu - ]: - # For Relu2/Silu: scale_c_fc1 = fc2_input_scale (broadcast to all experts) + if _fc31_scale_c_omits_dequant(module): + # scale_c_fc1 = quantScaleC only, broadcast to all experts. module.fc31_scale_c.data.copy_(module.fc2_input_scale.data.expand( module.expert_size_per_partition), non_blocking=True) @@ -5302,13 +5330,8 @@ def _finalize_shared_expert_alphas(self, module: torch.nn.Module): # The shared host copy of fc31_scale_c is consumed by online EPLB # when an expert is migrated into a local slot, so it must match - # the main-slot formula exactly (see load_quant_scales above). - # For Relu2/Silu: fc31_scale_c = fc2_input_scale (broadcast). - # For gated (SwiGlu): fc31_scale_c = fc2_input_scale * fc31_alpha. - if hasattr(module, - 'activation_type') and module.activation_type in [ - ActivationType.Relu2, ActivationType.Silu - ]: + # the main-slot formula exactly (see _compute_fc31_scale_c above). + if _fc31_scale_c_omits_dequant(module): local_shared_fc31_scale_c = module.fc2_input_scale.data.cpu( ).expand(num_shared).contiguous() else: diff --git a/tests/unittest/_torch/modules/moe/test_nvfp4_situ_moe.py b/tests/unittest/_torch/modules/moe/test_nvfp4_situ_moe.py new file mode 100644 index 000000000000..488fdd4e38ee --- /dev/null +++ b/tests/unittest/_torch/modules/moe/test_nvfp4_situ_moe.py @@ -0,0 +1,313 @@ +# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# SPDX-License-Identifier: Apache-2.0 +"""NVFP4 + TRTLLM-Gen SiTU MoE tests. + +SiTU on NVFP4 is served by the fused ``Bmm_E2m1_E2m1E2m1_..._siTuGlu_*`` FC1 +cubins (group-16 block scales); the MXFP4 variant is covered separately by +``test_kimi_k3_situ_moe.py``. The two paths differ in one way that is easy to +get wrong: NVFP4 carries *global* scale scalars into the kernel +(``output1_scale_scalar`` / ``output1_scale_gate_scalar``) while MXFP4 relies +purely on block scales. + +Because SiTuGlu is not linear in x0 (x0 feeds ``tanh(x0 / beta)``), trtllm-gen +requires the kernel's scaleC to be ``quantScaleC`` alone -- ``dequantScaleAb`` +is folded into the tanh/sigmoid arguments via scaleGate instead. Passing the +SwiGlu-style combined ``dequantScaleAb * quantScaleC`` multiplies the FC1 +output by ~6e-8, which pushes the per-block E4M3 scale factors below their +smallest subnormal, so the NVFP4 intermediate quantizes to *exactly zero* and +the MoE returns all zeros. + +That failure mode is the reason these tests assert a relative L2 error and a +nonzero-element count rather than only ``check_accuracy``'s tolerance band: an +all-zero output has ``relL2 == 1.0`` but still *passes* a +``rtol=0.1, atol=0.15, percent=0.97`` band, because these MoE outputs have a +standard deviation (~0.04) well below ``atol``. +""" + +import os + +import pytest +import torch +from _torch.modules.moe.quantize_utils import get_test_quant_params +from transformers import PretrainedConfig +from utils.util import getSMVersion + +from tensorrt_llm._torch.model_config import ModelConfig +from tensorrt_llm._torch.modules.fused_moe.fused_moe_trtllm_gen import TRTLLMGenFusedMoE +from tensorrt_llm._torch.modules.fused_moe.impl_contract import MoECommPlan, MoERunContext +from tensorrt_llm._torch.modules.fused_moe.interface import MoEWeightLoadingMode +from tensorrt_llm._torch.modules.fused_moe.quantization import NVFP4TRTLLMGenFusedMoEMethod +from tensorrt_llm._torch.modules.fused_moe.routing import RenormalizeMoeRoutingMethod +from tensorrt_llm._torch.modules.kimi_k3_moe._mlp import SituAndMul +from tensorrt_llm._torch.utils import ActivationType, ActType_TrtllmGen +from tensorrt_llm.mapping import Mapping +from tensorrt_llm.quantization.mode import QuantAlgo + +situ_supported = pytest.mark.skipif( + getSMVersion() not in (100, 103), + reason="The NVFP4 SiTU cubins only support SM100/SM103. Current SM is %d." % getSMVersion(), +) + +# Kimi K3 defaults. ``trtllm_gen_activation_alpha`` is the gate-side SiTU beta +# (the cubin's ``alpha``); ``trtllm_gen_activation_beta`` is the linear-side +# SiTU beta (the cubin's ``beta``). See modeling_kimi_linear.py. +GATE_ALPHA = 4.0 +LINEAR_BETA = 25.0 + + +def _build_backend( + *, num_experts, top_k, hidden_size, intermediate_size, quantize_util, quant_config, situ +): + """A single-rank TRTLLM-Gen NVFP4 MoE, with or without the SiTU override.""" + pretrained_config = PretrainedConfig() + pretrained_config.num_experts = num_experts + pretrained_config.hidden_size = hidden_size + pretrained_config.intermediate_size = intermediate_size + pretrained_config.torch_dtype = torch.bfloat16 + model_config = ModelConfig( + pretrained_config=pretrained_config, + quant_config=quant_config, + mapping=Mapping(world_size=1, tp_size=1, rank=0), + moe_backend="TRTLLM", + ) + situ_kwargs = {} + if situ: + situ_kwargs = dict( + trtllm_gen_activation_type=ActType_TrtllmGen.SiTu, + trtllm_gen_activation_alpha=GATE_ALPHA, + trtllm_gen_activation_beta=LINEAR_BETA, + ) + return TRTLLMGenFusedMoE( + routing_method=RenormalizeMoeRoutingMethod(top_k=top_k), + num_experts=num_experts, + hidden_size=hidden_size, + intermediate_size=intermediate_size, + dtype=torch.bfloat16, + reduce_results=True, + model_config=model_config, + init_load_balancer=False, + weight_loading_mode=MoEWeightLoadingMode.VANILLA, + # SiTU rides the generic SwiGLU geometry; the fused activation is + # selected by trtllm_gen_activation_type, not by activation_type. + activation_type=ActivationType.Swiglu, + **situ_kwargs, + ) + + +def _make_case(num_experts, top_k, hidden_size, intermediate_size, seq_len, situ, seed=0): + """Backend + reference module sharing one set of NVFP4 weights.""" + torch.manual_seed(seed) + device = "cuda" + x = torch.randn((seq_len, hidden_size), dtype=torch.bfloat16, device=device) * 0.5 + router_logits = torch.randn((seq_len, num_experts), dtype=torch.bfloat16, device=device) + + util_cls, quant_config, quant_kwargs = get_test_quant_params(QuantAlgo.NVFP4, x, "TRTLLM") + quant_kwargs.pop("ref_cls", None) + quantize_util = util_cls( + num_experts=num_experts, + dtype=torch.bfloat16, + intermediate_size=intermediate_size, + hidden_size=hidden_size, + quant_config=quant_config, + bias=False, + activation_type=ActivationType.Swiglu, + ) + weights = quantize_util.create_weights(**quant_kwargs) + + backend = _build_backend( + num_experts=num_experts, + top_k=top_k, + hidden_size=hidden_size, + intermediate_size=intermediate_size, + quantize_util=quantize_util, + quant_config=quant_config, + situ=situ, + ) + backend.load_weights([weights]) + backend.post_load_weights() + backend.cuda() + return backend, quantize_util, weights, x, router_logits + + +def _run_backend(backend, x, router_logits, top_k): + routing_method = RenormalizeMoeRoutingMethod(top_k=top_k) + with torch.inference_mode(): + selected, scales = routing_method.apply(router_logits) + x_quantized, x_sf = backend.quantize_input(x, post_quant_comm=False) + return backend.run_moe( + MoERunContext( + x=x_quantized, + x_sf=x_sf, + token_selected_experts=selected.to(torch.int32), + token_final_scales=scales.to(torch.bfloat16), + # Single GPU, no comm strategy: quantize_input ran locally and + # left the scale factors swizzled. + comm_plan=MoECommPlan( + input_sf_swizzled=True, + enable_alltoall=False, + moe_output=None, + payload_in_workspace=False, + ), + ) + ) + + +def _reference(quantize_util, weights, x, router_logits, top_k, situ): + ref = quantize_util.create_ref_module(RenormalizeMoeRoutingMethod(top_k=top_k)) + if situ: + # The in-tree SiTU activation is the source of truth: its ``beta`` is + # the gate side and ``linear_beta`` the linear side, matching how + # modeling_kimi_linear.py maps them onto the cubin's alpha/beta. + activation = SituAndMul(beta=GATE_ALPHA, linear_beta=LINEAR_BETA) + for expert in ref.experts: + expert.activation = activation + ref.load_weights([weights]) + ref.cuda() + with torch.inference_mode(): + return ref.forward(x, router_logits) + + +def _rel_l2(actual: torch.Tensor, reference: torch.Tensor) -> float: + """||actual - reference|| / ||reference||. An all-zero actual gives 1.0.""" + return ((actual.float() - reference.float()).norm() / reference.float().norm()).item() + + +@situ_supported +@pytest.mark.parametrize("num_tokens", [1, 8, 512]) +def test_nvfp4_situ_runner_has_valid_configs(num_tokens): + """The NVFP4 SiTU cubins must be reachable through tactic selection.""" + runner = torch.classes.trtllm.FP4BlockScaleMoERunner(int(ActType_TrtllmGen.SiTu)) + # (topK, hiddenSize, intermediateSize, numLocalExperts, numTokens) + tactics = runner.get_valid_configs(2, 512, 256, 8, num_tokens) + assert tactics, f"no valid NVFP4 SiTU tactic for num_tokens={num_tokens}" + + +@situ_supported +def test_nvfp4_situ_fc31_scale_c_drops_dequant_scale(): + """SiTuGlu is nonlinear in x0, so scaleC must be quantScaleC alone. + + Pins the trtllm-gen rule (``!isLinearInX0(mActType) && mFusedAct`` in + ``BatchedGemm/BatchedGemmTestUtils.h``) at the level where TRT-LLM + computes it, without needing to run a kernel. SwiGLU is the control: it + *is* linear in x0 and keeps the combined ``quantScaleC * dequantScaleAb``. + """ + shape = dict(num_experts=8, top_k=2, hidden_size=512, intermediate_size=256, seq_len=8) + situ_backend, *_ = _make_case(**shape, situ=True) + swiglu_backend, *_ = _make_case(**shape, situ=False) + + # fc2_input_scale is quantScaleC; fc31_alpha is dequantScaleAb (the + # scaleGate the kernel already folds into the tanh/sigmoid arguments). + torch.testing.assert_close( + situ_backend.fc31_scale_c.data.float(), + situ_backend.fc2_input_scale.data.float().expand_as(situ_backend.fc31_scale_c.data), + msg="SiTU fc31_scale_c must not carry the dequantScaleAb factor", + ) + torch.testing.assert_close( + swiglu_backend.fc31_scale_c.data.float(), + (swiglu_backend.fc2_input_scale.data * swiglu_backend.fc31_alpha.data).float(), + msg="SwiGLU fc31_scale_c must keep the combined scale", + ) + # The two conventions must actually differ, otherwise this test would pass + # vacuously on a build where dequantScaleAb happens to be 1. + assert not torch.allclose( + situ_backend.fc31_scale_c.data.float(), swiglu_backend.fc31_scale_c.data.float() + ) + + +@situ_supported +@pytest.mark.parametrize( + "num_experts,top_k,hidden_size,intermediate_size", + [ + pytest.param(8, 2, 512, 256, id="e8_k2_h512_i256"), + pytest.param(8, 2, 2048, 1024, id="e8_k2_h2048_i1024"), + pytest.param(256, 8, 2048, 1024, id="e256_k8_h2048_i1024"), + ], +) +@pytest.mark.parametrize("seq_len", [1, 16, 256], ids=lambda n: f"tokens{n}") +def test_nvfp4_situ_matches_reference(num_experts, top_k, hidden_size, intermediate_size, seq_len): + backend, quantize_util, weights, x, router_logits = _make_case( + num_experts, top_k, hidden_size, intermediate_size, seq_len, situ=True + ) + output = _run_backend(backend, x, router_logits, top_k) + reference = _reference(quantize_util, weights, x, router_logits, top_k, situ=True) + + # Regression guard for the all-zero FC1 output: check_accuracy's tolerance + # band alone passes on a zero tensor here (ref std ~0.04 << atol 0.15). + nonzero = int((output != 0).sum()) + assert nonzero > 0.99 * output.numel(), ( + f"NVFP4 SiTU produced a near-empty output ({nonzero}/{output.numel()} " + "nonzero) -- the FC1 intermediate most likely underflowed to zero" + ) + + rel_l2 = _rel_l2(output, reference) + assert rel_l2 < 0.1, ( + f"NVFP4 SiTU relative L2 error {rel_l2:.4f} is too large " + "(an output uncorrelated with the reference gives ~1.0)" + ) + + quantize_util.create_ref_module(RenormalizeMoeRoutingMethod(top_k=top_k)).check_accuracy( + output, reference + ) + + +_LAUNCH_EVIDENCE_SCRIPT = r""" +import torch +from test_nvfp4_situ_moe import _make_case, _run_backend, _rel_l2, _reference + +backend, quantize_util, weights, x, router_logits = _make_case( + 8, 2, 512, 256, 16, situ=True) +out = _run_backend(backend, x, router_logits, 2) +ref = _reference(quantize_util, weights, x, router_logits, 2, situ=True) +torch.cuda.synchronize() +assert int((out != 0).sum()) > 0, "all-zero output" +assert _rel_l2(out, ref) < 0.1, "reference mismatch" +""" + + +@situ_supported +def test_nvfp4_situ_launches_situ_cubin(): + """The FC1 kernel actually selected must be an NVFP4 siTuGlu cubin. + + Runs in a subprocess because the C++ logger level is fixed at process + start (TLLM_LOG_LEVEL) and TLLM_BATCHED_GEMM_PRINT_NAME logs at INFO. + """ + import subprocess + import sys + + env = dict(os.environ) + env["TLLM_BATCHED_GEMM_PRINT_NAME"] = "1" + env["TLLM_LOG_LEVEL"] = "INFO" + this_dir = os.path.dirname(os.path.abspath(__file__)) + unittest_root = os.path.abspath(os.path.join(this_dir, "..", "..", "..")) + env["PYTHONPATH"] = os.pathsep.join([this_dir, unittest_root, env.get("PYTHONPATH", "")]) + result = subprocess.run( + [sys.executable, "-c", _LAUNCH_EVIDENCE_SCRIPT], + capture_output=True, + text=True, + env=env, + cwd=this_dir, + timeout=600, + ) + log = result.stdout + result.stderr + assert result.returncode == 0, f"fused forward failed:\n{log[-4000:]}" + assert "siTuGlu" in log, ( + "expected the FC1 launch log to name a siTuGlu kernel; got:\n" + log[-4000:] + ) + # E2m1 in/out is what distinguishes the NVFP4 family from the MXFP4 one. + assert "E2m1_E2m1E2m1" in log or "e2m1_e2m1e2m1" in log.lower(), ( + "expected the NVFP4 (E2m1_E2m1E2m1) SiTU cubin family; got:\n" + log[-4000:] + ) + + +@situ_supported +def test_nvfp4_situ_padded_quant_method_is_selected(): + """SiTU must pick the padded NVFP4 method regardless of call ordering. + + ``_get_quant_method`` keys off ``is_situ_activation`` rather than + ``swiglu_alpha is not None`` precisely because create_weights fills the + alpha/beta slots afterwards. + """ + backend, *_ = _make_case(8, 2, 512, 256, 8, situ=True) + assert isinstance(backend.quant_method, NVFP4TRTLLMGenFusedMoEMethod) + assert backend.is_situ_activation + assert backend.scaling_vector_size == 16 diff --git a/tests/unittest/_torch/thop/serial/test_moe.py b/tests/unittest/_torch/thop/serial/test_moe.py index 8248d9d4d7f3..6d6369740e7c 100644 --- a/tests/unittest/_torch/thop/serial/test_moe.py +++ b/tests/unittest/_torch/thop/serial/test_moe.py @@ -66,6 +66,23 @@ def test_mxe4m3_mxe2m1_situ_runner_has_valid_configs(num_tokens): assert tactics, f"No valid SiTu tactic for num_tokens={num_tokens}" +@pytest.mark.skipif( + getSMVersion() not in (100, 103), + reason="The SiTu kernel only supports SM100/SM103. Current SM is %d." % + getSMVersion(), +) +@pytest.mark.parametrize("num_tokens", [1, 8, 512]) +def test_e2m1_e2m1_situ_runner_has_valid_configs(num_tokens): + """NVFP4 twin of the MXFP4 check above. + + Numerics for this path live in + tests/unittest/_torch/modules/moe/test_nvfp4_situ_moe.py. + """ + runner = torch.classes.trtllm.FP4BlockScaleMoERunner(ActType.SiTu.value) + tactics = runner.get_valid_configs(2, 512, 256, 8, num_tokens) + assert tactics, f"No valid NVFP4 SiTu tactic for num_tokens={num_tokens}" + + class moe_args: def __init__(self, From 6b7be7a59c74a1e1c45f35929dead8a9c77c3d2e Mon Sep 17 00:00:00 2001 From: rosong11 Date: Mon, 24 Aug 2026 10:23:43 +0000 Subject: [PATCH 2/4] [None][chore] Merge NVFP4 SiTu tests into the existing MoE test modules Signed-off-by: rosong11 --- .../test_lists/test-db/l0_b200.yml | 3 + .../test_lists/test-db/l0_b300.yml | 9 +- .../modules/moe/test_kimi_k3_situ_moe.py | 167 +++++++--- .../_torch/modules/moe/test_moe_backend.py | 186 ++++++++++- .../_torch/modules/moe/test_nvfp4_situ_moe.py | 313 ------------------ tests/unittest/_torch/thop/serial/test_moe.py | 37 ++- 6 files changed, 343 insertions(+), 372 deletions(-) delete mode 100644 tests/unittest/_torch/modules/moe/test_nvfp4_situ_moe.py diff --git a/tests/integration/test_lists/test-db/l0_b200.yml b/tests/integration/test_lists/test-db/l0_b200.yml index 6f0b5c320cc9..47c31ce67519 100644 --- a/tests/integration/test_lists/test-db/l0_b200.yml +++ b/tests/integration/test_lists/test-db/l0_b200.yml @@ -146,6 +146,9 @@ l0_b200: - unittest/_torch/modules/moe/test_moe_backend.py::test_megamoe_deepgemm_infers_kimi_situ_from_pretrained_config - unittest/_torch/modules/moe/test_moe_backend.py::test_megamoe_deepgemm_defaults_to_swiglu_without_situ_config - unittest/_torch/modules/moe/test_moe_backend.py::test_create_moe_forwards_megamoe_activation_options + - unittest/_torch/modules/moe/test_moe_backend.py::test_trtllm_gen_nvfp4_situ_selects_padded_quant_method + - unittest/_torch/modules/moe/test_moe_backend.py::test_trtllm_gen_nvfp4_situ_fc31_scale_c_drops_dequant_scale + - unittest/_torch/modules/moe/test_moe_backend.py::test_trtllm_gen_situ_rejects_quant_algos_without_fused_cubins # ------------- MoE: test_single_gpu (by backend) --------------- - unittest/_torch/modules/moe/test_moe_module.py::test_configurable_moe_single_gpu -k "CUTLASS and not None" - unittest/_torch/modules/moe/test_moe_module.py::test_configurable_moe_single_gpu -k "TRTLLM" diff --git a/tests/integration/test_lists/test-db/l0_b300.yml b/tests/integration/test_lists/test-db/l0_b300.yml index 9f5f64f93547..4e8996d43ba1 100644 --- a/tests/integration/test_lists/test-db/l0_b300.yml +++ b/tests/integration/test_lists/test-db/l0_b300.yml @@ -55,8 +55,10 @@ l0_b300: - unittest/_torch/modules/moe/test_kimi_k3_situ_moe.py::test_nvfp4_streaming_drains_staging_per_expert - unittest/_torch/modules/moe/test_kimi_k3_situ_moe.py::test_nvfp4_streamed_experts_forward_runs - unittest/_torch/modules/moe/test_kimi_k3_situ_moe.py::test_cutlass_situ_bf16_matches_reference - - unittest/_torch/modules/moe/test_kimi_k3_situ_moe.py::test_nvfp4_kernel_actually_applies_situ - - unittest/_torch/modules/moe/test_kimi_k3_situ_moe.py::test_nvfp4_streamed_experts_match_situ_reference[static_1.0] + - unittest/_torch/modules/moe/test_kimi_k3_situ_moe.py::test_nvfp4_kernel_actually_applies_situ[CUTLASS] + - unittest/_torch/modules/moe/test_kimi_k3_situ_moe.py::test_nvfp4_kernel_actually_applies_situ[TRTLLM] + - unittest/_torch/modules/moe/test_kimi_k3_situ_moe.py::test_nvfp4_experts_match_situ_reference[CUTLASS-static_1.0] + - unittest/_torch/modules/moe/test_kimi_k3_situ_moe.py::test_nvfp4_experts_match_situ_reference[TRTLLM-static_1.0] - unittest/_torch/modules/moe/test_kimi_k3_situ_moe.py::test_megamoe_streamed_coverage_survives_per_expert_drain - unittest/_torch/modules/moe/test_kimi_k3_situ_moe.py::test_megamoe_overrides_finalize_streamed_expert - unittest/_torch/modules/moe/test_kimi_k3_situ_moe.py::test_mega_format_transform_is_slot_blockwise @@ -64,6 +66,9 @@ l0_b300: - unittest/_torch/modules/moe/test_moe_backend.py::test_trtllm_bf16_unquantized_moe - unittest/_torch/modules/moe/test_moe_backend.py::test_trtllm_fp8_block_scales_fused_shared_experts - unittest/_torch/modules/moe/test_moe_backend.py::test_trtllm_fp8_block_scales_fuse_shared_expert_layout + - unittest/_torch/modules/moe/test_moe_backend.py::test_trtllm_gen_nvfp4_situ_selects_padded_quant_method + - unittest/_torch/modules/moe/test_moe_backend.py::test_trtllm_gen_nvfp4_situ_fc31_scale_c_drops_dequant_scale + - unittest/_torch/modules/moe/test_moe_backend.py::test_trtllm_gen_situ_rejects_quant_algos_without_fused_cubins # ------------- MoE: test_single_gpu (specific quant per backend) --------------- # CUTLASS backend: FP8, NVFP4, W4A8_MXFP4_MXFP8, W8A16 - unittest/_torch/modules/moe/test_moe_module.py::test_configurable_moe_single_gpu[e60_k4_h2048_i1408-seq=1-dtype=torch.bfloat16-backend=CUTLASS-quant=FP8-routing=Renormalize] diff --git a/tests/unittest/_torch/modules/moe/test_kimi_k3_situ_moe.py b/tests/unittest/_torch/modules/moe/test_kimi_k3_situ_moe.py index e3fca03ca9d3..3a2c979feb09 100644 --- a/tests/unittest/_torch/modules/moe/test_kimi_k3_situ_moe.py +++ b/tests/unittest/_torch/modules/moe/test_kimi_k3_situ_moe.py @@ -6,8 +6,11 @@ * runner-local ActType numeric stability (SwiGlu/Relu2/Silu unchanged, SiTu appended); -* the native SiTU runner returns valid tactics and actually launches a - kernel whose name contains ``siTuGlu``; +* the native SiTU runner actually launches a kernel whose name contains + ``siTuGlu``, and from the expected cubin family -- ``MxE4m3_MxE2m1`` for + the MXFP4 drop, ``E2m1_E2m1E2m1`` for the NVFP4 one (tactic availability + for both families is pinned at the op level in + tests/unittest/_torch/thop/serial/test_moe.py); * fused output matches the Python ``SituAndMul`` reference within MXFP8/MXFP4 quantization tolerance, for the default AND asymmetric non-default ``activation_situ_beta`` / ``activation_situ_linear_beta``; @@ -17,6 +20,12 @@ survive the fused path; * the fused path fails loudly without loaded weights (no silent random-weight fallback). + +The NVFP4 half of SiTU lives here too, on both FP4 backends: CUTLASS takes +SiTU as an ``ActivationType``, TRTLLM-Gen as an out-of-band +``trtllm_gen_activation_type`` served by the fused +``Bmm_E2m1_E2m1E2m1_..._siTuGlu_*`` FC1 cubins. Tests that apply to both are +parametrized over ``moe_backend`` rather than duplicated. """ import dataclasses @@ -268,17 +277,9 @@ def test_make_situ_alpha_beta_contract(): ) -@situ_supported -def test_situ_runner_returns_valid_tactics(): - runner = torch.classes.trtllm.MxE4m3MxE2m1BlockScaleMoERunner(int(ActType_TrtllmGen.SiTu), True) - # Representative low-latency and throughput shapes (topK, hidden, - # intermediate, localExperts, numTokens, validHidden, validIntermediate). - for num_tokens in (1, 8, 512): - tactics = runner.get_valid_configs(2, 512, 256, 8, num_tokens, 512, 256) - assert len(tactics) > 0, f"no valid SiTU tactic for num_tokens={num_tokens}" - - -_LAUNCH_EVIDENCE_SCRIPT = r""" +_LAUNCH_EVIDENCE_SCRIPTS = { + # MXFP4: the K3 fused block (MXFP8 activations x group-32 MXFP4 weights). + "mxfp4": r""" import torch from test_kimi_k3_situ_moe import _K3Config, _make_block_pair @@ -289,12 +290,46 @@ def test_situ_runner_returns_valid_tactics(): fused(x) torch.cuda.synchronize() assert fused._cubin_call_count == 1 -""" +""", + # NVFP4: the routed MoE on the TRTLLM-Gen backend (group-16 scales). + "nvfp4": r""" +import torch +from test_kimi_k3_situ_moe import ( + _TP_EXPERTS, + _TP_HIDDEN, + _TP_INTERMEDIATE, + _load_nvfp4_bank_for, + _make_nvfp4_expert_bank, + _make_nvfp4_moe, + _make_test_gate, +) + +gate = _make_test_gate() +moe = _make_nvfp4_moe(gate, moe_backend="TRTLLM") +bank = _make_nvfp4_expert_bank(_TP_EXPERTS, _TP_INTERMEDIATE, _TP_HIDDEN) +_load_nvfp4_bank_for(moe, bank, "TRTLLM") + +torch.manual_seed(5) +x = torch.randn(16, _TP_HIDDEN, dtype=torch.bfloat16, device="cuda") * 0.5 +out = moe.forward(x, gate.compute_logits(x), all_rank_num_tokens=None) +torch.cuda.synchronize() +assert torch.isfinite(out).all() +""", +} + +# What separates the two fused SiTu FC1 cubin families in the launch log. +_SITU_CUBIN_FAMILY = {"mxfp4": "MxE4m3_MxE2m1", "nvfp4": "E2m1_E2m1E2m1"} @situ_supported -def test_fused_forward_launches_situ_kernel(): - """Launch evidence: the FC1 kernel actually selected must be a siTuGlu cubin. +@pytest.mark.parametrize("fmt", ["mxfp4", "nvfp4"]) +def test_fused_forward_launches_situ_kernel(fmt): + """Launch evidence: the FC1 kernel selected must be a siTuGlu cubin from + the expected quantization family. + + ``siTuGlu`` alone only proves that *some* SiTu kernel ran. The family + substring is what separates the MXFP4 drop from the NVFP4 one, so a path + silently resolving to the other family still fails here. Runs in a subprocess because the C++ logger level is fixed at process start (TLLM_LOG_LEVEL) and TLLM_BATCHED_GEMM_PRINT_NAME logs at INFO. @@ -309,11 +344,11 @@ def test_fused_forward_launches_situ_kernel(): unittest_root = os.path.abspath(os.path.join(this_dir, "..", "..", "..")) env["PYTHONPATH"] = os.pathsep.join([this_dir, unittest_root, env.get("PYTHONPATH", "")]) result = subprocess.run( - [sys.executable, "-c", _LAUNCH_EVIDENCE_SCRIPT], + [sys.executable, "-c", _LAUNCH_EVIDENCE_SCRIPTS[fmt]], capture_output=True, text=True, env=env, - cwd=os.path.dirname(os.path.abspath(__file__)), + cwd=this_dir, timeout=600, ) log = result.stdout + result.stderr @@ -321,6 +356,10 @@ def test_fused_forward_launches_situ_kernel(): assert "siTuGlu" in log, ( "expected the FC1 launch log to name a siTuGlu kernel; got:\n" + log[-4000:] ) + family = _SITU_CUBIN_FAMILY[fmt] + assert family.lower() in log.lower(), ( + f"expected a {fmt.upper()} ({family}) SiTu cubin; got:\n" + log[-4000:] + ) @situ_supported @@ -956,14 +995,22 @@ def block_scales(*shape): return bank -def _make_nvfp4_moe(gate, num_experts=_TP_EXPERTS): +def _make_nvfp4_moe(gate, num_experts=_TP_EXPERTS, moe_backend="CUTLASS"): + """NVFP4 + SiTU routed MoE on either FP4 backend. + + CUTLASS takes SiTU as an ``ActivationType``; TRTLLM-Gen carries it out of + band as ``trtllm_gen_activation_type`` and serves it with the fused + ``Bmm_E2m1_E2m1E2m1_..._siTuGlu_*`` FC1 cubins (group-16 block scales). + ``_make_routed_moe`` already mirrors both of KimiK3MoERuntime's branches, + so the backend is the only variable. + """ from tensorrt_llm.models.modeling_utils import QuantAlgo, QuantConfig return _make_routed_moe( _TP_INTERMEDIATE, gate, num_experts=num_experts, - moe_backend="CUTLASS", + moe_backend=moe_backend, routed_quant_config=QuantConfig(quant_algo=QuantAlgo.NVFP4, group_size=_NVFP4_GROUP_SIZE), ) @@ -1016,6 +1063,26 @@ def _load_nvfp4_bank_whole(moe, bank): return moe.backend +def _load_nvfp4_bank_for(moe, bank, moe_backend): + """Load a bank and run the post-load transform, ready for ``forward``. + + CUTLASS goes through K3's per-expert streaming adapter -- the path the real + checkpoint loader takes, and the one the buffer-equality tests above pin. + TRTLLM-Gen's NVFP4 quant method does not override + ``prepare_streaming_expert_load``, so it takes the stock whole-checkpoint + path instead; the loader is not what these tests are about, the kernel is. + """ + if moe_backend == "CUTLASS": + backend = _stream_nvfp4_bank(moe, bank) + # _stream_nvfp4_bank already ran process_weights_after_loading, which + # the whole-checkpoint path leaves to post_load_weights below. + backend._weights_transformed = False + else: + backend = _load_nvfp4_bank_whole(moe, bank) + moe.post_load_weights() + return backend + + _NVFP4_LOADED_STATE = ( "w3_w1_weight", "w2_weight", @@ -1273,11 +1340,17 @@ def q(w, g): @nvfp4_moe_supported @pytest.mark.parametrize( - "input_scale", + "moe_backend,input_scale", [ - 1.0, + pytest.param("CUTLASS", 1.0, id="CUTLASS-static_1.0"), + pytest.param("TRTLLM", 1.0, id="TRTLLM-static_1.0"), + # derived_act_scale is a CUTLASS-path finding (see the reason below), + # so it stays pinned to CUTLASS rather than becoming a strict xfail the + # TRTLLM-Gen path would have to reproduce. pytest.param( + "CUTLASS", None, + id="CUTLASS-derived_act_scale", marks=pytest.mark.xfail( strict=True, reason="OPEN: the conventional NVFP4 activation global scale is the WORSE " @@ -1294,15 +1367,22 @@ def q(w, g): ), ), ], - ids=["static_1.0", "derived_act_scale"], ) -def test_nvfp4_streamed_experts_match_situ_reference(input_scale): - """Streamed NVFP4 experts against the golden activation. +def test_nvfp4_experts_match_situ_reference(moe_backend, input_scale): + """NVFP4 SiTU experts against the golden activation, on both FP4 backends. The streamed-vs-whole-checkpoint tests compare two loaders that share all the scale handling, so they agree whether or not that handling is right. This compares against the definition instead. + ``moe_backend`` is the discriminating dimension: CUTLASS runs SiTU as an + ``ActivationType``, TRTLLM-Gen runs the fused ``E2m1_E2m1E2m1 siTuGlu`` + FC1 cubins. The latter is the one that folds ``dequantScaleAb`` into the + tanh/sigmoid arguments instead of into scaleC; getting that wrong drives + the FC1 block scales below their smallest subnormal and the MoE returns + exact zeros, which the ``cosine > 0.95`` bound below rejects (an all-zero + output scores 0). + ``static_1.0`` is what nvidia/Kimi-K3-NVFP4 actually ships for ``input_scale``; ``derived_act_scale`` is the conventional value computed from the activations. Splitting them separates "the loader is wrong" from @@ -1329,10 +1409,8 @@ def test_nvfp4_streamed_experts_match_situ_reference(input_scale): ] bank = [_quantize_expert_to_nvfp4(w1[e], w2[e], w3[e], act_scale) for e in range(num_experts)] - moe = _make_nvfp4_moe(gate, num_experts=num_experts) - _stream_nvfp4_bank(moe, bank) - moe.backend._weights_transformed = False - moe.post_load_weights() + moe = _make_nvfp4_moe(gate, num_experts=num_experts, moe_backend=moe_backend) + _load_nvfp4_bank_for(moe, bank, moe_backend) router_logits = gate.compute_logits(x) actual = moe.forward(x, router_logits, all_rank_num_tokens=None).float() @@ -1343,7 +1421,8 @@ def test_nvfp4_streamed_experts_match_situ_reference(input_scale): cosine = torch.nn.functional.cosine_similarity(actual.flatten(), expected.flatten(), dim=0) rel_l2 = torch.linalg.vector_norm(actual - expected) / torch.linalg.vector_norm(expected) print( - f"nvfp4-streamed[{input_scale}] vs reference: cosine={cosine.item():.6f} rel_l2={rel_l2.item():.6f}" + f"nvfp4[{moe_backend}][{input_scale}] vs reference: " + f"cosine={cosine.item():.6f} rel_l2={rel_l2.item():.6f}" ) # The shipped configuration measures 0.9727 / 0.234. That residual is # activation quantization: this path takes the ACTIVATIONS to FP4 as well @@ -1382,7 +1461,8 @@ def _swiglu_reference_moe(x, router_logits, routing_method, w1, w2, w3, alpha, b @nvfp4_moe_supported -def test_nvfp4_kernel_actually_applies_situ(): +@pytest.mark.parametrize("moe_backend", ["CUTLASS", "TRTLLM"]) +def test_nvfp4_kernel_actually_applies_situ(moe_backend): """Which activation does the QUANTIZED kernel actually run? BF16 + SiTU already matches the golden reference, so the activation, the @@ -1393,6 +1473,13 @@ def test_nvfp4_kernel_actually_applies_situ(): wrong in exactly the way the GSM8K collapse showed, while every shape-and-buffer check stayed green. + Run for both FP4 backends: CUTLASS resolves SiTU through the activation + enum, TRTLLM-Gen through a distinct fused-cubin family + (``Bmm_E2m1_E2m1E2m1_..._siTuGlu_*``). A silent fallback is a different + failure on each, and this comparison is tolerance-free, so it catches + both -- including the degenerate all-zero FC1 output, which scores 0 + against both references and so fails the assertion below. + Reported rather than merely asserted: which reference the kernel is closer to is the diagnosis. """ @@ -1416,10 +1503,8 @@ def test_nvfp4_kernel_actually_applies_situ(): ] bank = [_quantize_expert_to_nvfp4(w1[e], w2[e], w3[e], act_scale) for e in range(num_experts)] - moe = _make_nvfp4_moe(gate, num_experts=num_experts) - _stream_nvfp4_bank(moe, bank) - moe.backend._weights_transformed = False - moe.post_load_weights() + moe = _make_nvfp4_moe(gate, num_experts=num_experts, moe_backend=moe_backend) + _load_nvfp4_bank_for(moe, bank, moe_backend) router_logits = gate.compute_logits(x) actual = moe.forward(x, router_logits, all_rank_num_tokens=None).float() @@ -1439,13 +1524,15 @@ def score(ref): situ_cos, situ_l2 = score(situ) swiglu_cos, swiglu_l2 = score(swiglu) print( - f"NVFP4 kernel vs SiTU ref: cosine={situ_cos:.6f} rel_l2={situ_l2:.6f}\n" - f"NVFP4 kernel vs SwiGLU ref: cosine={swiglu_cos:.6f} rel_l2={swiglu_l2:.6f}" + f"NVFP4[{moe_backend}] kernel vs SiTU ref: " + f"cosine={situ_cos:.6f} rel_l2={situ_l2:.6f}\n" + f"NVFP4[{moe_backend}] kernel vs SwiGLU ref: " + f"cosine={swiglu_cos:.6f} rel_l2={swiglu_l2:.6f}" ) assert situ_cos > swiglu_cos, ( - f"the NVFP4 kernel matches a SwiGLU reference better than the SiTU one " - f"(situ={situ_cos:.6f}, swiglu={swiglu_cos:.6f}): the quantized path is " - f"not applying SiTU" + f"the NVFP4 {moe_backend} kernel matches a SwiGLU reference better than " + f"the SiTU one (situ={situ_cos:.6f}, swiglu={swiglu_cos:.6f}): the " + f"quantized path is not applying SiTU" ) diff --git a/tests/unittest/_torch/modules/moe/test_moe_backend.py b/tests/unittest/_torch/modules/moe/test_moe_backend.py index fd6350b9d85e..2e3eda62d25b 100644 --- a/tests/unittest/_torch/modules/moe/test_moe_backend.py +++ b/tests/unittest/_torch/modules/moe/test_moe_backend.py @@ -73,11 +73,18 @@ FusedMoEMethodBase, NVFP4FusedMoEMethod, NVFP4MarlinFusedMoEMethod, + NVFP4TRTLLMGenFusedMoEBaseMethod, + NVFP4TRTLLMGenFusedMoEMethod, UnquantizedFusedMoEMethod, W4A8MXFP4MXFP8MegaMoEDeepGemmMethod, W4A16NVFP4CutlassFusedMoEMethod, ) -from tensorrt_llm._torch.utils import ActivationType, MxFp8QuantizedTensor, is_gated_activation +from tensorrt_llm._torch.utils import ( + ActivationType, + ActType_TrtllmGen, + MxFp8QuantizedTensor, + is_gated_activation, +) from tensorrt_llm._utils import get_sm_version, mpi_rank from tensorrt_llm.mapping import Mapping from tensorrt_llm.models.modeling_utils import QuantAlgo, QuantConfig @@ -480,6 +487,183 @@ def test_marlin_override_quant_config_degrades_per_layer(): assert report.degraded_from.reason is MoERejectReason.QUANT_UNSUPPORTED +# ============================================================================ +# TRTLLM-Gen SiTu backend contract +# ============================================================================ +# SiTu rides the generic SwiGLU geometry and is carried out of band +# (trtllm_gen_activation_type, not ActivationType), so the host-side wiring is +# easy to get wrong in ways no shape check catches. Kernel-level coverage -- +# tactic availability, launch evidence and SiTu-vs-SwiGLU numerics -- lives in +# test_kimi_k3_situ_moe.py (and thop/serial/test_moe.py for the runner). These +# are the contracts that hold without running a cubin. + +_situ_supported = pytest.mark.skipif( + get_sm_version() not in (100, 103), + reason="TRTLLM-Gen SiTu cubins only exist for SM100/SM103", +) + +# Kimi K3 defaults. The gate-side SiTu beta is the cubin's ``alpha``, the +# linear-side beta its ``beta``. See modeling_kimi_linear.py. +_SITU_GATE_ALPHA = 4.0 +_SITU_LINEAR_BETA = 25.0 + + +def _make_trtllm_gen_moe( + *, + quant_config: Optional[QuantConfig], + situ: bool, + num_experts: int = 8, + top_k: int = 2, + hidden_size: int = 512, + intermediate_size: int = 256, +) -> TRTLLMGenFusedMoE: + """A single-rank TRTLLM-Gen MoE, with or without the SiTu override.""" + pretrained_config = PretrainedConfig() + pretrained_config.num_experts = num_experts + pretrained_config.hidden_size = hidden_size + pretrained_config.intermediate_size = intermediate_size + pretrained_config.torch_dtype = torch.bfloat16 + model_config = ModelConfig( + pretrained_config=pretrained_config, + quant_config=quant_config, + mapping=Mapping(world_size=1, tp_size=1, rank=0), + moe_backend="TRTLLM", + ) + situ_kwargs = ( + dict( + trtllm_gen_activation_type=ActType_TrtllmGen.SiTu, + trtllm_gen_activation_alpha=_SITU_GATE_ALPHA, + trtllm_gen_activation_beta=_SITU_LINEAR_BETA, + ) + if situ + else {} + ) + return TRTLLMGenFusedMoE( + routing_method=RenormalizeMoeRoutingMethod(top_k=top_k), + num_experts=num_experts, + hidden_size=hidden_size, + intermediate_size=intermediate_size, + dtype=torch.bfloat16, + reduce_results=True, + model_config=model_config, + init_load_balancer=False, + weight_loading_mode=MoEWeightLoadingMode.VANILLA, + # SiTu rides the generic SwiGLU geometry; the fused activation is + # selected by trtllm_gen_activation_type, not by activation_type. + activation_type=ActivationType.Swiglu, + **situ_kwargs, + ) + + +def _make_loaded_nvfp4_trtllm_gen_moe(situ: bool) -> TRTLLMGenFusedMoE: + """The same MoE with NVFP4 weights and quant scales actually loaded. + + The scale contracts below are written by ``load_quant_scales``, so they + only exist after a real load; the stock quantize utils supply the weights. + """ + num_experts, hidden_size, intermediate_size = 8, 512, 256 + torch.manual_seed(0) + x = torch.randn((8, hidden_size), dtype=torch.bfloat16, device="cuda") * 0.5 + util_cls, quant_config, quant_kwargs = get_test_quant_params(QuantAlgo.NVFP4, x, "TRTLLM") + quant_kwargs.pop("ref_cls", None) + quantize_util = util_cls( + num_experts=num_experts, + dtype=torch.bfloat16, + intermediate_size=intermediate_size, + hidden_size=hidden_size, + quant_config=quant_config, + bias=False, + activation_type=ActivationType.Swiglu, + ) + backend = _make_trtllm_gen_moe( + quant_config=quant_config, + situ=situ, + num_experts=num_experts, + hidden_size=hidden_size, + intermediate_size=intermediate_size, + ) + backend.load_weights([quantize_util.create_weights(**quant_kwargs)]) + backend.post_load_weights() + return backend.cuda() + + +@_situ_supported +def test_trtllm_gen_nvfp4_situ_selects_padded_quant_method() -> None: + """``_get_quant_method`` must key off ``is_situ_activation``. + + SiTu fills the swiglu_alpha/swiglu_beta slots from ``create_weights``, + i.e. after ``_get_quant_method`` has already run, so keying off + ``swiglu_alpha is not None`` would make the selected method depend on + *when* it is resolved. Plain SwiGLU is the control: it still gets the + unpadded base method, so this is not asserting a constant. + """ + situ = _make_loaded_nvfp4_trtllm_gen_moe(situ=True) + assert situ.is_situ_activation + assert isinstance(situ.quant_method, NVFP4TRTLLMGenFusedMoEMethod) + assert situ.scaling_vector_size == 16 + + plain = _make_loaded_nvfp4_trtllm_gen_moe(situ=False) + assert not plain.is_situ_activation + assert type(plain.quant_method) is NVFP4TRTLLMGenFusedMoEBaseMethod + + +@_situ_supported +def test_trtllm_gen_nvfp4_situ_fc31_scale_c_drops_dequant_scale() -> None: + """SiTuGlu is nonlinear in x0, so scaleC must be quantScaleC alone. + + Pins the trtllm-gen host rule (``!isLinearInX0(mActType) && mFusedAct`` in + ``BatchedGemm/BatchedGemmTestUtils.h``) at the level where TRT-LLM + computes it, without needing to run a kernel. SwiGLU is the control: it + *is* linear in x0 and keeps the combined ``quantScaleC * dequantScaleAb``. + + Getting this wrong is not a small numeric error. The extra dequantScaleAb + factor (~6e-8) drives the FC1 output's per-block E4M3 scale factors below + their smallest subnormal, so the NVFP4 intermediate quantizes to exactly + zero and the whole MoE returns zeros -- an output that still passes a + ``rtol=0.1, atol=0.15`` tolerance band, because these MoE outputs have a + standard deviation (~0.04) well below ``atol``. + """ + situ = _make_loaded_nvfp4_trtllm_gen_moe(situ=True) + swiglu = _make_loaded_nvfp4_trtllm_gen_moe(situ=False) + + # fc2_input_scale is quantScaleC; fc31_alpha is dequantScaleAb (the + # scaleGate the kernel already folds into the tanh/sigmoid arguments). + torch.testing.assert_close( + situ.fc31_scale_c.data.float(), + situ.fc2_input_scale.data.float().expand_as(situ.fc31_scale_c.data), + msg="SiTu fc31_scale_c must not carry the dequantScaleAb factor", + ) + torch.testing.assert_close( + swiglu.fc31_scale_c.data.float(), + (swiglu.fc2_input_scale.data * swiglu.fc31_alpha.data).float(), + msg="SwiGLU fc31_scale_c must keep the combined scale", + ) + # The two conventions must actually differ, otherwise this test would pass + # vacuously on a build where dequantScaleAb happens to be 1. + assert not torch.allclose(situ.fc31_scale_c.data.float(), swiglu.fc31_scale_c.data.float()) + + +@_situ_supported +@pytest.mark.parametrize( + "quant_algo", + [ + pytest.param(None, id="unquantized"), + pytest.param(QuantAlgo.FP8, id="fp8"), + pytest.param(QuantAlgo.W4A16_MXFP4, id="w4a16_mxfp4"), + ], +) +def test_trtllm_gen_situ_rejects_quant_algos_without_fused_cubins(quant_algo) -> None: + """There is no standalone SiTu activation kernel. + + SiTu exists only as a fused FC1 epilogue, in the NVFP4 (group-16) and + W4A8_MXFP4_MXFP8 (group-32) cubin families. Anything else has to be + rejected at construction rather than silently resolving to SwiGLU, which + is structurally wrong output that no shape check would catch. + """ + with pytest.raises(ValueError, match="requires one of .* quantization"): + _make_trtllm_gen_moe(quant_config=QuantConfig(quant_algo=quant_algo), situ=True) + + def test_megamoe_cutedsl_post_load_weights_uses_staged_hooks(): moe = MegaMoECuteDsl.__new__(MegaMoECuteDsl) torch.nn.Module.__init__(moe) diff --git a/tests/unittest/_torch/modules/moe/test_nvfp4_situ_moe.py b/tests/unittest/_torch/modules/moe/test_nvfp4_situ_moe.py deleted file mode 100644 index 488fdd4e38ee..000000000000 --- a/tests/unittest/_torch/modules/moe/test_nvfp4_situ_moe.py +++ /dev/null @@ -1,313 +0,0 @@ -# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. -# SPDX-License-Identifier: Apache-2.0 -"""NVFP4 + TRTLLM-Gen SiTU MoE tests. - -SiTU on NVFP4 is served by the fused ``Bmm_E2m1_E2m1E2m1_..._siTuGlu_*`` FC1 -cubins (group-16 block scales); the MXFP4 variant is covered separately by -``test_kimi_k3_situ_moe.py``. The two paths differ in one way that is easy to -get wrong: NVFP4 carries *global* scale scalars into the kernel -(``output1_scale_scalar`` / ``output1_scale_gate_scalar``) while MXFP4 relies -purely on block scales. - -Because SiTuGlu is not linear in x0 (x0 feeds ``tanh(x0 / beta)``), trtllm-gen -requires the kernel's scaleC to be ``quantScaleC`` alone -- ``dequantScaleAb`` -is folded into the tanh/sigmoid arguments via scaleGate instead. Passing the -SwiGlu-style combined ``dequantScaleAb * quantScaleC`` multiplies the FC1 -output by ~6e-8, which pushes the per-block E4M3 scale factors below their -smallest subnormal, so the NVFP4 intermediate quantizes to *exactly zero* and -the MoE returns all zeros. - -That failure mode is the reason these tests assert a relative L2 error and a -nonzero-element count rather than only ``check_accuracy``'s tolerance band: an -all-zero output has ``relL2 == 1.0`` but still *passes* a -``rtol=0.1, atol=0.15, percent=0.97`` band, because these MoE outputs have a -standard deviation (~0.04) well below ``atol``. -""" - -import os - -import pytest -import torch -from _torch.modules.moe.quantize_utils import get_test_quant_params -from transformers import PretrainedConfig -from utils.util import getSMVersion - -from tensorrt_llm._torch.model_config import ModelConfig -from tensorrt_llm._torch.modules.fused_moe.fused_moe_trtllm_gen import TRTLLMGenFusedMoE -from tensorrt_llm._torch.modules.fused_moe.impl_contract import MoECommPlan, MoERunContext -from tensorrt_llm._torch.modules.fused_moe.interface import MoEWeightLoadingMode -from tensorrt_llm._torch.modules.fused_moe.quantization import NVFP4TRTLLMGenFusedMoEMethod -from tensorrt_llm._torch.modules.fused_moe.routing import RenormalizeMoeRoutingMethod -from tensorrt_llm._torch.modules.kimi_k3_moe._mlp import SituAndMul -from tensorrt_llm._torch.utils import ActivationType, ActType_TrtllmGen -from tensorrt_llm.mapping import Mapping -from tensorrt_llm.quantization.mode import QuantAlgo - -situ_supported = pytest.mark.skipif( - getSMVersion() not in (100, 103), - reason="The NVFP4 SiTU cubins only support SM100/SM103. Current SM is %d." % getSMVersion(), -) - -# Kimi K3 defaults. ``trtllm_gen_activation_alpha`` is the gate-side SiTU beta -# (the cubin's ``alpha``); ``trtllm_gen_activation_beta`` is the linear-side -# SiTU beta (the cubin's ``beta``). See modeling_kimi_linear.py. -GATE_ALPHA = 4.0 -LINEAR_BETA = 25.0 - - -def _build_backend( - *, num_experts, top_k, hidden_size, intermediate_size, quantize_util, quant_config, situ -): - """A single-rank TRTLLM-Gen NVFP4 MoE, with or without the SiTU override.""" - pretrained_config = PretrainedConfig() - pretrained_config.num_experts = num_experts - pretrained_config.hidden_size = hidden_size - pretrained_config.intermediate_size = intermediate_size - pretrained_config.torch_dtype = torch.bfloat16 - model_config = ModelConfig( - pretrained_config=pretrained_config, - quant_config=quant_config, - mapping=Mapping(world_size=1, tp_size=1, rank=0), - moe_backend="TRTLLM", - ) - situ_kwargs = {} - if situ: - situ_kwargs = dict( - trtllm_gen_activation_type=ActType_TrtllmGen.SiTu, - trtllm_gen_activation_alpha=GATE_ALPHA, - trtllm_gen_activation_beta=LINEAR_BETA, - ) - return TRTLLMGenFusedMoE( - routing_method=RenormalizeMoeRoutingMethod(top_k=top_k), - num_experts=num_experts, - hidden_size=hidden_size, - intermediate_size=intermediate_size, - dtype=torch.bfloat16, - reduce_results=True, - model_config=model_config, - init_load_balancer=False, - weight_loading_mode=MoEWeightLoadingMode.VANILLA, - # SiTU rides the generic SwiGLU geometry; the fused activation is - # selected by trtllm_gen_activation_type, not by activation_type. - activation_type=ActivationType.Swiglu, - **situ_kwargs, - ) - - -def _make_case(num_experts, top_k, hidden_size, intermediate_size, seq_len, situ, seed=0): - """Backend + reference module sharing one set of NVFP4 weights.""" - torch.manual_seed(seed) - device = "cuda" - x = torch.randn((seq_len, hidden_size), dtype=torch.bfloat16, device=device) * 0.5 - router_logits = torch.randn((seq_len, num_experts), dtype=torch.bfloat16, device=device) - - util_cls, quant_config, quant_kwargs = get_test_quant_params(QuantAlgo.NVFP4, x, "TRTLLM") - quant_kwargs.pop("ref_cls", None) - quantize_util = util_cls( - num_experts=num_experts, - dtype=torch.bfloat16, - intermediate_size=intermediate_size, - hidden_size=hidden_size, - quant_config=quant_config, - bias=False, - activation_type=ActivationType.Swiglu, - ) - weights = quantize_util.create_weights(**quant_kwargs) - - backend = _build_backend( - num_experts=num_experts, - top_k=top_k, - hidden_size=hidden_size, - intermediate_size=intermediate_size, - quantize_util=quantize_util, - quant_config=quant_config, - situ=situ, - ) - backend.load_weights([weights]) - backend.post_load_weights() - backend.cuda() - return backend, quantize_util, weights, x, router_logits - - -def _run_backend(backend, x, router_logits, top_k): - routing_method = RenormalizeMoeRoutingMethod(top_k=top_k) - with torch.inference_mode(): - selected, scales = routing_method.apply(router_logits) - x_quantized, x_sf = backend.quantize_input(x, post_quant_comm=False) - return backend.run_moe( - MoERunContext( - x=x_quantized, - x_sf=x_sf, - token_selected_experts=selected.to(torch.int32), - token_final_scales=scales.to(torch.bfloat16), - # Single GPU, no comm strategy: quantize_input ran locally and - # left the scale factors swizzled. - comm_plan=MoECommPlan( - input_sf_swizzled=True, - enable_alltoall=False, - moe_output=None, - payload_in_workspace=False, - ), - ) - ) - - -def _reference(quantize_util, weights, x, router_logits, top_k, situ): - ref = quantize_util.create_ref_module(RenormalizeMoeRoutingMethod(top_k=top_k)) - if situ: - # The in-tree SiTU activation is the source of truth: its ``beta`` is - # the gate side and ``linear_beta`` the linear side, matching how - # modeling_kimi_linear.py maps them onto the cubin's alpha/beta. - activation = SituAndMul(beta=GATE_ALPHA, linear_beta=LINEAR_BETA) - for expert in ref.experts: - expert.activation = activation - ref.load_weights([weights]) - ref.cuda() - with torch.inference_mode(): - return ref.forward(x, router_logits) - - -def _rel_l2(actual: torch.Tensor, reference: torch.Tensor) -> float: - """||actual - reference|| / ||reference||. An all-zero actual gives 1.0.""" - return ((actual.float() - reference.float()).norm() / reference.float().norm()).item() - - -@situ_supported -@pytest.mark.parametrize("num_tokens", [1, 8, 512]) -def test_nvfp4_situ_runner_has_valid_configs(num_tokens): - """The NVFP4 SiTU cubins must be reachable through tactic selection.""" - runner = torch.classes.trtllm.FP4BlockScaleMoERunner(int(ActType_TrtllmGen.SiTu)) - # (topK, hiddenSize, intermediateSize, numLocalExperts, numTokens) - tactics = runner.get_valid_configs(2, 512, 256, 8, num_tokens) - assert tactics, f"no valid NVFP4 SiTU tactic for num_tokens={num_tokens}" - - -@situ_supported -def test_nvfp4_situ_fc31_scale_c_drops_dequant_scale(): - """SiTuGlu is nonlinear in x0, so scaleC must be quantScaleC alone. - - Pins the trtllm-gen rule (``!isLinearInX0(mActType) && mFusedAct`` in - ``BatchedGemm/BatchedGemmTestUtils.h``) at the level where TRT-LLM - computes it, without needing to run a kernel. SwiGLU is the control: it - *is* linear in x0 and keeps the combined ``quantScaleC * dequantScaleAb``. - """ - shape = dict(num_experts=8, top_k=2, hidden_size=512, intermediate_size=256, seq_len=8) - situ_backend, *_ = _make_case(**shape, situ=True) - swiglu_backend, *_ = _make_case(**shape, situ=False) - - # fc2_input_scale is quantScaleC; fc31_alpha is dequantScaleAb (the - # scaleGate the kernel already folds into the tanh/sigmoid arguments). - torch.testing.assert_close( - situ_backend.fc31_scale_c.data.float(), - situ_backend.fc2_input_scale.data.float().expand_as(situ_backend.fc31_scale_c.data), - msg="SiTU fc31_scale_c must not carry the dequantScaleAb factor", - ) - torch.testing.assert_close( - swiglu_backend.fc31_scale_c.data.float(), - (swiglu_backend.fc2_input_scale.data * swiglu_backend.fc31_alpha.data).float(), - msg="SwiGLU fc31_scale_c must keep the combined scale", - ) - # The two conventions must actually differ, otherwise this test would pass - # vacuously on a build where dequantScaleAb happens to be 1. - assert not torch.allclose( - situ_backend.fc31_scale_c.data.float(), swiglu_backend.fc31_scale_c.data.float() - ) - - -@situ_supported -@pytest.mark.parametrize( - "num_experts,top_k,hidden_size,intermediate_size", - [ - pytest.param(8, 2, 512, 256, id="e8_k2_h512_i256"), - pytest.param(8, 2, 2048, 1024, id="e8_k2_h2048_i1024"), - pytest.param(256, 8, 2048, 1024, id="e256_k8_h2048_i1024"), - ], -) -@pytest.mark.parametrize("seq_len", [1, 16, 256], ids=lambda n: f"tokens{n}") -def test_nvfp4_situ_matches_reference(num_experts, top_k, hidden_size, intermediate_size, seq_len): - backend, quantize_util, weights, x, router_logits = _make_case( - num_experts, top_k, hidden_size, intermediate_size, seq_len, situ=True - ) - output = _run_backend(backend, x, router_logits, top_k) - reference = _reference(quantize_util, weights, x, router_logits, top_k, situ=True) - - # Regression guard for the all-zero FC1 output: check_accuracy's tolerance - # band alone passes on a zero tensor here (ref std ~0.04 << atol 0.15). - nonzero = int((output != 0).sum()) - assert nonzero > 0.99 * output.numel(), ( - f"NVFP4 SiTU produced a near-empty output ({nonzero}/{output.numel()} " - "nonzero) -- the FC1 intermediate most likely underflowed to zero" - ) - - rel_l2 = _rel_l2(output, reference) - assert rel_l2 < 0.1, ( - f"NVFP4 SiTU relative L2 error {rel_l2:.4f} is too large " - "(an output uncorrelated with the reference gives ~1.0)" - ) - - quantize_util.create_ref_module(RenormalizeMoeRoutingMethod(top_k=top_k)).check_accuracy( - output, reference - ) - - -_LAUNCH_EVIDENCE_SCRIPT = r""" -import torch -from test_nvfp4_situ_moe import _make_case, _run_backend, _rel_l2, _reference - -backend, quantize_util, weights, x, router_logits = _make_case( - 8, 2, 512, 256, 16, situ=True) -out = _run_backend(backend, x, router_logits, 2) -ref = _reference(quantize_util, weights, x, router_logits, 2, situ=True) -torch.cuda.synchronize() -assert int((out != 0).sum()) > 0, "all-zero output" -assert _rel_l2(out, ref) < 0.1, "reference mismatch" -""" - - -@situ_supported -def test_nvfp4_situ_launches_situ_cubin(): - """The FC1 kernel actually selected must be an NVFP4 siTuGlu cubin. - - Runs in a subprocess because the C++ logger level is fixed at process - start (TLLM_LOG_LEVEL) and TLLM_BATCHED_GEMM_PRINT_NAME logs at INFO. - """ - import subprocess - import sys - - env = dict(os.environ) - env["TLLM_BATCHED_GEMM_PRINT_NAME"] = "1" - env["TLLM_LOG_LEVEL"] = "INFO" - this_dir = os.path.dirname(os.path.abspath(__file__)) - unittest_root = os.path.abspath(os.path.join(this_dir, "..", "..", "..")) - env["PYTHONPATH"] = os.pathsep.join([this_dir, unittest_root, env.get("PYTHONPATH", "")]) - result = subprocess.run( - [sys.executable, "-c", _LAUNCH_EVIDENCE_SCRIPT], - capture_output=True, - text=True, - env=env, - cwd=this_dir, - timeout=600, - ) - log = result.stdout + result.stderr - assert result.returncode == 0, f"fused forward failed:\n{log[-4000:]}" - assert "siTuGlu" in log, ( - "expected the FC1 launch log to name a siTuGlu kernel; got:\n" + log[-4000:] - ) - # E2m1 in/out is what distinguishes the NVFP4 family from the MXFP4 one. - assert "E2m1_E2m1E2m1" in log or "e2m1_e2m1e2m1" in log.lower(), ( - "expected the NVFP4 (E2m1_E2m1E2m1) SiTU cubin family; got:\n" + log[-4000:] - ) - - -@situ_supported -def test_nvfp4_situ_padded_quant_method_is_selected(): - """SiTU must pick the padded NVFP4 method regardless of call ordering. - - ``_get_quant_method`` keys off ``is_situ_activation`` rather than - ``swiglu_alpha is not None`` precisely because create_weights fills the - alpha/beta slots afterwards. - """ - backend, *_ = _make_case(8, 2, 512, 256, 8, situ=True) - assert isinstance(backend.quant_method, NVFP4TRTLLMGenFusedMoEMethod) - assert backend.is_situ_activation - assert backend.scaling_vector_size == 16 diff --git a/tests/unittest/_torch/thop/serial/test_moe.py b/tests/unittest/_torch/thop/serial/test_moe.py index 6d6369740e7c..e2bffdbe9a2b 100644 --- a/tests/unittest/_torch/thop/serial/test_moe.py +++ b/tests/unittest/_torch/thop/serial/test_moe.py @@ -53,17 +53,18 @@ def test_act_type_enum_values_stable(): assert ActType.SiTu.value == 3 -@pytest.mark.skipif( - getSMVersion() not in (100, 103), - reason="The SiTu kernel only supports SM100/SM103. Current SM is %d." % - getSMVersion(), -) -@pytest.mark.parametrize("num_tokens", [1, 8, 512]) -def test_mxe4m3_mxe2m1_situ_runner_has_valid_configs(num_tokens): +def _mxfp4_situ_tactics(num_tokens): + """MXFP8 activations x MXFP4 weights: group-32 scales, padded shapes.""" runner = torch.classes.trtllm.MxE4m3MxE2m1BlockScaleMoERunner( ActType.SiTu.value, True) - tactics = runner.get_valid_configs(2, 512, 256, 8, num_tokens, 512, 256) - assert tactics, f"No valid SiTu tactic for num_tokens={num_tokens}" + return runner.get_valid_configs(2, 512, 256, 8, num_tokens, 512, 256) + + +def _nvfp4_situ_tactics(num_tokens): + """NVFP4 in/out: group-16 scales, and no valid hidden/intermediate sizes + (padding is absorbed by the padded weight buffers instead).""" + runner = torch.classes.trtllm.FP4BlockScaleMoERunner(ActType.SiTu.value) + return runner.get_valid_configs(2, 512, 256, 8, num_tokens) @pytest.mark.skipif( @@ -71,16 +72,20 @@ def test_mxe4m3_mxe2m1_situ_runner_has_valid_configs(num_tokens): reason="The SiTu kernel only supports SM100/SM103. Current SM is %d." % getSMVersion(), ) +@pytest.mark.parametrize("get_tactics", + [_mxfp4_situ_tactics, _nvfp4_situ_tactics], + ids=["mxfp4", "nvfp4"]) @pytest.mark.parametrize("num_tokens", [1, 8, 512]) -def test_e2m1_e2m1_situ_runner_has_valid_configs(num_tokens): - """NVFP4 twin of the MXFP4 check above. +def test_situ_runner_has_valid_configs(num_tokens, get_tactics): + """Both fused SiTu FC1 cubin families must be reachable through tactic + selection. There is no standalone SiTu activation kernel, so an empty + tactic list here means the whole path is unreachable. - Numerics for this path live in - tests/unittest/_torch/modules/moe/test_nvfp4_situ_moe.py. + Numerics and launch evidence for both families live in + tests/unittest/_torch/modules/moe/test_kimi_k3_situ_moe.py. """ - runner = torch.classes.trtllm.FP4BlockScaleMoERunner(ActType.SiTu.value) - tactics = runner.get_valid_configs(2, 512, 256, 8, num_tokens) - assert tactics, f"No valid NVFP4 SiTu tactic for num_tokens={num_tokens}" + assert get_tactics( + num_tokens), f"No valid SiTu tactic for num_tokens={num_tokens}" class moe_args: From 9118e0cf4269f8086a1e5ab5214ffbe9996e7751 Mon Sep 17 00:00:00 2001 From: rosong11 Date: Tue, 25 Aug 2026 07:31:05 +0000 Subject: [PATCH 3/4] [None][feat] Enable the TRTLLM-Gen MoE backend on the whole SM100 family The trtllm-gen batchedGemm drop ships 601 sm_100f cubins alongside 40 sm_100a and 40 sm_103a ones. sm_100f is family-compatible, so it loads on any SM in [100, 110) -- but the stack rejected everything except SM100/SM103, which left SM107 unserved even though the kernels run there. Three gates had to move, and only the C++ one was load-bearing: * KernelRunner.cpp isSMCompatible() accepted gpuSM in {90, 100, 103} and TLLM_THROW'd otherwise. It is called unconditionally from the TrtllmGenBatchedGemmRunner constructor, which blockScaleMoe/runner.cu builds for every FP4/MXFP4 MoE, so widening only the Python side would have turned a clean backend reject into a throw at runner construction. It now accepts sm_100f across the family while keeping sm_100a/sm_103a locked to their own arch. * TRTLLMGenFusedMoE.can_implement() gated the whole backend on {100, 103}; this is what actually kept SM107 out, so the SiTu-local check below it was never reachable on SM107. * _validate_backend_local_activation()'s SiTu check, for consistency. Both Python gates now use the existing tensorrt_llm._utils.is_sm_100f helper (100 <= sm < 110), which also replaces the open-coded range in the fused route-quant fast path so the file no longer carries two spellings of the same predicate. thop/fp4BlockScaleMoe.cpp and mxFp4BlockScaleMoe.cpp already used isSM100Family() and needed no change. The SiTu test skip guards move with them, otherwise the tests would silently skip on the hardware this enables. Behavior is unchanged for every SM the stack served before: is_sm_100f(90), (120) and (121) are all False exactly as {100, 103} membership was, and the C++ SM100/SM103 branches are byte-for-byte equivalent. The only difference is SM in {101, 102, 104..109}, which previously rejected and now runs sm_100f. SM107 was validated separately on Rubin silicon. Verified here on B200/SM100 against a local sm100-real build that nothing regressed: thop/serial/test_moe.py -k situ 9 passed test_moe_backend.py -k trtllm_gen*situ 5 passed test_kimi_k3_situ_moe.py (full module) 69 passed, 1 xfailed test_moe_backend.py minus the waived matrix 55 passed, 4 skipped NVFP4 SiTu tactic count on SM100 is unchanged at 72. Signed-off-by: rosong11 --- .../batchedGemm/KernelRunner.cpp | 26 ++++++++++++++----- .../modules/fused_moe/fused_moe_trtllm_gen.py | 20 +++++++------- .../moe/kimi_k3_ref_moe/_moe_kernels.py | 10 ++++--- .../_torch/modules/moe/test_moe_backend.py | 6 ++--- tests/unittest/_torch/thop/serial/test_moe.py | 8 +++--- 5 files changed, 44 insertions(+), 26 deletions(-) diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/KernelRunner.cpp b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/KernelRunner.cpp index 782772fa28c1..19e9fc9c7a21 100644 --- a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/KernelRunner.cpp +++ b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/KernelRunner.cpp @@ -45,13 +45,27 @@ using tensorrt_llm::common::fmtstr; constexpr bool isSMCompatible(int gpuSM, SmVersion kernelSM) { - if (gpuSM == 103) + // Blackwell family. Family-compatible (sm_100f) cubins load on any SM in + // [100, 110); arch-specific ones are locked to the SM they were built for. + // The batchedGemm drop ships Sm100f, Sm100a and Sm103a only, so family + // members other than 100/103 are served by Sm100f alone. + // Range kept in sync with tensorrt_llm::common::isSM100Family (cudaUtils.h), + // which is not constexpr and so cannot be reused here. + if (gpuSM >= 100 && gpuSM < 110) { - return kernelSM == SmVersion::Sm100f || kernelSM == SmVersion::Sm103a; - } - else if (gpuSM == 100) - { - return kernelSM == SmVersion::Sm100f || kernelSM == SmVersion::Sm100a; + if (kernelSM == SmVersion::Sm100f) + { + return true; + } + if (gpuSM == 100) + { + return kernelSM == SmVersion::Sm100a; + } + if (gpuSM == 103) + { + return kernelSM == SmVersion::Sm103a; + } + return false; } else if (gpuSM == 90) { diff --git a/tensorrt_llm/_torch/modules/fused_moe/fused_moe_trtllm_gen.py b/tensorrt_llm/_torch/modules/fused_moe/fused_moe_trtllm_gen.py index 49bb019074df..baabf629f4d6 100644 --- a/tensorrt_llm/_torch/modules/fused_moe/fused_moe_trtllm_gen.py +++ b/tensorrt_llm/_torch/modules/fused_moe/fused_moe_trtllm_gen.py @@ -21,7 +21,7 @@ import torch from torch import nn -from tensorrt_llm._utils import get_sm_version +from tensorrt_llm._utils import get_sm_version, is_sm_100f from tensorrt_llm.logger import logger from tensorrt_llm.models.modeling_utils import QuantAlgo @@ -141,7 +141,7 @@ class TRTLLMGenFusedMoE(MoE): @classmethod def can_implement(cls, p: MoEProblem, d: MoEDeployment) -> MoEEligibility: - """TRTLLM-Gen kernels: SM100/SM103, bfloat16 activations. + """TRTLLM-Gen kernels: the SM100 (Blackwell) family, bfloat16 activations. Quantized coverage is ``_SUPPORTED_QUANT_ALGOS``. The unquantized BF16 path is served by a FlashInfer kernel, so it is available only where @@ -150,11 +150,14 @@ def can_implement(cls, p: MoEProblem, d: MoEDeployment) -> MoEEligibility: sm_version = d.env.sm quant_algo = p.quant_algo - # TRTLLMGenFusedMoE requires SM in {100, 103} - if sm_version not in {100, 103}: + # The cubin drop is sm_100f (family-compatible) plus arch-specific + # sm_100a/sm_103a, so the whole SM100 family is servable; the C++ + # selector (KernelRunner.cpp isSMCompatible) picks sm_100f on family + # members without their own arch build. + if not is_sm_100f(sm_version): return _reject( MoERejectReason.SM_UNSUPPORTED, - f"TRTLLMGenFusedMoE requires SM100 or SM103, got SM{sm_version}" + f"TRTLLMGenFusedMoE requires the SM100 family, got SM{sm_version}" ) # forward_impl asserts x.dtype == torch.bfloat16 @@ -380,8 +383,8 @@ def _validate_backend_local_activation(self) -> None: raise ValueError( "TRTLLM-Gen SiTu requires bfloat16 activations, got " f"{self.dtype}.") - if get_sm_version() not in {100, 103}: - raise ValueError("TRTLLM-Gen SiTu requires SM100 or SM103, got " + if not is_sm_100f(): + raise ValueError("TRTLLM-Gen SiTu requires the SM100 family, got " f"SM{get_sm_version()}.") quant_algo = (None if self.quant_config is None else self.quant_config.quant_algo) @@ -714,8 +717,7 @@ def try_fused_kimi_route_quant( or isinstance(x, MxFp8QuantizedTensor)): return None - sm_version = get_sm_version() - if (not 100 <= sm_version < 110 or not self.has_w4a8_mxfp4_mxfp8 + if (not is_sm_100f() or not self.has_w4a8_mxfp4_mxfp8 or not isinstance(self.op_backend, TRTLLMOpBackend) or not isinstance(self.routing_method, DeepSeekV3MoeRoutingMethod)): diff --git a/tests/unittest/_torch/modules/moe/kimi_k3_ref_moe/_moe_kernels.py b/tests/unittest/_torch/modules/moe/kimi_k3_ref_moe/_moe_kernels.py index f21b81328722..9f95757f955e 100644 --- a/tests/unittest/_torch/modules/moe/kimi_k3_ref_moe/_moe_kernels.py +++ b/tests/unittest/_torch/modules/moe/kimi_k3_ref_moe/_moe_kernels.py @@ -74,8 +74,8 @@ def get_moe_sm_version() -> int: def is_native_situ_supported() -> bool: - """Native SiTU cubins are Blackwell sm_100f (SM100/SM103) only.""" - return get_moe_sm_version() in (100, 103) + """Native SiTU cubins are Blackwell sm_100f: the whole SM100 family.""" + return 100 <= get_moe_sm_version() < 110 def assert_native_situ_supported( @@ -88,8 +88,10 @@ def assert_native_situ_supported( if not torch.cuda.is_available(): raise RuntimeError("native SiTU MoE requires CUDA; no CUDA device is available") sm = get_moe_sm_version() - if sm not in (100, 103): - raise RuntimeError(f"native SiTU MoE requires SM100/SM103 (Blackwell); running on SM{sm}") + if not 100 <= sm < 110: + raise RuntimeError( + f"native SiTU MoE requires the SM100 (Blackwell) family; running on SM{sm}" + ) if group_size != SCALING_VECTOR_SIZE: raise RuntimeError( f"native SiTU MoE requires MXFP4 group_size {SCALING_VECTOR_SIZE}, got {group_size}" diff --git a/tests/unittest/_torch/modules/moe/test_moe_backend.py b/tests/unittest/_torch/modules/moe/test_moe_backend.py index 2e3eda62d25b..987ef15cc46f 100644 --- a/tests/unittest/_torch/modules/moe/test_moe_backend.py +++ b/tests/unittest/_torch/modules/moe/test_moe_backend.py @@ -85,7 +85,7 @@ MxFp8QuantizedTensor, is_gated_activation, ) -from tensorrt_llm._utils import get_sm_version, mpi_rank +from tensorrt_llm._utils import get_sm_version, is_sm_100f, mpi_rank from tensorrt_llm.mapping import Mapping from tensorrt_llm.models.modeling_utils import QuantAlgo, QuantConfig @@ -498,8 +498,8 @@ def test_marlin_override_quant_config_degrades_per_layer(): # are the contracts that hold without running a cubin. _situ_supported = pytest.mark.skipif( - get_sm_version() not in (100, 103), - reason="TRTLLM-Gen SiTu cubins only exist for SM100/SM103", + not is_sm_100f(), + reason="TRTLLM-Gen SiTu cubins are sm_100f: the SM100 family only", ) # Kimi K3 defaults. The gate-side SiTu beta is the cubin's ``alpha``, the diff --git a/tests/unittest/_torch/thop/serial/test_moe.py b/tests/unittest/_torch/thop/serial/test_moe.py index e2bffdbe9a2b..e50c3557432e 100644 --- a/tests/unittest/_torch/thop/serial/test_moe.py +++ b/tests/unittest/_torch/thop/serial/test_moe.py @@ -68,8 +68,8 @@ def _nvfp4_situ_tactics(num_tokens): @pytest.mark.skipif( - getSMVersion() not in (100, 103), - reason="The SiTu kernel only supports SM100/SM103. Current SM is %d." % + not (100 <= getSMVersion() < 110), + reason="The SiTu kernels are sm_100f (SM100 family). Current SM is %d." % getSMVersion(), ) @pytest.mark.parametrize("get_tactics", @@ -2699,8 +2699,8 @@ def test_moe_mxe2m1_weights(num_tokens, hidden_size, intermediate_size, @pytest.mark.skipif( - getSMVersion() not in (100, 103), - reason="The SiTu kernel only supports SM100/SM103. Current SM is %d." % + not (100 <= getSMVersion() < 110), + reason="The SiTu kernels are sm_100f (SM100 family). Current SM is %d." % getSMVersion(), ) @pytest.mark.parametrize("num_tokens", [1, 8, 256]) From 41359647ada0e0ad26a53ce01073e7d6b17c08a5 Mon Sep 17 00:00:00 2001 From: rosong11 Date: Tue, 25 Aug 2026 08:57:43 +0000 Subject: [PATCH 4/4] [None][fix] Validate the SiTu MoE-TP shard against the resolved NVFP4 alignment The MoE-TP shard check in _validate_backend_local_activation read NVFP4TRTLLMGenFusedMoEMethod.weight_alignment, i.e. the class attribute, which is only a starting point: create_weights() replaces it with 128 or 256 depending on hidden_size and intermediate_size_per_partition. A per-rank shard divisible by the unresolved default (32) but not by the resolved alignment therefore passed validation and reached a loader that cannot lay it out -- hidden=1536 / intermediate=192 is the smallest example, since 192 % 32 == 0 but the resolved alignment is 128. The check cannot simply read the instance attribute instead: it runs from __init__, deliberately before create_weights(), so that an illegal shard is rejected before any weight buffer is allocated. So the resolution moves into a resolve_alignments() classmethod that both call sites share; it depends only on hidden_size and intermediate_size_per_partition, both known at __init__. This is NVFP4-only. MXFP4WeightTRTLLMGenFusedMoEMethod.weight_alignment (128) is never reassigned, so the MXFP4 branch of the same check was already correct and is left alone. Sharing the resolution also makes create_weights() idempotent. It previously rounded up from self.weight_alignment, so a second call started from the value the first one had selected; it now always starts from the class default. Two tests: one pins that resolve_alignments() predicts what create_weights() actually selects, across shapes that resolve to 32, 128 and 256; the other pins that 192 is divisible by the class default but not by its resolved alignment, so restoring the class-attribute read fails here rather than in a multi-GPU accuracy run. An end-to-end MoE-TP rejection test is not possible in a single process -- Mapping(world_size>1) fails in AllReduce setup inside super().__init__(), before this validation runs. Verified on B200/SM100: test_moe_backend.py minus the waived matrix 59 passed, 4 skipped test_moe_module.py -k "TRTLLM and NVFP4" 61 passed, 80 skipped test_kimi_k3_situ_moe.py (full module, x5) 69 passed, 1 xfailed resolve_alignments() matches create_weights() on all of (512,256), (3584,3072), (2880,2880), (2048,1024), (1536,192), (4096,192) and (7168,2048). Signed-off-by: rosong11 --- .../modules/fused_moe/fused_moe_trtllm_gen.py | 15 ++++-- .../_torch/modules/fused_moe/quantization.py | 53 +++++++++++++------ .../_torch/modules/moe/test_moe_backend.py | 49 +++++++++++++++++ 3 files changed, 98 insertions(+), 19 deletions(-) diff --git a/tensorrt_llm/_torch/modules/fused_moe/fused_moe_trtllm_gen.py b/tensorrt_llm/_torch/modules/fused_moe/fused_moe_trtllm_gen.py index baabf629f4d6..30b1b5931b8f 100644 --- a/tensorrt_llm/_torch/modules/fused_moe/fused_moe_trtllm_gen.py +++ b/tensorrt_llm/_torch/modules/fused_moe/fused_moe_trtllm_gen.py @@ -401,9 +401,18 @@ def _validate_backend_local_activation(self) -> None: # per-rank shard to stay a whole multiple of the quant method's # weight alignment so per-shard scale groups and the padded # weight buffers line up without fractional groups. - alignment = (NVFP4TRTLLMGenFusedMoEMethod.weight_alignment - if quant_algo == QuantAlgo.NVFP4 else - W4A8MXFP4MXFP8TRTLLMGenFusedMoEMethod.weight_alignment) + if quant_algo == QuantAlgo.NVFP4: + # NVFP4 picks its alignment from the layer shape in + # create_weights (32 -> 128 or 256), which runs after this + # check. Ask for the resolved value: the class attribute is + # only the starting point, and validating against it would + # admit shards the loader cannot lay out. + alignment, _ = NVFP4TRTLLMGenFusedMoEMethod.resolve_alignments( + self.hidden_size, self.intermediate_size_per_partition) + else: + # MXFP4's alignment is a fixed class attribute. + alignment = ( + W4A8MXFP4MXFP8TRTLLMGenFusedMoEMethod.weight_alignment) if (self.intermediate_size % self.tp_size != 0 or self.intermediate_size_per_partition % alignment != 0): raise ValueError( diff --git a/tensorrt_llm/_torch/modules/fused_moe/quantization.py b/tensorrt_llm/_torch/modules/fused_moe/quantization.py index b661790ead75..cdea7e6013e6 100644 --- a/tensorrt_llm/_torch/modules/fused_moe/quantization.py +++ b/tensorrt_llm/_torch/modules/fused_moe/quantization.py @@ -5347,9 +5347,43 @@ def _finalize_shared_expert_alphas(self, module: torch.nn.Module): class NVFP4TRTLLMGenFusedMoEMethod(NVFP4TRTLLMGenFusedMoEBaseMethod): + # Starting point only: create_weights resolves both from the layer shape + # via resolve_alignments() below. Read the class attributes directly only + # when you mean "the unresolved default". weight_alignment = 32 input_hidden_alignment = 32 + @classmethod + def resolve_alignments(cls, hidden_size: int, + intermediate_size_per_partition: int): + """The alignments create_weights will select for this shape. + + Exposed as a classmethod because TRTLLMGenFusedMoE validates its + MoE-TP shard from __init__, i.e. before create_weights has run and + shadowed the class attributes with instance ones. Reading the class + attributes from there would validate against the unresolved default + (32) and admit shards this method cannot lay out. + + Returns ``(weight_alignment, input_hidden_alignment)``. + """ + # Padding is only enabled for hidden_size > 1024, since there are + # small unit tests that expect no padding. + if hidden_size > 1024 and hidden_size % 256 != 0: + # For now let's keep input alignment same as weight alignment. + # There are practical reasons that this might be a different + # value. See the comment in MXFP4WeightTRTLLMGenFusedMoEMethod + # for more details. + return 256, 256 + # Weight scales require M % 128 in + # get_shuffle_matrix_sf_a_row_indices. If the intermediate size after + # padding does not satisfy that, drop to a 128 weight alignment. + padded = ( + (intermediate_size_per_partition + cls.weight_alignment - 1) // + cls.weight_alignment * cls.weight_alignment) + if padded % 128 != 0: + return 128, cls.input_hidden_alignment + return cls.weight_alignment, cls.input_hidden_alignment + def get_weights_shapes(self, module: torch.nn.Module, weight_vec_size: int, block_scales_vec_size: int): @@ -5403,22 +5437,9 @@ def _round_up(self, x, alignment): return (x + alignment - 1) // alignment * alignment def create_weights(self, module: torch.nn.Module): - # Here we only enable padding for hidden_size > 1024 since there are small unit tests that expect no padding. - if module.hidden_size > 1024 and module.hidden_size % 256 != 0: - self.weight_alignment = 256 - # For now let's keep input alignment same as weight alignment. There are practical reasons that this might be a different value. - # See the comment in MXFP4WeightTRTLLMGenFusedMoEMethod for more details. - self.input_hidden_alignment = 256 - - else: - # Weight scales require M % 128 in get_shuffle_matrix_sf_a_row_indices. - # Check if intermediate_size after padding satisfies this requirement. - # If not, set weight_alignment to 128. - intermediate_size_padded = self._round_up( - module.intermediate_size_per_partition, self.weight_alignment) - if intermediate_size_padded % 128 != 0: - self.weight_alignment = 128 - + self.weight_alignment, self.input_hidden_alignment = ( + self.resolve_alignments(module.hidden_size, + module.intermediate_size_per_partition)) super().create_weights(module, bias_dtype=torch.float32) def setup_quant_scales(self, module: torch.nn.Module): diff --git a/tests/unittest/_torch/modules/moe/test_moe_backend.py b/tests/unittest/_torch/modules/moe/test_moe_backend.py index 987ef15cc46f..6730d3448089 100644 --- a/tests/unittest/_torch/modules/moe/test_moe_backend.py +++ b/tests/unittest/_torch/modules/moe/test_moe_backend.py @@ -664,6 +664,55 @@ def test_trtllm_gen_situ_rejects_quant_algos_without_fused_cubins(quant_algo) -> _make_trtllm_gen_moe(quant_config=QuantConfig(quant_algo=quant_algo), situ=True) +@_situ_supported +@pytest.mark.parametrize( + "hidden_size,intermediate_size,expected", + [ + pytest.param(512, 256, (32, 32), id="no_padding"), + pytest.param(2880, 2880, (256, 256), id="hidden_not_256_aligned"), + pytest.param(1536, 192, (128, 32), id="intermediate_needs_128"), + ], +) +def test_nvfp4_trtllm_gen_resolve_alignments_matches_create_weights( + hidden_size, intermediate_size, expected +) -> None: + """``resolve_alignments`` must predict what ``create_weights`` selects. + + ``TRTLLMGenFusedMoE`` validates its MoE-TP shard from ``__init__``, i.e. + before ``create_weights`` has shadowed the class attributes with + shape-resolved instance ones, so the check has to predict them. If the two + ever drift apart the shard contract is validated against the wrong number. + """ + predicted = NVFP4TRTLLMGenFusedMoEMethod.resolve_alignments(hidden_size, intermediate_size) + assert predicted == expected + + backend = _make_trtllm_gen_moe( + quant_config=QuantConfig(quant_algo=QuantAlgo.NVFP4), + situ=True, + hidden_size=hidden_size, + intermediate_size=intermediate_size, + ) + assert ( + backend.quant_method.weight_alignment, + backend.quant_method.input_hidden_alignment, + ) == predicted + + +@_situ_supported +def test_nvfp4_trtllm_gen_class_alignment_would_admit_bad_shards() -> None: + """Why the MoE-TP check cannot read the class attribute. + + ``hidden=1536``/``intermediate=192`` resolves to a 128 weight alignment. + 192 is divisible by the class default (32) but not by 128, so validating + against ``NVFP4TRTLLMGenFusedMoEMethod.weight_alignment`` admits a shard + the loader cannot lay out. Pins the gap, so reverting the check to the + class attribute fails here rather than in a multi-GPU accuracy run. + """ + resolved, _ = NVFP4TRTLLMGenFusedMoEMethod.resolve_alignments(1536, 192) + assert 192 % NVFP4TRTLLMGenFusedMoEMethod.weight_alignment == 0 + assert 192 % resolved != 0 + + def test_megamoe_cutedsl_post_load_weights_uses_staged_hooks(): moe = MegaMoECuteDsl.__new__(MegaMoECuteDsl) torch.nn.Module.__init__(moe)