Refactor cuda-bindings setup.py to defer CUDA_HOME check to build time by kkraus14 · Pull Request #1652 · NVIDIA/cuda-python

kkraus14 · 2026-02-19T05:34:53Z

Summary

Introduces cuda_bindings/build_hooks.py as a custom PEP 517 build backend that defers all CUDA-dependent logic (header parsing, code generation, cythonization) to build_wheel/build_editable
Rewrites cuda_bindings/setup.py as a thin shim (matching cuda_core/setup.py) that only provides the ext_modules bridge and parallel build_ext
Changes pyproject.toml build-backend from setuptools.build_meta to build_hooks

Metadata queries (uv lock, uv sync, pip install --dry-run) no longer require CUDA_HOME or CUDA_PATH to be set.

Follows the same pattern already used by cuda_core/build_hooks.py.

Test plan

Verify python -c "import build_hooks" succeeds without CUDA_HOME (confirmed locally)
Verify uv lock / uv pip compile works without CUDA_HOME
Verify CUDA_HOME=/usr/local/cuda uv build --wheel produces a working wheel
Verify pip install -e . with CUDA_HOME set works for editable installs
CI wheel builds pass on Linux and Windows

Closes #1635

Made with Cursor

copy-pr-bot · 2026-02-19T05:34:57Z

Auto-sync is disabled for ready for review pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

copy-pr-bot · 2026-02-19T05:38:38Z

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

Move all CUDA-dependent logic (header parsing, code generation, cythonization) from setup.py into a custom PEP 517 build backend (build_hooks.py). Metadata queries (uv lock, uv sync, pip install --dry-run) no longer require CUDA_HOME or CUDA_PATH to be set. This follows the same pattern already used by cuda_core/build_hooks.py: - Metadata hooks delegate directly to setuptools - build_wheel/build_editable defer to _build_cuda_bindings() - setup.py becomes a thin shim for ext_modules + parallel build_ext Closes NVIDIA#1635 Co-authored-by: Cursor <cursoragent@cursor.com>

kkraus14 · 2026-02-19T05:47:57Z

/ok to test

mdboom

A few minor comments, none critical.

cuda_bindings/build_hooks.py

- Move extra_link_args (strip) handling alongside extra_compile_args in the non-debug build path instead of a separate post-loop - Rename compiler_directives to cython_directives for clarity - Restore deprecated PARALLEL_LEVEL env var support in setup.py for build_ext parallelism parity with the old code Co-authored-by: Cursor <cursoragent@cursor.com>

kkraus14 · 2026-02-19T18:18:18Z

/ok to test

…ome-to-build-time Co-authored-by: Cursor <cursoragent@cursor.com> # Conflicts: # cuda_bindings/setup.py

kkraus14 · 2026-02-19T20:59:18Z

/ok to test

Check the compiler once in build_extensions (plural) and fix up all extensions in one pass, instead of checking per-extension in build_extension (singular). Co-authored-by: Cursor <cursoragent@cursor.com>

kkraus14 · 2026-02-19T21:01:52Z

/ok to test

kkraus14 · 2026-02-19T21:34:48Z

cuda_bindings/setup.py

+        if _is_clang(self.compiler):
+            for ext in self.extensions:
+                ext.extra_compile_args = [a for a in ext.extra_compile_args if a != "-fno-var-tracking-assignments"]


@mdboom I moved this so we only check this once instead of per extension. Would be great if you could take a look in case I have the wrong idea.

Yes, this should be fine. It would be nicer to do it where we set the flags in the first place, but we don't know what the compiler is yet at that point.

github-actions · 2026-02-20T14:00:15Z

Backport failed for 12.9.x, because it was unable to cherry-pick the commit(s).

Please cherry-pick the changes locally and resolve any conflicts.

git fetch origin 12.9.x
git worktree add -d .worktree/backport-1652-to-12.9.x origin/12.9.x
cd .worktree/backport-1652-to-12.9.x
git switch --create backport-1652-to-12.9.x
git cherry-pick -x 9e60c6fcff8ba2e44b7f54994db4a803d45e35f7

github-actions · 2026-02-20T14:13:52Z

Doc Preview CI
Preview removed because the pull request was closed or merged.

NVIDIA#1652) * Refactor cuda-bindings setup.py to defer CUDA_HOME check to build time Move all CUDA-dependent logic (header parsing, code generation, cythonization) from setup.py into a custom PEP 517 build backend (build_hooks.py). Metadata queries (uv lock, uv sync, pip install --dry-run) no longer require CUDA_HOME or CUDA_PATH to be set. This follows the same pattern already used by cuda_core/build_hooks.py: - Metadata hooks delegate directly to setuptools - build_wheel/build_editable defer to _build_cuda_bindings() - setup.py becomes a thin shim for ext_modules + parallel build_ext Closes NVIDIA#1635 Co-authored-by: Cursor <cursoragent@cursor.com> * Address review feedback and restore PARALLEL_LEVEL compat - Move extra_link_args (strip) handling alongside extra_compile_args in the non-debug build path instead of a separate post-loop - Rename compiler_directives to cython_directives for clarity - Restore deprecated PARALLEL_LEVEL env var support in setup.py for build_ext parallelism parity with the old code Co-authored-by: Cursor <cursoragent@cursor.com> * Move clang detection to build_extensions for single check Check the compiler once in build_extensions (plural) and fix up all extensions in one pass, instead of checking per-extension in build_extension (singular). Co-authored-by: Cursor <cursoragent@cursor.com> --------- Co-authored-by: Cursor <cursoragent@cursor.com>

#1652) (#1666) * Refactor cuda-bindings setup.py to defer CUDA_HOME check to build time Move all CUDA-dependent logic (header parsing, code generation, cythonization) from setup.py into a custom PEP 517 build backend (build_hooks.py). Metadata queries (uv lock, uv sync, pip install --dry-run) no longer require CUDA_HOME or CUDA_PATH to be set. This follows the same pattern already used by cuda_core/build_hooks.py: - Metadata hooks delegate directly to setuptools - build_wheel/build_editable defer to _build_cuda_bindings() - setup.py becomes a thin shim for ext_modules + parallel build_ext Closes #1635 * Address review feedback and restore PARALLEL_LEVEL compat - Move extra_link_args (strip) handling alongside extra_compile_args in the non-debug build path instead of a separate post-loop - Rename compiler_directives to cython_directives for clarity - Restore deprecated PARALLEL_LEVEL env var support in setup.py for build_ext parallelism parity with the old code * Move clang detection to build_extensions for single check Check the compiler once in build_extensions (plural) and fix up all extensions in one pass, instead of checking per-extension in build_extension (singular). --------- Co-authored-by: Cursor <cursoragent@cursor.com>

kkraus14 marked this pull request as draft February 19, 2026 05:38

kkraus14 force-pushed the kkraus14/defer-cuda-home-to-build-time branch from 30954a9 to dc8be9b Compare February 19, 2026 05:46

kkraus14 added the to-be-backported Trigger the bot to raise a backport PR upon merge label Feb 19, 2026

This comment has been minimized.

Sign in to view

kkraus14 marked this pull request as ready for review February 19, 2026 17:26

mdboom added the cuda.bindings Everything related to the cuda.bindings module label Feb 19, 2026

mdboom approved these changes Feb 19, 2026

View reviewed changes

cuda_bindings/build_hooks.py Outdated Show resolved Hide resolved

cuda_bindings/build_hooks.py Outdated Show resolved Hide resolved

Merge remote-tracking branch 'origin/main' into kkraus14/defer-cuda-h…

cadd2f2

…ome-to-build-time Co-authored-by: Cursor <cursoragent@cursor.com> # Conflicts: # cuda_bindings/setup.py

Move clang detection to build_extensions for single check

a8e3a99

Check the compiler once in build_extensions (plural) and fix up all extensions in one pass, instead of checking per-extension in build_extension (singular). Co-authored-by: Cursor <cursoragent@cursor.com>

kkraus14 commented Feb 19, 2026

View reviewed changes

rparolin assigned kkraus14 Feb 20, 2026

kkraus14 merged commit 9e60c6f into NVIDIA:main Feb 20, 2026
86 checks passed

kkraus14 mentioned this pull request Feb 20, 2026

[12.9.x backport] Refactor cuda-bindings setup.py to defer CUDA_HOME check to build time #1666

Merged

1 task

Comments

Conversation

kkraus14 commented Feb 19, 2026

Summary

Test plan

Uh oh!

copy-pr-bot bot commented Feb 19, 2026

Uh oh!

copy-pr-bot bot commented Feb 19, 2026

Uh oh!

kkraus14 commented Feb 19, 2026

Uh oh!

This comment has been minimized.

mdboom left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

kkraus14 commented Feb 19, 2026

Uh oh!

kkraus14 commented Feb 19, 2026

Uh oh!

kkraus14 commented Feb 19, 2026

Uh oh!

kkraus14 Feb 19, 2026

Choose a reason for hiding this comment

Uh oh!

mdboom Feb 20, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

github-actions bot commented Feb 20, 2026

Uh oh!

github-actions bot commented Feb 20, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants