[rwgk]

Summary

Integrate the imported CUDA 12 sources with the shared build and test tooling, while retaining the CUDA 12 header parser, runtime templates, and legacy modules. This layer carries forward the reviewed source adaptations from #2737; it does not activate CUDA 12 in CI.

• Use the canonical shared build helper, package it in standalone sdists, and preserve release-tag README links.
• Retain the Windows SDK constant declaration required by Cython warnings-as-errors. Independent runtime maintenance and its regression tests are isolated in cuda.bindings: backport CUDA 12 runtime maintenance fixes #3182.
• Adapt test fixtures, cuFile isolation, and documentation paths. NVML runner/test maintenance is isolated in CI: test and document both bindings majors on main #3176. Preserve main's existing CUDA 13 and Core test safeguards.
• Port test(cuda.bindings): reuse worker-thread contexts under pytest-run-parallel #3122's thread-local stream capture to three CUDA 12 tests and recognize both bindings roots with the shared bindings pytest marker.
• Enable CUDA 12 pre-commit coverage in this layer, with narrow exceptions for generated sources and historical documentation.
• Remove obsolete package-local benchmarks in favor of the shared harness, and document Python 3.15 support.
• Keep README, PyPI description, and build-error documentation links on the published CUDA 12.9.7 snapshot; latest follows the current CUDA major.

The preceding import remains the exact upstream source snapshot. This diff isolates its integration changes.

Stack / Review Boundary

Layer 2 of 8. Previous: #3172. Next: #3174. All layers and integrated validation.

The eight PRs are review boundaries for one change. #3182 contains the final integrated source state; earlier integrated validation is recorded in #3178. The stack is rebased onto main dfad7e3658a9. Full hosted CI has not been rerun after this rebase.

Suggested Review Order

1. cuda_bindings_12/build_hooks.py
2. cuda_bindings_12/_build_shared.py
3. cuda_bindings_12/tests/conftest.py
4. shared pytest marker classification

Validation

At this PR's head, full local pre-commit run --all-files --show-diff-on-failure passes, including CUDA 12 checks and Lychee. Direct HTTP checks verified the pinned documentation landing page, overview, installation page, and source-build anchors; the 12.9.8–12.9.10 overview and installation pages return 404. The comments explain the intentional snapshot and identify the copies to keep aligned. No full CI was triggered.

Earlier local checks passed the CUDA 12 build-helper suite (75 passed, 3 skipped), validated all 133 generated-file seals across both roots, and rejected simulated seal removal or tampering. A focused Cython 3.2.9 translation check confirms that the Windows const declaration is required when warnings are errors.

Earlier integrated-stack validation: 279 focused build tests passed (9 skipped), all 62 sealed CUDA 12 generated files match the imported upstream versions, and shared pytest setup was compared across all three copies. The three capture tests also passed 360 local parallel invocations across 2/4 threads and normal/PTDS stream modes; six path-classification checks cover the shared marker update. Hosted validation passed every CUDA 12 row, including all free-threaded Python targets. Earlier native validation and remaining hosted checks are recorded in #3178; #3182 contains the final source state after this scope split.

Checklist

• Relevant existing and focused regression tests cover the integration behavior.
• This description identifies the layer's scope and validation boundary.

[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.

[mdboom]

A bunch of changes here seem to be moving to new precommit linter/formatter rules. That is fine, but we should re-enable the running of pre-commit on these files in this PR as well if that's the case. (It was disabled on the cuda_bindings_12 directory in the previous PR in this stack).

[mdboom]

Ditto.

[rwgk]

codex GPT-6.1-Sol ultra on behalf of Ralf:

This declaration is needed for the build integration: #3173 enables Cython warnings-as-errors, and #2463 originally introduced that setting together with the const DWORD declaration.

I checked Cython 3.2.9 translation with warnings treated as errors. The imported declaration fails with Assignment in pxd file will not be executed. Suggest declaring as const.; the retained declaration passes. This check requires no Windows SDK or native compilation.

The paired changes are the Windows SDK constant declaration and Cython warning policy.

I have retained it here and left this thread open for your review.

[rwgk]

@mdboom I wasn't actually sure myself if this PR is independently mergeable (I explicitly told my agent that's not requirement when building this stack initially), but below is what my agent thinks after analyzing your feedback. I'll work on the "Agree" items first. Please let me know any feedback on the other items. I'll try to work through all bit by bit, based on the information I'll have as I get to each part.

---

codex GPT-6.1-Sol ultra

I would address the feedback before merging #3173. It appears safe to land as inactive preparation, but it is not yet a self-contained, validated CUDA 12 integration.

I fetched all 11 inline comments, Mike’s review summary, and the conversation. Mike submitted his review with GitHub’s “Changes requested” status.

Independent Mergeability

Two distinctions matter:

1. Building CUDA 12 directly still gives the wrong package version. Its SCM configuration still matches every v* tag. At the exact cuda.bindings: integrate CUDA 12 source with shared build tooling #3173 head, git describe selects v13.4.3-375-g494296d9fd0. The version selection and other packaging corrections arrive in cuda.bindings: configure independent CUDA 12 and CUDA 13 packages #3174. This is inherited from the inactive import, but means cuda.bindings: integrate CUDA 12 source with shared build tooling #3173 alone is not a usable CUDA 12 package. Current SCM configuration

2. Pre-commit still excludes the entire CUDA 12 tree. Mike is right: we are adapting files for modern tooling while their checks remain disabled until cuda.bindings: configure independent CUDA 12 and CUDA 13 packages #3174. The green Linux/Windows pre-commit checks therefore do not validate those files. The separate security scan has been restored. Current exclusion

I found no dependency on later-layer code that would break existing CUDA 13 functionality. The shared build helper already exists on main; the changes outside CUDA 12 are additive or comments.

Mike’s Comments

Feedback	Assessment and recommended response
Re-enable pre-commit	Agree. Move the necessary lint configuration and narrow historical/generated-file exceptions from #3174 into this PR.
Delete cuda_bindings_12/benchmarks	Agree. The shared benchmark harness supersedes it. Update the README instructions alongside the deletion.
Byte-string compiler options are unrelated and lack tests	The scope concern is valid. This backports #1665, and tests are included, in test_nvjitlink.py and test_nvvm.py. Nevertheless, it is an independent behavior improvement rather than migration plumbing.
Coredump changes in both utils.pxd and utils.pxi	These retain borrowed setter data and free getter-owned storage, carried forward from the #2399 maintenance work. Useful fixes, but independently reviewable runtime changes.
Windows SDK constant declaration	Keep and explain. This comes from #2463 and supports the newly enabled Cython warnings-as-errors setting. It directly relates to build integration.
Handle-getter exception handling	Backports #2551, with a focused regression test already included. Again, useful but independent of migration.
NVML test changes	They incorporate current main’s accommodations for restricted device visibility, unsupported PCI information, and newer hardware. That explains their origin, but does not establish that all are necessary in this integration PR. Place them with CI activation or explain the specific runner requirement.
Python support should extend through 3.15 — both pages	Agree. The imported 12.9.10 release notes already announce Python 3.15 wheels. Coordinate the later package classifiers too.
Hardcoded installation URL in build_hooks.py	Valid maintenance concern. However, blindly using latest sends CUDA 12 users to CUDA 13 documentation. Preserve an appropriate CUDA 12 target unless the linked guidance is demonstrably version-independent.
Hardcoded README documentation version	Agree. It hardcodes 12.9.7, while the build error links to 12.9.10. Address README and DESCRIPTION.rst consistently. #3165’s release-tag rewriting covers GitHub source links, not these documentation URLs.

Direction I Recommend

Keep #3173 focused on independently safe preparation:

• Enable its pre-commit coverage now.
• Retain build integration and changes demonstrably required by it.
• Move the separable bytes, coredump, and handle-getter backports—with their tests—into an explicit maintenance layer.
• Place runner-specific test modernization with CI activation where practical.
• Address the benchmark and documentation feedback.

Version/package selection can remain in #3174 if #3173 explicitly remains inactive preparation. If we want developers to build and use CUDA 12 immediately after merging #3173, its minimum packaging corrections must move forward too.

The original stack deliberately allowed intermediate review boundaries that were not fully integrated. Now that we are merging individual layers, tightening this boundary would make the review substantially clearer.

[rwgk]

codex GPT-6.1-Sol ultra on behalf of Ralf:

The first three agreed follow-ups are pushed, in separate commits:

• Remove the obsolete CUDA 12 benchmarks, with a README pointer to the shared harness.
• Document Python 3.15 support in both reviewed pages. The matching classifier is updated in cuda.bindings: configure independent CUDA 12 and CUDA 13 packages #3174.
• Enable CUDA 12 pre-commit coverage in this PR, addressing the review's pre-commit request. The blanket exclusion is removed. Generated sources and historical MyST documentation retain narrow exceptions, while SPDX and generated-file seal checks cover both bindings roots.

Full local pre-commit run --all-files --show-diff-on-failure passes at this PR's head, including Lychee. The CUDA 12 build-helper tests pass (75 passed, 3 skipped); all 133 generated-file seals validate across both roots, and simulated removal or tampering of a CUDA 12 seal is rejected. Full local pre-commit also passes on the restacked #3174.

The remaining stack is restacked and published. Package/version selection and CUDA 12 CI activation remain in their later layers. The runtime-scope and documentation-URL discussions remain open. No full CI run has been triggered for this update.

[rwgk]

codex GPT-6.1-Sol ultra on behalf of Ralf:

The next scope-reduction pass is pushed:

• Draft cuda.bindings: backport CUDA 12 runtime maintenance fixes #3182 now contains the three independent CUDA 12 runtime backports, in three focused commits with their existing regression coverage and upstream provenance.
• The four NVML test/helper adaptations move unchanged to CI: test and document both bindings majors on main #3176, alongside activation of those tests on the current CI runners.
• The Windows const DWORD declaration stays here because the paired Cython warnings-as-errors setting otherwise rejects the imported declaration. A Cython 3.2.9 translation check confirms this; the thread remains open for Mike's review.

The stack is restacked and published, and the affected PR descriptions reflect the new boundaries. The new top's complete Git tree is identical to the pre-extraction integrated snapshot: no final source, test, packaging, or CI behavior changed.

Full local pre-commit passes on both the narrowed #3173 and the new top. The #3173 build-helper suite passes with 75 passed and 3 skipped. No full CI was triggered. The five relocated-change threads are resolved; the Windows declaration and two documentation-URL discussions remain open.

[rwgk]

codex GPT-6.1-Sol ultra on behalf of Ralf:

Implemented the agreed documentation policy in dc6166b31297: README, the PyPI description, and build-error guidance consistently use the published CUDA 12.9.7 snapshot. Comments explain why the links are pinned and identify the three files to keep aligned.

Direct HTTP checks verified the landing page, overview, installation page, and source-build anchors. Full local pre-commit passes on both #3173 and the final #3182 head, including Lychee. Exact Git tree comparisons confirm that every restacked layer gained only this three-file documentation change.

The stack is published and the PR descriptions are updated. Both documentation-URL threads are resolved; the Windows declaration thread remains open for Mike's review. All live PRs remain draft, and no full CI was triggered.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

• Configuration used: Repository: NVIDIA/cuda-python/.coderabbit.yaml
• Review profile: CHILL
• Plan: Enterprise
• Run ID: 48795a97-1577-472b-85fa-6850adbf9e6d

📥 Commits

Reviewing files that changed from the base of the PR and between dfad7e3 and 2f01e98.

📒 Files selected for processing: 84

• .github/labeler.yml
• .github/workflows/bandit.yml
• .gitignore
• .pre-commit-config.yaml
• benchmarks/cuda_bindings/runner/runtime.py
• cuda_bindings_12/DESCRIPTION.rst
• cuda_bindings_12/LICENSE
• cuda_bindings_12/MANIFEST.in
• cuda_bindings_12/README.md
• cuda_bindings_12/_build_shared.py
• cuda_bindings_12/benchmarks/conftest.py
• cuda_bindings_12/benchmarks/kernels.py
• cuda_bindings_12/benchmarks/pytest.ini
• cuda_bindings_12/benchmarks/test_cupy.py
• cuda_bindings_12/benchmarks/test_launch_latency.py
• cuda_bindings_12/benchmarks/test_numba.py
• cuda_bindings_12/benchmarks/test_pointer_attributes.py
• cuda_bindings_12/build_hooks.py
• cuda_bindings_12/cuda/bindings/_internal/utils.pxd
• cuda_bindings_12/cuda/bindings/_lib/windll.pxd
• cuda_bindings_12/docs/README.md
• cuda_bindings_12/docs/build_docs.sh
• cuda_bindings_12/docs/source/_static/javascripts/version_dropdown.js
• cuda_bindings_12/docs/source/_templates/sidebar/variant-selector.html
• cuda_bindings_12/docs/source/contribute.md
• cuda_bindings_12/docs/source/install.md
• cuda_bindings_12/docs/source/motivation.md
• cuda_bindings_12/docs/source/overview.md
• cuda_bindings_12/docs/source/release.rst
• cuda_bindings_12/docs/source/release/11.7.1-notes.md
• cuda_bindings_12/docs/source/release/11.8.0-notes.md
• cuda_bindings_12/docs/source/release/12.9.1-notes.rst
• cuda_bindings_12/docs/source/release/12.9.2-notes.rst
• cuda_bindings_12/docs/source/release/12.9.3-notes.rst
• cuda_bindings_12/docs/source/release/12.9.4-notes.rst
• cuda_bindings_12/docs/source/release/12.9.5-notes.rst
• cuda_bindings_12/docs/source/release/12.9.6-notes.rst
• cuda_bindings_12/docs/source/release/12.9.7-notes.rst
• cuda_bindings_12/docs/source/release/12.X.Y-notes.rst
• cuda_bindings_12/docs/source/support.rst
• cuda_bindings_12/docs/source/tips_and_tricks.rst
• cuda_bindings_12/docs/versions.json
• cuda_bindings_12/examples/0_Introduction/clock_nvrtc_test.py
• cuda_bindings_12/examples/0_Introduction/simpleCubemapTexture_test.py
• cuda_bindings_12/examples/0_Introduction/simpleP2P_test.py
• cuda_bindings_12/examples/0_Introduction/simpleZeroCopy_test.py
• cuda_bindings_12/examples/0_Introduction/systemWideAtomics_test.py
• cuda_bindings_12/examples/0_Introduction/vectorAddDrv_test.py
• cuda_bindings_12/examples/0_Introduction/vectorAddMMAP_test.py
• cuda_bindings_12/examples/2_Concepts_and_Techniques/streamOrderedAllocation_test.py
• cuda_bindings_12/examples/3_CUDA_Features/globalToShmemAsyncCopy_test.py
• cuda_bindings_12/examples/3_CUDA_Features/simpleCudaGraphs_test.py
• cuda_bindings_12/examples/4_CUDA_Libraries/conjugateGradientMultiBlockCG_test.py
• cuda_bindings_12/examples/common/common.py
• cuda_bindings_12/examples/common/helper_cuda.py
• cuda_bindings_12/examples/common/helper_string.py
• cuda_bindings_12/examples/extra/isoFDModelling_test.py
• cuda_bindings_12/examples/extra/jit_program_test.py
• cuda_bindings_12/examples/extra/numba_emm_plugin.py
• cuda_bindings_12/examples/pytest.ini
• cuda_bindings_12/pyproject.toml
• cuda_bindings_12/setup.py
• cuda_bindings_12/site-packages/_cuda_bindings_redirector.py
• cuda_bindings_12/tests/conftest.py
• cuda_bindings_12/tests/cython/build_tests.bat
• cuda_bindings_12/tests/cython/build_tests.py
• cuda_bindings_12/tests/cython/build_tests.sh
• cuda_bindings_12/tests/pytest.ini
• cuda_bindings_12/tests/test_build_hooks.py
• cuda_bindings_12/tests/test_cuda.py
• cuda_bindings_12/tests/test_cudart.py
• cuda_bindings_12/tests/test_cufile.py
• cuda_bindings_12/tests/test_graphics_apis.py
• cuda_bindings_12/tests/test_kernelParams.py
• cuda_bindings_12/tests/test_nvfatbin.py
• cuda_bindings_12/tests/test_utils.py
• cuda_bindings_13/MANIFEST.in
• cuda_bindings_13/tests/conftest.py
• cuda_core/MANIFEST.in
• cuda_core/tests/conftest.py
• cuda_python_test_helpers/cuda_python_test_helpers/_pytest_plugin.py
• ruff.toml
• toolshed/check_generated_file_seals.py
• toolshed/check_spdx.py

💤 Files with no reviewable changes: 8

• cuda_bindings_12/benchmarks/pytest.ini
• cuda_bindings_12/docs/versions.json
• cuda_bindings_12/benchmarks/test_cupy.py
• cuda_bindings_12/benchmarks/test_numba.py
• cuda_bindings_12/benchmarks/conftest.py
• cuda_bindings_12/benchmarks/test_pointer_attributes.py
• cuda_bindings_12/benchmarks/test_launch_latency.py
• cuda_bindings_12/benchmarks/kernels.py

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

• New Features

  • Added CUDA 12.9.1–12.9.7 release documentation, including compatibility notes, fixes, and upgrade guidance.
  • Added stricter CUDA header and toolchain validation during builds.
  • Added support for Python 3.10–3.15 in CUDA 12 documentation.
  • Improved CUDA context creation compatibility across CUDA versions.

• Documentation

  • Updated installation, version-selection, benchmark, and published documentation links.
  • Centralized documentation version data across CUDA binding versions.

• Chores

  • Removed the CUDA 12-specific benchmark suite in favor of the shared benchmark harness.

CUDA 12 binding builds and tests now use shared infrastructure, and repository checks cover the CUDA 12 tree. Documentation and release records were updated. The package’s local benchmark suite was removed, and the shared benchmark runtime now selects a context-creation signature by CUDA version.

Changes: CUDA 12 Bindings Integration

Layer / File(s)	Summary
Build and package flow cuda_bindings_12/build_hooks.py, cuda_bindings_12/setup.py, cuda_bindings_12/MANIFEST.in, cuda_bindings_12/_build_shared.py, cuda_bindings_12/LICENSE, cuda_bindings_12/cuda/bindings/_internal/utils.pxd, cuda_bindings_12/cuda/bindings/_lib/windll.pxd, cuda_core/MANIFEST.in, cuda_bindings_13/MANIFEST.in	CUDA 12 builds validate toolkit headers and use shared toolchain, Cython-cache, and debug-build logic. Package manifests and selected declarations were updated.
Test setup and validation cuda_bindings_12/tests/*, cuda_bindings_12/examples/pytest.ini, cuda_bindings_12/tests/cython/*, cuda_bindings_13/tests/conftest.py, cuda_core/tests/conftest.py, cuda_python_test_helpers/cuda_python_test_helpers/_pytest_plugin.py	CUDA 12 tests register the shared pytest plugin and add a Cython build driver and build-hook tests. CUDA and library tests add or update markers, fixtures, and assertions. Pytest bootstrap comments are synchronized across package trees.
Documentation and release records cuda_bindings_12/DESCRIPTION.rst, cuda_bindings_12/README.md, cuda_bindings_12/docs/*, cuda_bindings_12/docs/source/*, cuda_bindings_12/examples/*	Documentation links and version selection use the CUDA 12.9.7 snapshot and shared version data. Installation and support ranges, release notes, example copyright notices, and documentation formatting were updated.
Repository-wide tooling .github/labeler.yml, .github/workflows/bandit.yml, .gitignore, .pre-commit-config.yaml, cuda_bindings_12/pyproject.toml, ruff.toml, toolshed/*	Repository labels, lint and hook configuration, generated-file checks, and SPDX validation now include CUDA 12 binding paths.
Benchmark suite and runtime compatibility cuda_bindings_12/benchmarks/*, cuda_bindings_12/README.md, benchmarks/cuda_bindings/runner/runtime.py	The CUDA 12 package’s local benchmark files and configuration were removed. Its README points to the shared harness, whose context setup now selects a signature based on cuda.CUDA_VERSION.

Possibly related issues

• NVIDIA/cuda-python#1967: The labeler rule now includes cuda_bindings_12/**, which changes the paths that can trigger the label behavior described in the issue. It does not change restricted-path label handling.
• NVIDIA/cuda-python#2621: The Apache License appendix changes the CUDA 12 package licensing surface. It does not address the issue’s separate questions about wheel licensing and build-time toolkit dependencies.

Suggested reviewers: mdboom, danielfrg

Priority: ⬇️ Low

Change: Feature

Merge Risk: ⚪ Minimal · up to 2f01e

No issue introduced by this change was established that needs to be fixed before merge; normal checks remain appropriate.

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

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

✨ Finishing Touches: 💡 1

🛠️ Fix failing CI checks: 💡

• Commit to this branch
• Create a new PR

[github-actions]

Doc Preview CI
🚀 View preview at https://nvidia.github.io/cuda-python/pr-preview/pr-3173/
https://nvidia.github.io/cuda-python/pr-preview/pr-3173/cuda-core/
https://nvidia.github.io/cuda-python/pr-preview/pr-3173/cuda-bindings/
https://nvidia.github.io/cuda-python/pr-preview/pr-3173/cuda-pathfinder/
Preview will be ready when the GitHub Pages deployment is complete.

[rwgk]

@mdboom I think I addressed all your feedback, and this PR is now tested with the full CI. Could you please take another look?

I cannot set it to auto-merge (apparently that doesn't work for stacks). If it looks good, could you please go ahead and squash-merge this PR as usual?