[CI] Add sm80 to the clang20 CUDA pull builds' arch list (#194685)
The two `linux-jammy-cuda13.{0,2}-cudnn9-py3.10-clang20` build jobs never set `cuda-arch-list`, so they inherit `_linux-build.yml`'s default of `7.5`. Since #194502 disables `USE_FLASH_ATTENTION` when no resolved arch reaches sm80, they lost the SYSTEM marking that `caffe2/CMakeLists.txt` puts on `third_party/cutlass/include`, and clang-20 started promoting `-Wunused-function` in cutlass headers to errors.
Requesting `8.0` restores the SYSTEM include; `7.5` stays so these builds still cover Turing.
Related: #194645 also touches how the cutlass include dirs are marked.
Authored with assistance from Claude (Claude Code).
Pull Request resolved: https://github.com/pytorch/pytorch/pull/194685
Approved by: https://github.com/pytorchgreenlight H
Huy Do committed
cf8577d42b4750c22ad369fd8cef2b8e84bce74e
Parent: 0800934
Committed by PyTorch MergeBot <pytorchmergebot@users.noreply.github.com>
on 8/25/2026, 2:21:58 AM