Skip to content

CUDA backend: keep wide attention heads that are not a power of 2 off the Triton SDPA kernel - #24427

Merged
shoumikhin merged 3 commits into
pytorch:mainfrom
shoumikhin:cuda-sdpa-wide-head
Nov 2, 2026
Merged

CUDA backend: keep wide attention heads that are not a power of 2 off the Triton SDPA kernel#24427
shoumikhin merged 3 commits into
pytorch:mainfrom
shoumikhin:cuda-sdpa-wide-head

Say only the head-size limit this change adds in the triton::sdpa doc…

82b00ec
Select commit
Loading
Failed to load commit list.
Meta CLA / Meta CLA Check succeeded Nov 2, 2026 in 0s

Contributor License Agreement is valid!

View more details on Meta CLA