feat: add SM_121 (GB10 consumer Blackwell) support for FA4#3125

Open

TyGu1 wants to merge 7 commits into

NVIDIA:mainfrom

TyGu1:feature/sm121-blackwell-support

Commits on Jun 12, 2026

feat: add SM_121 (GB10 consumer Blackwell) support for FA4
TyGu1
committed
fix: address SM_121 code review findings
TyGu1
committed
build: find NCCL headers unconditionally for pip-wheel CUDA installs

TyGu1
and
claude
committed
build: add get_nccl_include_dirs() for PyTorch extension on pip-wheel CUDA

TyGu1
and
claude
committed
build: fix get_cuda_include_dirs() to supplement toolkit with pip wheel includes

TyGu1
and
claude
committed
test: scope test_dpa_fa4_sm121 to MHA-only configs; add test results

TyGu1
and
claude
committed
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot]
committed