[rwgk]

— codex GPT-6 on behalf of Ralf —

Layer 8 of 8. Previous: #3178. Functionally independent of the CI layers; stacked only for ordering. Stack map at the bottom.

These fixes are also isolated in #3355 (12.9.x) and #3356 (main). They remain in this stack until #3356 lands; rebasing afterward will remove the duplicate runtime changes from this layer.

What this PR does

Ports three helper fixes that CUDA 13 already has to the CUDA 12 runtime, with their tests. The diff is +48/-4 across 8 files, including a one-line CI workaround: the existing Core pool-trim test runs serially for the known driver synchronization issue #3350. All test cases remain enabled.

Fix	CUDA 13 origin	Test here
Accept bytes as well as str compiler options in the nested-resource helper	#1665, #1698	nvJitLink bytes-options test; NVVM parameterized case
Keep borrowed coredump setter buffers alive; free getter-owned storage in __dealloc__	#2399	existing coredump API tests (no instrumented leak test)
Propagate exceptions raised inside a registered handle getter instead of reporting "unknown type"	#2551	regression: an internal KeyError propagates

Where to look

• cuda_bindings_12/cuda/bindings/_internal/utils.pyx, _lib/utils.pxi, _lib/utils.pxd, utils/__init__.py: read each side by side with its CUDA 13 counterpart.

Validation

• The complete stack passed full CI at this PR's pre-rebase head 341d286d (run 40997121153, attempt 3, 105 jobs), including the CUDA 12 suites that select these tests.
• Known gap: the existing coredump tests pass with or without the ownership fix; proving the leak is gone would need an instrumented test.
• Current head ede2016affa47ad3b6484a92e16e9fd872447509. Cumulative local validation at cuda.bindings: backport CUDA 12 runtime maintenance fixes #3182: 361 CI-tool tests and full pre-commit passed. fix(bindings): backport CUDA 12 utility runtime fixes #3355 passed all 58 CI jobs, including the CUDA 12 GPU regressions. Full CI on alternative #3322 passed (105 successful jobs; one expected skip).
• The same pool-trim serialization passed in that stack's Windows 3.14t/A100-MCDM job on driver 596.36: DeviceMR and PinnedMR passed; ManagedMR had its normal unsupported-device skip.

Stack map: review scopes and merge units

Layer	PR	Scope	Merge unit
1	#3172	import cuda_bindings_12/, byte-identical to the 12.9.x branch	merged
2	#3173	integrate the import with the shared build tooling	merged
3	#3174	registry, packaging, version selectors, developer environment, guidance	group: #3174 + #3175 + #3176 land together
4	#3175	build both majors from main; planner; version preview bootstrap	group
5	#3176	test, document, nightly, coverage; first CI-complete state	group
6	#3177	releases driven by the tagged source's configuration	after the group
7	#3178	maintenance policy, manual backport workflow, retire backport_branch	after #3177
8	#3182	three CUDA 12 runtime fixes	independent of the CI layers; stacked only for ordering

Layers 3 to 5 are review boundaries, not working intermediate states: #3174 alone breaks CI (workflows still read the removed cuda.build.version), and #3175 alone fails actionlint (callers pass an input the test workflows only accept in #3176). The plan is to review them separately and land the reviewed group as one merge. #3176 is the full-CI target for the group.

The staged stack #3314-#3322 is the alternative in which every layer is mergeable on its own; both stacks reach the same end state, except that the staged stack 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 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 341d286

[coderabbitai]

No actionable comments were generated in the recent review.

Recent review info

Run configuration

• Source mode: Auto
• Repository: NVIDIA/cuda-python
• Branch: rwgk/cb12-on-main-08-runtime-maintenance
• File: .coderabbit.yaml
• Commit: 0b5dbe3
• Layers: Repository YAML
• Review profile: CHILL
• Plan: Enterprise
• Run ID: d6901a71-cb7a-4bd7-92c9-f65b9736ed2b

Commits

Reviewing files that changed from the base of the PR and between e2a6f30 and 0b5dbe3.

Files selected for processing: 7

• cuda_bindings_12/cuda/bindings/_internal/utils.pyx
• cuda_bindings_12/cuda/bindings/_lib/utils.pxd
• cuda_bindings_12/cuda/bindings/_lib/utils.pxi
• cuda_bindings_12/cuda/bindings/utils/__init__.py
• cuda_bindings_12/tests/test_nvjitlink.py
• cuda_bindings_12/tests/test_nvvm.py
• cuda_bindings_12/tests/test_utils.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

• Bug Fixes
  • CUDA options supplied as bytes are now handled correctly, including architecture and NVVM compiler options.
  • Nested character values accept only strings or bytes; strings are encoded as UTF-8, while unsupported types raise a TypeError identifying the received type.
  • Errors raised by a registered CUDA handle getter now propagate instead of being reported as an unknown type.

Nested character resources now accept strings and bytes. The cuCoredumpSettings helper retains borrowed setter buffers. get_cuda_native_handle now allows KeyError from a registered getter to propagate.

Changes: CUDA bindings runtime fixes

Layer / File(s)	Summary
Nested character values cuda_bindings_12/cuda/bindings/_internal/utils.pyx, cuda_bindings_12/tests/test_nvjitlink.py, cuda_bindings_12/tests/test_nvvm.py	Nested char elements accept exact str or bytes values. Strings are UTF-8 encoded, while bytes are passed unchanged. Tests cover byte-string options for nvJitLink and NVVM.
cuCoredumpSettings buffer lifetime cuda_bindings_12/cuda/bindings/_lib/utils.pxd, cuda_bindings_12/cuda/bindings/_lib/utils.pxi	The helper retains init_value for setter-side file and pipe settings. Its destructor frees _charstar only for getters.
Native handle getter errors cuda_bindings_12/cuda/bindings/utils/__init__.py, cuda_bindings_12/tests/test_utils.py	A missing registered getter still raises TypeError. A KeyError raised by a registered getter now propagates. The test checks the propagated error message.

Possibly related PRs

• NVIDIA/cuda-python#3173: Integrates the CUDA 12 source tree with shared build and test tooling; its scope defers runtime backports to this PR.

Suggested reviewers: mdboom, juenglin

Change: Bug fix

Merge Risk: Minimal · up to 0b5db

No merge-blocking issue was identified in the runtime maintenance changes.

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