Repository navigation

[Perf] [AMD] qwen3.5: Fuse activation quant into small-M FP8 block-scale GEMM on gfx950 - #42229

Open

chuyeh wants to merge 8 commits into

sgl-project:mainfrom

chuyeh:feat/rocm-smallm-fp8-blockscale-gemm

Open

[Perf] [AMD] qwen3.5: Fuse activation quant into small-M FP8 block-scale GEMM on gfx950#42229
chuyeh wants to merge 8 commits into
sgl-project:mainfrom
chuyeh:feat/rocm-smallm-fp8-blockscale-gemm

Commits on Oct 7, 2026
- [AMD] Fuse the 1x128 activation quant into a gfx950 small-M FP8 block-scale GEMM
  
  chuyeh
  and
  cursoragent
  committed
  d9d4318
- Reuse smallm_moe_gfx950's _Kernel for the function lookup
  
  chuyeh
  and
  cursoragent
  committed
  9156162
Commits on Oct 8, 2026
- Merge branch 'main' into feat/rocm-smallm-fp8-blockscale-gemm
  yichiche
  authored
  69206c9
- [AMD] Gate the small-M FP8 test on AITER
  chuyeh
  committed
  4468dbe
- Sync main and simplify small-M FP8 dispatch.
  chuyeh
  committed
  c77976c
Commits on Oct 15, 2026
- Add Qwen3.5 TP8 shapes to the small-M FP8 block-scale GEMM.
  
  chuyeh
  and
  cursoragent
  committed
  7b56ab4
- Merge remote-tracking branch 'upstream/main' into feat/rocm-smallm-fp8-blockscale-gemm
  chuyeh
  committed
  3358a25
Commits on Oct 16, 2026
- Merge remote-tracking branch 'upstream/main' into feat/rocm-smallm-fp8-blockscale-gemm
  chuyeh
  committed
  3b71892