Repository navigation

[Quant] Serve 32-wide-K ue8m0 block-FP8 linears through the FlashInfer MXFP8 GEMMs - #40039

Merged

hnyls2002 merged 8 commits into

mainfrom

lsyin/dsv4-quant-block-fp8-mxfp8

Sep 18, 2026

Commits on Sep 17, 2026
- block-fp8 dispatch by weight block size; triton ue8m0 activation scales
  hnyls2002
  committed
  dbc2c0b
- block-fp8-as-mxfp8 helpers; Mxfp8SwizzledInput
  hnyls2002
  committed
  b14f037
- Fp8LinearMethod: mxfp8 view of 32-wide-K ue8m0 block weights; keep_plain_weight_layout
  hnyls2002
  committed
  931f95f
- Fp8LinearMethod.apply: route mxfp8 view; accept Mxfp8SwizzledInput
  hnyls2002
  committed
  49e50bd
- fold quant tests into existing suites
  hnyls2002
  committed
  b78d34d
- trim quant comments
  hnyls2002
  committed
  fe76c57
Commits on Sep 18, 2026
- pin cpu device in e8m0 test
  hnyls2002
  committed
  6b71471
- Merge remote-tracking branch 'origin/main' into lsyin/dsv4-quant-block-fp8-mxfp8
  hnyls2002
  committed
  862a2d5