[rwgk]

Highlights

Apply three existing current-major helper fixes to the CUDA 12 runtime. Each fix has its own commit, keeping runtime changes separate from the packaging/CI migration.

• Accept byte-string compiler options alongside strings in the nested-resource helper.
• Retain borrowed coredump setter buffers and free getter-owned storage.
• Propagate exceptions raised inside registered handle getters instead of misreporting an unknown handle type.

Provenance and Review Scope

These are backports of fixes already implemented for the current bindings major. They are separate from, and not prerequisites for, making #3174–#3176 function together.

Fix	Earlier upstream change	Coverage retained here
Byte-string compiler options	#1665, #1698	nvJitLink bytes-options test and NVVM parameterized case
Coredump buffer ownership	#2399	Existing API tests; no new instrumented memory/lifetime regression test
Registered handle-getter errors	#2551	Regression proving an internal KeyError propagates

An ordinary coredump round trip would also pass with the getter leak; the existing functional tests do not prove its absence. This is a validation gap for the memory-ownership fix.

Stack / Review Boundary

Previous: #3178. The stack overview lists all six open PRs and the intended workflow.

#3174 through #3176 are intended to land together, with separate review scopes and cumulative CI at #3176; the first two are incomplete intermediate states. Release routing (#3177), maintenance policy (#3178), and the runtime fixes (#3182) follow that group. The complete proposed source tree is at #3182.

Baseline: main b4a5e5513a1f.

Counterpart Review

Compared the CUDA 12 _internal/utils.pyx, _lib/utils.pxd/.pxi, utils/__init__.py, and their focused tests with the corresponding CUDA 13 implementations. These are handwritten helper adaptations of the existing current-major fixes; generated API differences are preserved. The process-global coredump test serialization is handled separately in #3176 for both roots.

Validation

At this published head, 341 CI-tool tests pass in a full Git checkout, together with full local pre-commit run --all-files, including actionlint, Lychee, and generated-file seals. Those are tooling/source checks, not native CUDA runtime tests.

Full hosted CI also passed for the complete stack at the unchanged #3182 head 775392e1a6b372b8d046a95466513bd114b05f6a: run 40456627846, attempt 2. The only retry was a self-hosted H100 runner disconnect during driver setup, before package tests began. Its replacement passed, together with the final gate; exact-head auxiliary checks are also green or intentionally skipped. This tests the cumulative stack, rather than hosted release publication or backport automation.

The focused runtime regressions listed above are selected by the CUDA 12 test coverage. Existing coredump API tests do not instrument allocations or prove the absence of a leak; that ownership-validation gap remains explicit. The three runtime changes stay isolated from the preceding CI and documentation revisions.

Checklist

• Existing regression tests move with the corresponding runtime changes.
• The description records provenance, scope, and validation limits.

[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 7856417

[github-actions]

Doc Preview CI
🚀 View preview at https://nvidia.github.io/cuda-python/pr-preview/pr-3182/
https://nvidia.github.io/cuda-python/pr-preview/pr-3182/cuda-core/
https://nvidia.github.io/cuda-python/pr-preview/pr-3182/cuda-bindings/
https://nvidia.github.io/cuda-python/pr-preview/pr-3182/cuda-pathfinder/
Preview will be ready when the GitHub Pages deployment is complete.

[rwgk]

/ok to test 4f6212a

[rwgk]

/ok to test dd7d842

[rwgk]

/ok to test 775392e