Repository navigation
Fix Python 3.15 CUDA and Windows wheel builds - #23932
Merged
Merged
Conversation
Use the available torchaudio 2.11.1 release for Python 3.15 CUDA 13.2 installs. Persist the PyPI extra index from the shared wheel pre-build hook so CUDA and Windows smoke installs can resolve runtime dependencies. Authored with OpenAI Codex.
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/23932
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit f8bd186 with merge base cbcc5ef ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
Bot
added
the
CLA Signed
This PR needs a
|
rascani
approved these changes
Oct 16, 2026
No cp315 Windows pytorch-tokenizers wheel is published, so pip falls back to the old 1.0.1 source distribution in the plain smoke shell and cannot identify ClangCL. Build the pinned source first through the Visual Studio environment helper using the already installed build dependencies. Authored with OpenAI Codex.
Build the pinned tokenizer source from a temporary short path. MSBuild file tracking in CMake compiler probes fails on the nested source build path even inside the Visual Studio environment. Authored with OpenAI Codex.
JacobSzwejbka
changed the title
This branch was successfully deployed
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
Combine the remaining Python 3.15 CUDA and Windows wheel fixes, including the Windows tokenizer changes previously proposed in #23940.
CUDA 13.2 installs request torchaudio 2.11.0, which has no Python 3.15 wheel. Select the available 2.11.1 release for Python 3.15. Preserve the CUDA 13.4 package selection and source-pinned installs.
Move the PyPI extra index setting from the CPU Linux environment script into the shared wheel pre-build hook. Export it for the hook and persist it through
GITHUB_ENVso later CUDA Linux and Windows smoke installs can resolve dependencies such asexpecttestwhile using PyTorch's index. Preserve existing extra indices.PyPI has no cp315 Windows pytorch-tokenizers wheel, so pip falls back to an older source distribution whose isolated build cannot find the compiler in the smoke shell. Install the pinned tokenizer source through the existing Visual Studio environment helper with build isolation disabled. Copy it to a short temporary directory to keep CMake compiler probes below the Windows path limit. Dereference submodule symlinks during the copy because MSYS cannot recreate them on every Windows runner, and remove the copy after installation.
This provisions the tokenizer dependency for CI. PyPI still needs an upstream cp315 Windows tokenizer wheel for installation without building that dependency from source.
Validation
The 13 installer regression tests pass on Python 3.14 and 3.15. Lintrunner, bash syntax validation, and git diff checks pass. A source-copy check confirms that the two symlinks which failed on Windows CUDA become regular files with identical contents. Earlier checks verified torchaudio wheel availability and metadata, preservation of extra indices through the hook, and actual pip resolution of
expecttestagainst the PyTorch indices with PyPI enabled.The earlier full Python 3.15 wheel builds and smoke tests passed for Linux x86_64 CUDA 13.2 and 13.4 (https://github.com/pytorch/executorch/actions/runs/38936913096), Linux aarch64 CUDA 13.2 and 13.4 (https://github.com/pytorch/executorch/actions/runs/38936920119), and Windows CPU (https://github.com/pytorch/executorch/actions/runs/38961222041). Both Windows CUDA configurations then failed copying submodule symlinks before compilation (https://github.com/pytorch/executorch/actions/runs/38964126110); this combined revision includes the copy fix.
Fresh validation of the combined revision is running for all seven configurations. Windows CPU: https://github.com/pytorch/executorch/actions/runs/38968122117. Windows CUDA 13.2 and 13.4: https://github.com/pytorch/executorch/actions/runs/38968126439. Linux x86_64 CUDA 13.2 and 13.4: https://github.com/pytorch/executorch/actions/runs/38968132148. Linux aarch64 CUDA 13.2 and 13.4: https://github.com/pytorch/executorch/actions/runs/38968137580. These are pending results, not passes.
The validation branch has identical source fixes and temporarily restricts the wheel matrix and its guard to Python 3.15. It omits additional Windows CUDA program export jobs while retaining the wheel build and smoke suite. Those validation-only edits are not included in this PR.
Authored with OpenAI Codex.