fix(GFX1100-TG200): skinny_gemm OOB write + MoE silu gate rounding - #2894

Closed

ghazni101 wants to merge 17 commits into

mudler:mainfrom

ghazni101:row/GFX1100-TG200-BUGFIX

Commits on Sep 8, 2026

feat(KERNEL-QUANT-CIQ-GEMM-ROCM): land the W1 keep-quant providers on kROCM
ghazni101
committed
8ef6ec3
fix(KERNEL-QUANT-CIQ-GEMM-ROCM): keep grouped Q8_0 on the kROCM provider

mudler
authored and
ghazni101
committed
f1819b2
Revert "fix(KERNEL-QUANT-CIQ-GEMM-ROCM): keep grouped Q8_0 on the kROCM provider"

mudler
authored and
ghazni101
committed
a6eb567
fix(rocm): delegate grouped Q8_0 to *Gdn kernel (#2927)
ghazni101
committed
a338124
fix(rocm): export the keep-quant providers and drop the duplicate registrar
ghazni101
committed
35063b2
spec(BACKEND-ROCM): define Qwen3.5 gfx1100 numerical gate

VikashLoomba
authored and
ghazni101
committed
6a3932d
spec(BACKEND-ROCM): repair Qwen3.5 numerical plan

VikashLoomba
authored and
ghazni101
committed
9476aaa
spec(BACKEND-ROCM): correct Qwen3.5 capture contracts

VikashLoomba
authored and
ghazni101
committed
fc7c284
perf(GFX1100-TG200): T4a MMVQ K-quant decode GEMV arm
ghazni101
committed
8116aa7
feat(GFX1100-TG200): lever-C fuses Q8_K activation quant into the RmsNorm epilogue
ghazni101
committed
b8ee53e
T27: warp-cooperative QuantizeQ8KK for decode (+2.06%, byte-identical)
ghazni101
committed
d83e7f0
feat(GFX1100-TG200): T24 LDS-buffered quant epilogue in RmsNormRowCoopKernel
ghazni101
committed
d58c923
feat(GFX1100-TG200): T21 keep-quant for V-head row-permuted GDN projections
ghazni101
committed
aa6f4ae
fix(rocm): gate the coop rmsnorm vector body to bf16 gamma, seed quant amax at 0
ghazni101
committed
a0b7333
T25: keep ssm_out as Q5_K with runtime input permutation
ghazni101
committed
4960fdf
fix(rocm): count the fused MMVQ route once per call
ghazni101
committed
a6be41b
fix(GFX1100-TG200): guard the skinny-GEMM tail store against the partial M tile
ghazni101
committed
c3095f8

Navigation Menu

fix(GFX1100-TG200): skinny_gemm OOB write + MoE silu gate rounding - #2894

fix(GFX1100-TG200): skinny_gemm OOB write + MoE silu gate rounding#2894
ghazni101 wants to merge 17 commits into
mudler:mainfrom
ghazni101:row/GFX1100-TG200-BUGFIX

Commits on Sep 8, 2026

feat(KERNEL-QUANT-CIQ-GEMM-ROCM): land the W1 keep-quant providers on kROCM

fix(KERNEL-QUANT-CIQ-GEMM-ROCM): keep grouped Q8_0 on the kROCM provider

Revert "fix(KERNEL-QUANT-CIQ-GEMM-ROCM): keep grouped Q8_0 on the kROCM provider"

fix(rocm): delegate grouped Q8_0 to *Gdn kernel (#2927)

fix(rocm): export the keep-quant providers and drop the duplicate registrar

spec(BACKEND-ROCM): define Qwen3.5 gfx1100 numerical gate

spec(BACKEND-ROCM): repair Qwen3.5 numerical plan

spec(BACKEND-ROCM): correct Qwen3.5 capture contracts

perf(GFX1100-TG200): T4a MMVQ K-quant decode GEMV arm

feat(GFX1100-TG200): lever-C fuses Q8_K activation quant into the RmsNorm epilogue

T27: warp-cooperative QuantizeQ8KK for decode (+2.06%, byte-identical)

feat(GFX1100-TG200): T24 LDS-buffered quant epilogue in RmsNormRowCoopKernel

feat(GFX1100-TG200): T21 keep-quant for V-head row-permuted GDN projections

fix(rocm): gate the coop rmsnorm vector body to bf16 gamma, seed quant amax at 0

T25: keep ssm_out as Q5_K with runtime input permutation

fix(rocm): count the fused MMVQ route once per call

fix(GFX1100-TG200): guard the skinny-GEMM tail store against the partial M tile