Repository navigation

[Metal] Add affine qqmm with int8 NAX MMAs - #4802

Open

XXXXRT666 wants to merge 4 commits into

ml-explore:mainfrom

XXXXRT666:metal-nax-affine-int-mma

Open

[Metal] Add affine qqmm with int8 NAX MMAs#4802
XXXXRT666 wants to merge 4 commits into
ml-explore:mainfrom
XXXXRT666:metal-nax-affine-int-mma

Commits on Nov 3, 2026
- [Metal] Route qqmm without global scales like quantized_matmul
  
  XXXXRT666
  authored and
  zcbenz
  committed
  cab962d
- [Metal] Add affine qqmm with int8 NAX MMAs
  XXXXRT666
  committed
  6c3997b
- [Metal] Route affine qqmm to int8 MMAs from 1.5 tiles per GPU core
  XXXXRT666
  committed
  63217fa
- Support affine in QQLinear and qqmm vjp with quantized weights
  XXXXRT666
  committed
  21f5e03