Integrate Rubin wgrad cutedsl kernel and fix an issue in rebased rubin branch - #456

Merged

Anerudhan merged 10 commits into

NVIDIA:developfrom

WanZzzzzz:rubin-wgrad-rebased

Jul 30, 2026

Commits on Jul 29, 2026

rubin wgrad kernels
WanZzzzzz
committed
c6f787b
multiple of 256 for fp4 test
WanZzzzzz
committed
2f46834
fp4 specific issue
WanZzzzzz
committed
42c1135
create a fresh C TMA operation inside the helper kernel, matching upstream Rubin, and added MXFP8 discrete-accumulation coverage
WanZzzzzz
committed
72fce02
further fix
WanZzzzzz
committed
a4e8aa6
restricted the final store/reduce to the leader CTA
WanZzzzzz
committed
5a40c2b
uses the upstream-style V coordinate from block_idx_in_cluster
WanZzzzzz
committed
0e1bbf1
fix
WanZzzzzz
committed
5aaffee
fp4 discrete-accumulation coverage
WanZzzzzz
committed
21a8656

Commits on Jul 30, 2026

Fixed the import in test_grouped_gemm_wgrad.py
WanZzzzzz
committed
f1ffafa