Add FlyDSL grouped groupwise FP8 GEMM for ROCm - #456

Closed

aryaman-gupta wants to merge 18 commits into

meta-pytorch:mainfrom

aryaman-gupta:flydsl-groupwise-grouped

Commits on Jul 27, 2026

Vendor shared FlyDSL kernel infra modules

aryaman-gupta
and
claude
committed
3f8c943
Add FlyDSL f8f8bf16_groupwise_grouped_preshuffle op

aryaman-gupta
and
claude
committed
5dbe52c
Add plain-B FlyDSL grouped GEMM path + wide-MFMA SW scaling

aryaman-gupta
and
claude
committed
16fe99a
Back the plain f8f8bf16_groupwise_grouped op with FlyDSL

aryaman-gupta
and
claude
committed
0428482
Resolve grouped-GEMM group ownership in-kernel; add tuned tile tables

aryaman-gupta
and
claude
committed
876b616
Tune grouped GEMM tiles with FlyDSL autotune behind an env gate

aryaman-gupta
and
claude
committed
6e8837a
Move FlyDSL helper imports to mslk.flydsl.common / mslk.flydsl.jit
aryaman-gupta
committed
ee47260
Guard the preshuffle LDS budget and make the limit arch-aware
aryaman-gupta
committed
e466424
Cover the FlyDSL plain grouped GEMM from gemm_test
aryaman-gupta
committed
76a4dd0
Read scale_a from its per-group layout in the grouped GEMM
aryaman-gupta
committed
048277f
Refresh grouped GEMM docs and drop the unused kernel argument
aryaman-gupta
committed
b2c7f30
Validate preshuffle_b_mfma shapes and fix its tile extents
aryaman-gupta
committed
79def36
Validate operand dtypes and read m_sizes as int64 in the grouped GEMM
aryaman-gupta
committed
75fa3ac

Commits on Jul 28, 2026

Commits on Jul 31, 2026

Run ROCm CI when the FlyDSL GEMM ops change
aryaman-gupta
committed
8a0e6a2