Repository navigation

FlyDSL persistent work-stealing grouped GEMM enablement (gfx950) - #746

Open

aris134 wants to merge 28 commits into

devfrom

amartin/flydsl-mxfp8-group-gemm

Open

FlyDSL persistent work-stealing grouped GEMM enablement (gfx950)#746
aris134 wants to merge 28 commits into
devfrom
amartin/flydsl-mxfp8-group-gemm

Commits on Sep 29, 2026
- Add FlyDSL persistent work-stealing MXFP8 grouped GEMM (gfx950)
  
  aris134
  and
  claude
  committed
  287c1d1
Commits on Sep 30, 2026
- Add FlyDSL grouped MXFP8 dgrad (NN) backward path
  
  aris134
  and
  claude
  committed
  e4fc6bd
- Add FlyDSL grouped MXFP8 wgrad (NT) via uniform padding
  
  aris134
  and
  claude
  committed
  10f48f8
Commits on Oct 14, 2026
Commits on Oct 15, 2026
- Address grouped GEMM weights via per-expert pointers, dropping the host concat
  aris134
  committed
  14d85a4
- Enable FP16 grouped GEMM, add a FlyDSL dispatch table, and wire a test backend
  aris134
  committed
  4d5d448
Commits on Oct 16, 2026
- Cut FlyDSL grouped-GEMM host overhead across all dtypes
  aris134
  committed
  0dee39c
Commits on Oct 17, 2026
- Hoist _SpanCAI
  aris134
  committed
  3112c12
Commits on Oct 18, 2026
- Write FlyDSL FP8 grouped wgrad output in place, skip the copy-out loop
  aris134
  committed
  d290583
Commits on Oct 19, 2026
- Merge origin/dev into amartin/flydsl-mxfp8-group-gemm
  aris134
  committed
  35711ab
Commits on Oct 20, 2026
- Merge remote-tracking branch 'origin/dev' into amartin/flydsl-mxfp8-group-gemm
  aris134
  committed
  a710a9a
- Read FlyDSL FP8 grouped operands and scales in place via per-expert pointers
  aris134
  committed
  071d251
Commits on Oct 21, 2026
- Share per-expert in-place operand reads across all FlyDSL grouped dtypes
  aris134
  committed
  0c783e4
- Make FlyDSL grouped GEMM dispatch non-blocking and cheaper on the host
  aris134
  committed
  d14a87f
- remove redundant checks, clean up some code comments
  aris134
  committed
  bcab563
- address bot comments
  aris134
  committed
  ff139a4
- address bot comments
  aris134
  committed
  6635016
- Merge remote-tracking branch 'origin/dev' into amartin/flydsl-mxfp8-group-gemm
  aris134
  committed
  7f3d0ee
- address bot comments
  aris134
  committed
  e169938
- remove formatting-only changes from quantization.py
  aris134
  committed
  9efa0d8
- Replace deprecated fx.Index with fx.Int64 in FlyDSL GEMM kernels
  aris134
  committed
  2114ec4
- Merge remote-tracking branch 'origin/dev' into amartin/flydsl-mxfp8-group-gemm
  aris134
  committed
  954519c
Commits on Oct 27, 2026
- Merge remote-tracking branch 'origin/dev' into amartin/flydsl-mxfp8-group-gemm
  aris134
  committed
  5312860
Commits on Oct 28, 2026
- address reviewer comments
  aris134
  committed
  87b343c

Commits

Commits on Sep 29, 2026

Commits on Sep 30, 2026

Commits on Oct 14, 2026

Commits on Oct 15, 2026

Commits on Oct 16, 2026

Commits on Oct 17, 2026

Commits on Oct 18, 2026

Commits on Oct 19, 2026

Commits on Oct 20, 2026

Commits on Oct 21, 2026

Commits on Oct 27, 2026

Commits on Oct 28, 2026