Repository navigation

[ROCm][Perf] MXFP8 GEMM on native 32x32 block scales for gfx950 - #58510

Merged

AndreasKaratzas merged 7 commits into

vllm-project:mainfrom

Fangzhou-Ai:rocm-mxfp8-block32-gemm

Sep 25, 2026

Merged

[ROCm][Perf] MXFP8 GEMM on native 32x32 block scales for gfx950#58510
AndreasKaratzas merged 7 commits into
vllm-project:mainfrom
Fangzhou-Ai:rocm-mxfp8-block32-gemm

Commits on Sep 24, 2026
- [ROCm][Perf] MXFP8 GEMM on native 32x32 block scales for gfx950
  
  Fangzhou-Ai
  and
  claude
  committed
  05701ee
- [ROCm] Read compact 32x32 wo_a scales in the grouped MXFP8 wo_a GEMM
  
  Fangzhou-Ai
  and
  claude
  committed
  3b46e2d
- [ROCm][Perf] Sum MXFP8 block32 GEMM split-K partials within the launch
  
  Fangzhou-Ai
  and
  claude
  committed
  c3c0032
- Merge branch 'main' into rocm-mxfp8-block32-gemm
  Fangzhou-Ai
  authored
  225afd7
- [ROCm] Compact wo_a test scales with the linear kernel's helper
  
  Fangzhou-Ai
  and
  claude
  committed
  17c6a9d
- Merge branch 'main' into rocm-mxfp8-block32-gemm
  Fangzhou-Ai
  authored
  984c634
Commits on Sep 25, 2026
- Merge branch 'main' into rocm-mxfp8-block32-gemm
  Fangzhou-Ai
  authored
  1112cce