Add CUDA 13.4 to the nightly domain-library build matrix - #8477
Merged
Conversation
torchvision and torchaudio take their CUDA matrix from here rather than from pytorch/pytorch, so CUDA 13.4 has to be added separately. pytorch added it in #177316's successor; the torch wheels are built and published for linux x86 and sbsa, so the domain libraries now have something to build against. Nightly only: CUDA_ARCHES_DICT["test"] and ["release"] are left alone, and STABLE_CUDA_VERSIONS stays at 13.0. The container images this generates, manylinux2_28-builder:cuda13.4 and manylinuxaarch64-builder:cuda13.4, are both built by pytorch/pytorch's build-manywheel-images.yml. Windows is excluded. pytorch sets CUDA_ARCHES_NO_WINDOWS = ["13.4"] because there is no Windows CUDA 13.4 installer on the ossci-windows bucket, so no Windows torch wheel exists for cu134. Without the matching exclusion here the generator emits eight Windows cu134 entries whose builds would fail resolving torch. Mirror the constant and filter it out of the Windows arches for both the wheel and libtorch matrices. Test Plan: Reference files regenerated and the suite passes: ``` PYTHONPATH=. python3 tools/tests/test_generate_binary_build_matrix.py --update-reference-files PYTHONPATH=. python3 -m unittest discover -s tools/tests -p "test_generate_binary_build_matrix.py" ``` 11 tests, OK. cu134 entry counts after the change: linux 8, linux-aarch64 8, windows 0. The aarch64 entries resolve to manylinuxaarch64-builder:cuda13.4, which is the path that matters for Grace/Vera-class ARM hosts. Authored with assistance from an AI coding assistant (Claude). Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
@tinglvv is attempting to deploy a commit to the Meta Open Source Team on Vercel. A member of the Team first needs to authorize it. |
meta-cla
Bot
added
the
CLA Signed
…pute_50'
get_cuda_arch_list() matched "13.0" and "13.2" exactly, so 13.4 fell through to
the legacy fallback list "5.0;6.0;7.0;7.5;8.0;8.6;9.0". CUDA 13 removed sm_50
and sm_60, so every CUDA translation unit failed:
/usr/local/cuda-13.4/bin/nvcc ...
nvcc fatal : Unsupported gpu architecture 'compute_50'
subprocess.CalledProcessError: Command '['ninja', '-v']' returned non-zero exit status 1
RuntimeError: Error compiling objects for extension
seen on pytorch/vision build-wheel-py3_10-cuda-aarch6413_4-aarch64.
Match the whole 13.x series with startswith("13.") rather than adding 13.4 to
the exact-match chain, so the next CUDA minor does not silently fall through to
a list containing architectures the toolkit no longer supports.
Add direct coverage for the arch list on 13.0/13.2/13.4 x linux/linux-aarch64:
the existing tests compare get_cuda_variables() against get_cuda_arch_list()
output, so they cannot catch a wrong list.
atalman
approved these changes
Aug 10, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
torchvision and torchaudio take their CUDA matrix from here rather than from pytorch/pytorch, so CUDA 13.4 has to be added separately. pytorch added it in #177316's successor; the torch wheels are built and published for linux x86 and sbsa, so the domain libraries now have something to build against.
Nightly only: CUDA_ARCHES_DICT["test"] and ["release"] are left alone, and STABLE_CUDA_VERSIONS stays at 13.0. The container images this generates, manylinux2_28-builder:cuda13.4 and manylinuxaarch64-builder:cuda13.4, are both built by pytorch/pytorch's build-manywheel-images.yml.
Windows is excluded. pytorch sets CUDA_ARCHES_NO_WINDOWS = ["13.4"] because there is no Windows CUDA 13.4 installer on the ossci-windows bucket, so no Windows torch wheel exists for cu134. Without the matching exclusion here the generator emits eight Windows cu134 entries whose builds would fail resolving torch. Mirror the constant and filter it out of the Windows arches for both the wheel and libtorch matrices.
Test Plan:
Reference files regenerated and the suite passes:
Authored with assistance from an AI coding assistant (Claude).
cc @atalman