Permute-Free Grouped GEMM for MoE (bf16, gfx950) - #694

Draft

sudhu2k wants to merge 44 commits into

devfrom

sudhu/permute-free-groupedgemm-flydsl

Commits on Jul 22, 2026

Add optional FlyDSL dependency for ROCm PyTorch builds via NVTE_USE_FLYDSL
aris134
committed
cf54ce6
Wire FlyDSL into the TransformerEngine MXFP8 GEMM dispatch path
aris134
committed
2ed8285
Add initial support for TN BF16 FlyDSL GEMM backend
aris134
committed
17b9b74
Add initial support for TN FP16 FlyDSL GEMM backend
aris134
committed
155e81e
Add initial support for TN FP8 FlyDSL GEMM backend
aris134
committed
492a29c
Add initial support for TN FP32 FlyDSL GEMM backend
aris134
committed
6ef76b3
add layout support for NN/NT FlyDSL GEMM
aris134
committed
9a9462e
add support for bf16/fp32 output types for flydsl mxfp8 gemm
aris134
committed
4c9543f
add support for bf16/fp32 output types for flydsl fp8 gemm
aris134
committed
b9eaa8a
add broad output dtype support for flydsl fp8/fp16/bf16 gemms
aris134
committed
71c4ef4
add mixed e4m3/e5m2 fp8 dtype support for flydsl fp8 GEMM
aris134
committed
aa19610
add mixed e4m3/e5m2 fp8 dtype support for flydsl mxfp8 GEMM
aris134
committed
2ee10d3
add pytorch flydsl gemm tests
aris134
committed
979f38c
add the actual pytorch flydsl gemm tests
aris134
committed
e1896cd
add e2e flydsl test in test_numerics
aris134
committed
838e64f
add flydsl gemm fallback support
aris134
committed
9670650

Commits on Jul 23, 2026

Add direct FP8 NN/NT FlyDSL GEMM specializations
aris134
committed
ede7ace

Commits on Jul 28, 2026

feat(flydsl): add FP8 NN and NT GEMM specializations
aris134
committed
842bfa2
Add direct MXFP8 NN/NT FlyDSL GEMM specializations
aris134
committed
1ba15ed
gemm wrappers patch
aris134
committed
2524a08
correct FP8 NT storage contract
aris134
committed
de6d22a
Unify FP8 TN, NN, and NT GEMM paths
aris134
committed
38ccfb2
unify mxfp8 gemm shape variants
aris134
committed
f1a5213
FlyDSL MXFP8: fuse scale prepacking into dedicated GPU kernels
aris134
committed
d36c6c2

Commits on Jul 29, 2026

Normalize MXFP8 TN to direct rowwise storage
aris134
committed
fbc6862
improve backward bf16 gemms with shape specialization
aris134
committed
4d7d273
add shape specialization for flydsl fp16 gemm backend
aris134
committed
e928833
add todo comments
aris134
committed
9d6cbfb
add missing EOLs
aris134
committed
e5f4a0c
Merge remote-tracking branch 'origin/dev' into amartin/flydsl-gemm-integration
aris134
committed
ef46d52
remove calls to old utility get_tolerances and use dtype_tols instead for flydsl test
aris134
committed
0948d9a
add gpu arch gating to flyDSL GEMM backend enablement
aris134
committed
63c5c4c

Commits on Aug 3, 2026

Pin FlyDSL below 0.3 for buffer_ops compatibility
aris134
committed
12d06c7

Commits on Aug 5, 2026

Commits on Aug 6, 2026