[rwgk]

— codex GPT-6 on behalf of Ralf —

Layer 4 of 7 of the staged "CUDA 12 bindings on main" stack. Previous: #3317. Next: #3319. Stack map at the bottom.

What this PR does

Cuts the last dependency on the 12.9.x branch inside PR CI. Everything that needs CUDA 12 bindings during a main run now takes the wheel built in the same run.

• cuda.core's cu12 half is built against the same-run CUDA 12 wheel. The lookup-run-id --branch 12.9.x artifact download disappears from build-wheel.yml.
• Core's cu12 Cython test extensions are built and uploaded (<core artifact>-cu12-tests), with stale-output cleanup between the two ABI builds.
• A CUDA 12 metapackage wheel is built (CUDA_PYTHON_BUILD_MAJOR=12) next to the CUDA 13 one.
• The temporary Maintenance bindings jobs from ci: build and test local CUDA 12 bindings #3317 go away. Instead the regular test matrices annotate each row from the registry (BINDINGS_ROOT, BINDINGS_SOURCE, artifact names): 12.9.1 rows test bindings, Core, and metapackage from this run's wheels; 13.0.x / 12.6.x rows keep testing Core against published floor bindings.
• sdist checks run once per package root (matrix over the registry), with the SCM preview bootstrap for the CUDA 12 root.
• cuda_python/setup.py + pyproject.toml: a selector table {"12": "v12.9.*", "13": "v13.4.*"}; CUDA_PYTHON_BUILD_MAJOR (default 13) picks the line; dependencies follow the resolved version, so a CUDA 12 sdist rebuilt outside Git keeps CUDA 12 dependencies. The CUDA 12 metapackage depends on bindings only; CUDA 13 also on Core and Pathfinder. Covered by test_cuda_python_setup.py.

After this merges

• PR CI is self-contained for both CUDA majors; nothing reads the 12.9.x branch any more except nightly for older artifact sources (handled in ci: retire legacy branch-based bindings maintenance #3322).
• Releases are unchanged: CUDA 13 from main as today, CUDA 12 still from 12.9.x tooling until ci: release CUDA 12 bindings from main #3319.
• Steady-state runner work is about the same as ci: build and test local CUDA 12 bindings #3317: the separate maintenance test jobs are replaced by richer 12.9.1 rows in the regular jobs.

Where to look

• .github/workflows/build-wheel.yml: "Constrain maintenance cuda.core to the same-run cuda.bindings wheel" and "Build maintenance cuda.core Cython tests".
• .github/workflows/test-wheel-linux.yml: the jq block that annotates matrix rows (mirrored in the Windows file).
• cuda_python/setup.py: line selection and version-derived dependencies.

All changes (14 files, +331 / -237)

• .github/workflows/build-wheel.yml: same-run CUDA 12 constraint for Core, maintenance metapackage wheel, maintenance Core Cython tests, removal of the prior-branch lookup.
• .github/workflows/ci.yml: removal of the three temporary jobs; sdist jobs matrixed over package_roots; bindings-config passed to all test jobs.
• .github/workflows/test-sdist-linux.yml, test-sdist-windows.yml: bindings-package input replaces cuda-version; per-root build, Pathfinder sdist only for the current root.
• .github/workflows/test-wheel-linux.yml, test-wheel-windows.yml: registry-driven row annotation; bindings-only input removed.
• ci/tools/env-vars: MAINTENANCE_CUDA_MAJOR, registry-driven test routing, Core Cython artifact suffix, metapackage artifact name.
• cuda_python/setup.py, cuda_python/pyproject.toml, ci/tools/tests/test_cuda_python_setup.py (new).
• Docs: cuda_python/AGENTS.md, cuda_python/docs/README.md, cuda_core/docs/source/install.rst.

Validation

• Full CI passed at the pre-rebase head 2f2eb43e (run 41188280375).
• Current head 55b92f8674510105d2184ad371df4b75ba71e3f0 includes the follow-up fixes. Local validation: 306 CI-tool tests, actionlint passed. Full CI and native Pixi source/GPU checks have not been rerun for it.

Stack map: what main can do after each layer merges

Layer	PR	After it merges, main ...
1	#3314	behaves exactly as before; CI has one reader for the two bindings roots; CUDA 13 versions can no longer match a v12.* tag
2	#3315	lets a developer build, test, and document cuda_bindings_12/; the PR-time CPU source-build smoke covers both roots; three CUDA 12 runtime fixes are in
3	#3317	builds the CUDA 12 bindings wheel itself and runs its GPU tests; cuda.core and the metapackage still consume 12.9.x artifacts
4	#3318	uses same-run CUDA 12 wheels everywhere; PR CI no longer reads the 12.9.x branch
5	#3319	can release CUDA 12 from a v12.9.N tag on main; CUDA 13 releases are unchanged
6	#3321	builds and tests only what a change needs, per bindings root, reusing baseline artifacts
7	#3322	carries the schema-2 registry, a manual-only backport workflow, a metadata consistency hook, and the maintenance docs; 12.9.x is frozen

Every layer was designed to be mergeable on its own and had a green full CI run at its pre-rebase head on 2026-10-31. The stack is the "individually mergeable pieces" alternative to #3174-#3182; both stacks reach the same end state except that this one keeps compatibility routes for older artifact sources and retained CUDA 13 tags.

[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.

[rwgk]

/ok to test 2f2eb43

[github-actions]

Doc Preview CI
🚀 View preview at https://nvidia.github.io/cuda-python/pr-preview/pr-3318/
https://nvidia.github.io/cuda-python/pr-preview/pr-3318/cuda-core/
https://nvidia.github.io/cuda-python/pr-preview/pr-3318/cuda-bindings/
https://nvidia.github.io/cuda-python/pr-preview/pr-3318/cuda-pathfinder/
Preview will be ready when the GitHub Pages deployment is complete.