torch: put torch.sdpa on the cuDNN Python API (CUDNN provider + dense backward) - #554

Merged

vedaanta merged 7 commits into

NVIDIA:developfrom

vedaanta:vagarwalla/cudnn-torch-provider

Sep 1, 2026

Commits on Sep 1, 2026

torch: add the "CUDNN" torch.nn.attention provider (cudnn.torch)

vedaanta
and
claude
committed
b79b8dc
fix(torch): drop the 2*B D2H syncs from the varlen backward

vedaanta
and
claude
committed
e2f10b6
docs(agents): make the no-D2H rule verifiable, not just greppable

vedaanta
and
claude
committed
bc17fbb
feat(sdpa): dense BHSD backward in cudnn::sdpa_bwd

vedaanta
and
claude
committed
c6e5329
fix(sdpa): normalize dense operands; forward the deterministic flag

vedaanta
and
claude
committed
f83c131
torch: route the provider's dense backward through cudnn::sdpa_bwd

vedaanta
and
claude
committed
4cc8100
docs(agents): record the concrete test traps hit while wiring the torch ops

vedaanta
and
claude
committed
7a8e583