Repository navigation

extend MXFP4 recipe to GroupedLinear and a8w4 (Triton grouped GEMM kernels) - #745

Open

matthiasdiener wants to merge 31 commits into

devfrom

mdiener/mxfp4-a8w4-recipe

Open

extend MXFP4 recipe to GroupedLinear and a8w4 (Triton grouped GEMM kernels)#745
matthiasdiener wants to merge 31 commits into
devfrom
mdiener/mxfp4-a8w4-recipe

Commits on Sep 28, 2026
- add a8w4/mxfp4 reference recipe
  matthiasdiener
  committed
  349e95b
- enable fast path for groupedlinear
  matthiasdiener
  committed
  f4c9b08
- address copilot review
  matthiasdiener
  committed
  a42c6c1
- weight caching
  matthiasdiener
  committed
  7ff27ca
- fused downcast
  matthiasdiener
  committed
  94be09d
Commits on Sep 29, 2026
- adjust comments
  matthiasdiener
  authored
  898ef53
- swizzle a8w4
  matthiasdiener
  committed
  41e94da
- use native quantizer for a4w4
  matthiasdiener
  committed
  92a19b8
- address copilot reviews
  matthiasdiener
  committed
  c893901
- address comments
  matthiasdiener
  committed
  fcd002d
- wire in out
  matthiasdiener
  committed
  9aaa49b
- address review
  matthiasdiener
  committed
  aef244a
- validate buffer
  matthiasdiener
  committed
  4f44d8a
- address comments
  matthiasdiener
  committed
  8da1066
- check quantizer type before accessing MXFP4-only attributes
  
  matthiasdiener
  and
  Copilot
  authored
  69b9ba6
- address review
  matthiasdiener
  committed
  1580d77
- only quantize a4w4 col weight when input grad is needed
  matthiasdiener
  committed
  d026ae2
- isolate MXFP4 weight-cache namespace; gate a4w4 on grad_output quantize
  matthiasdiener
  committed
  b08bd19
- add tests
  matthiasdiener
  committed
  c29e03e
Commits on Oct 9, 2026
- use get_device_compute_capability for the gfx950 gate; drop ImportError guards
  matthiasdiener
  committed
  f7f98a5
- Merge remote-tracking branch 'origin/dev' into mdiener/mxfp4-a8w4-recipe
  matthiasdiener
  committed
  2c4172e
Commits on Oct 13, 2026
- fix post-2.19 tests
  matthiasdiener
  committed
  c4c0e64
Commits on Oct 19, 2026
Commits on Oct 20, 2026
Commits on Oct 27, 2026
- Merge remote-tracking branch 'origin/dev' into mdiener/mxfp4-a8w4-recipe
  matthiasdiener
  committed
  4c5a272
- centralize gfx1250 Triton knob override in _amd_knob_overrides
  matthiasdiener
  committed
  8780523

Commits

Commits on Sep 28, 2026

Commits on Sep 29, 2026

Commits on Oct 9, 2026

Commits on Oct 13, 2026

Commits on Oct 19, 2026

Commits on Oct 20, 2026

Commits on Oct 27, 2026