[None][feat] Rubin kernels & attention: DSV4/DSA, CuteDSL GEMM - #19184

Merged

reasonsolo merged 21 commits into

NVIDIA:mainfrom

reasonsolo:user/lizhiz/mb-kernels

Sep 19, 2026

Commits on Sep 17, 2026

[None][feat] Rubin kernels & attention: DSV4/DSA, Helix, CuteDSL GEMM
reasonsolo
committed
ba1f808
[None][feat] restore the Rubin fused-FC12 NVFP4 MoE op registration
reasonsolo
committed
860350e
[None][fix] drop two stray args from the mixed-cluster grid call
reasonsolo
committed
6f843fa
[None][fix] repair two defects in the GatedMLP SwiGLU fallback test
reasonsolo
committed
f79af97
[None][fix] gate the fused NVFP4 SwiGLU epilogue on swiglu_limit
reasonsolo
committed
ce8e2b7
[None][fix] accept q_rope_applied in the mla_rope_generation fake impl
reasonsolo
committed
f847447
[None][fix] restore SM107 in the CuteDSL MLA decode reduction path
reasonsolo
committed
bf6a9fe
[None][chore] move the Helix verify-group work out of this PR
reasonsolo
committed
29b1b2f
[None][fix] rebase the new DSA TopK tests onto main's production code
reasonsolo
committed
cc0120b
[None][feat] complete the SiTU plumbing for the gather grouped GEMM ops
reasonsolo
committed
eb7f559
[None][fix] restore the mHC Phase-4 store-to-load coherence fix
reasonsolo
committed
67f81c8
[None][fix] keep Flux on the shared fused NVFP4 SwiGLU predicate
reasonsolo
committed
3fed8f9
[None][fix] stop the gather act-fusion test staging weights through int32
reasonsolo
committed
f01972b
[None][fix] address review: f32x2 arch bound and FC12 activation guard
reasonsolo
committed
2d75179
[None][fix] register the missing fp8_prequantized_swap_ab_gemm op
reasonsolo
committed
57b6791
[None][fix] repair the C++ unit-test build and the fused FP8 SwiGLU gate
reasonsolo
committed
639140f
[None][fix] align the mla_rope_generation fake with the native schema
reasonsolo
committed
1111762

Navigation Menu

[None][feat] Rubin kernels & attention: DSV4/DSA, CuteDSL GEMM - #19184

[None][feat] Rubin kernels & attention: DSV4/DSA, CuteDSL GEMM#19184
reasonsolo merged 21 commits into
NVIDIA:mainfrom
reasonsolo:user/lizhiz/mb-kernels

Commits on Sep 17, 2026

[None][feat] Rubin kernels & attention: DSV4/DSA, Helix, CuteDSL GEMM

[None][feat] restore the Rubin fused-FC12 NVFP4 MoE op registration

[None][fix] drop two stray args from the mixed-cluster grid call

[None][fix] repair two defects in the GatedMLP SwiGLU fallback test

[None][fix] gate the fused NVFP4 SwiGLU epilogue on swiglu_limit

[None][fix] accept q_rope_applied in the mla_rope_generation fake impl

[None][fix] restore SM107 in the CuteDSL MLA decode reduction path

[None][chore] move the Helix verify-group work out of this PR

[None][fix] rebase the new DSA TopK tests onto main's production code

[None][feat] complete the SiTU plumbing for the gather grouped GEMM ops

[None][fix] restore the mHC Phase-4 store-to-load coherence fix

[None][fix] keep Flux on the shared fused NVFP4 SwiGLU predicate

[None][fix] stop the gather act-fusion test staging weights through int32

[None][fix] address review: f32x2 arch bound and FC12 activation guard

[None][fix] register the missing fp8_prequantized_swap_ab_gemm op

[None][fix] repair the C++ unit-test build and the fused FP8 SwiGLU gate

[None][fix] align the mla_rope_generation fake with the native schema

Commits on Sep 18, 2026

[None][fix] drop the re-introduced trtllm-gen FMHA CGA smem-reduction clamp

[None][fix] rebase the GVR prior indexer test onto main's TopK contract

[None][fix] teach the Flux guard fixture the shared fused SwiGLU predicate

[None][fix] refresh the Rubin MoE op-schema expectations for SiTU