[rwgk]

Description

Closes #3246.

Follow-on to #3209.

Run cython-lint serially so its cross-file rules see matching .pxd declarations and .pyx implementations together. With parallel pre-commit batches, the same files can pass or fail depending on the worker count because the linter keeps this information within one process.

Remove the ineffective inline keyword from the three stream helpers identified in the issue (Stream_handle_is_default_token, Stream_is_default_token, and Stream_is_legacy_default_token) and from _resolve_memcpy_operand in _graph_node.pyx, which serial linting also exposed. Their signatures, implementations, and exception/GIL specifications otherwise stay the same.

Runtime impact

On my main development machine, the added require_serial: true adds about 1.3 seconds to an all-files run for cython-lint only, with negligible impact on small commits.

Measured wall-clock times, using warm hook environments and the median of three runs:

Files checked	Parallel	Serial
All 108 eligible files	1.27 s	2.53 s
Two files (_stream.pyx + .pxd)	0.71 s	0.72 s

My main development machine has 128 Xeon CPUs and uses pre-commit 3.6.2. For up to four files, pre-commit already uses one process, so ordinary small commits see essentially no difference.

pre-commit run --all-files for all hooks except lychee:

• Before this PR: 15.4 s
• After this PR: 17.5 s

Validation

• Full pre-commit run --all-files.
• A focused cython-lint check reports all four violations against the original source and passes against the updated source.

Full CI is deferred while this PR is a draft.

Checklist

• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[copy-pr-bot]

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.

[coderabbitai]

No actionable comments were generated in the recent review.

Recent review info

Run configuration

• Source mode: Auto
• Repository: NVIDIA/cuda-python
• Branch: rwgk/fix-cython-lint-cross-file-checks
• File: .coderabbit.yaml
• Commit: 764873b
• Layers: Repository YAML
• Review profile: CHILL
• Plan: Enterprise
• Run ID: 9621772c-aa8f-45ef-9898-e082ef636672

Commits

Reviewing files that changed from the base of the PR and between f9212dd and 764873b.

Files selected for processing: 3

• .pre-commit-config.yaml
• cuda_core/cuda/core/_stream.pyx
• cuda_core/cuda/core/graph/_graph_node.pyx

Included review availability: This review used your included allowance. Your plan provides up to 12 included reviews per hour; 11 remain after this review.

📝 Summary

Summary by CodeRabbit

• Chores
  • Code checks now process related files sequentially, providing more consistent validation.
  • No changes to application behavior or features are included in this update.

The cython-lint pre-commit hook now runs serially. The inline qualifier was removed from two stream helpers and one graph helper. Stream_is_legacy_default_token remains unchanged.

Changes: Cython lint remediation

Layer / File(s)	Summary
Run cython-lint serially .pre-commit-config.yaml	The cython-lint hook now runs serially so its cross-file checks process declarations and implementations in one invocation.
Remove inline qualifiers cuda_core/cuda/core/_stream.pyx, cuda_core/cuda/core/graph/_graph_node.pyx	The inline qualifier was removed from Stream_handle_is_default_token, Stream_is_default_token, and _resolve_memcpy_operand. The adjacent Stream_is_legacy_default_token declaration remains unchanged.

Assessment against linked issues

Objective	Addressed	Explanation
Fix the cross-file lint blind spot and remove the reported ineffective inline qualifiers [#3246]	❌	Stream_is_legacy_default_token remains inline, so one reported violation is not removed.

Out-of-scope changes

Code Change	Explanation
Remove inline from _resolve_memcpy_operand (cuda_core/cuda/core/graph/_graph_node.pyx:926)	Issue #3246 does not identify this helper. It may be another instance of the same lint violation, but the supplied context does not establish that connection.

Possibly related PRs

• NVIDIA/cuda-python#3209: It introduced the cython-lint version whose cross-file rule is involved here.
• NVIDIA/cuda-python#3173: It expanded cython-lint coverage to CUDA 12 and CUDA 13 source trees, which are subject to the serial execution change.

Suggested reviewers: andy-jost, mdboom

Priority: Low

Change: Bug fix · Severity of issue fixed: Low

Merge Risk: Minimal · up to 76487

No actionable merge-blocking risk remains in the reviewed changes.

- [ ] Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

_{Comment @coderabbitai help to get the list of available commands.}

[mdboom]

The way these are being fixed here is a performance regression. Prior to this change, _resolve_memcpy_operand was inlined within other functions in the same _graph_node.pyx file, but it wasn't inlined where it was cimport'ed from _subclasses.pyx. Now, it's not inlined anywhere.

Instead, we need to move the function definitions entirely to the .pxd files which should silence the lint (haven't checked) but also keep these functions inlined as intended.

There is an additional bug in the precommit hook. While putting all of the files in a single process /helps/ it doesn't entirely fix the bug. If only the .pyx file or the .pxd file changes, only the changed files are sent to the precommit hook, so again the thing this is trying to detect could slip through the cracks. The fix is to either run all files all the time or, if possible, configure precommit to always look at .pyx/.pxd pairs.