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

Commits

  1. Commits on Nov 2, 2026