fix(cuda.core): fall back to driver when nvJitLink < 12.3 is installed

[atiaomar1978-hub]

Summary

Fixes #2408.

When nvJitLink 12.0–12.2 is installed, Linker.which_backend() and Linker construction raised FunctionNotFoundError: function nvJitLinkVersion is not found instead of warning and falling back to the driver (cuLink) backend.

This is a regression introduced in cuda-core 0.7.0. cuda-core 0.6.0 handled this correctly.

Root cause

nvJitLink 12.0–12.2 only export versioned symbols (__nvJitLink*_12_X); the unversioned nvJitLinkVersion export was added in 12.3.

Since cuda-core >= 0.7.0, _decide_nvjitlink_or_driver() probed availability via:

_optional_cuda_import(
    "cuda.bindings.nvjitlink",
    probe_function=lambda module: module.version(),
)

On nvJitLink 12.0–12.2, module.version() calls the missing unversioned symbol and raises FunctionNotFoundError. _optional_cuda_import() only treats DynamicLibNotFoundError as "unavailable", so the exception escapes instead of triggering driver fallback.

cuda-core 0.6.0 used _inspect_function_pointer("__nvJitLinkVersion") (symbol pointer inspection, no function call), which handled this correctly.

Fix

Stop probing via module.version(). Import cuda.bindings.nvjitlink without a probe function and rely on the existing _nvjitlink_has_version_symbol() helper for the >= 12.3 check.

Reproduction (before fix)

pip install "cuda-core==1.1.0" "cuda-bindings==12.9.7" "cuda-toolkit[nvjitlink]==12.0.1"
python -c "from cuda.core import Linker; print(Linker.which_backend())"

Result: FunctionNotFoundError: function nvJitLinkVersion is not found

Independently confirmed on Windows x86_64 (no NVIDIA GPU) with Python 3.12.10 and nvidia-nvjitlink-cu12==12.0.140.

Expected (and restored) behavior:

>>> Linker.which_backend()
RuntimeWarning: ... is too old (<12.3) ...
'driver'

Tests added

• test_decide_nvjitlink_or_driver_falls_back_when_nvjitlink_too_old
• test_decide_nvjitlink_or_driver_selects_nvjitlink_when_version_symbol_present
• test_which_backend_falls_back_when_nvjitlink_too_old (Linker API regression test for [BUG]: Linker raises FunctionNotFoundError instead of falling back to the driver when nvJitLink < 12.3 is installed #2408)
• Updated existing optional-import tests to assert no version() probe on import

Test plan

• Added unit/integration coverage for nvJitLink < 12.3 fallback path
• Ran cuda_pathfinder/tests/test_optional_cuda_import.py locally (5 passed)
• cuda_core tests require NVIDIA driver/GPU locally (nvcuda.dll); CI will run full suite
• Manual repro with cuda-toolkit[nvjitlink]==12.0.1 returns "driver" with RuntimeWarning

CI note

The failing PR has assignee, labels, and milestone check is a metadata gate, not a code/test failure. External contributors cannot add the required bug label, assignee, or milestone. Could a maintainer please add:

• Label: bug
• Milestone: cuda.core 1.2.0 (or appropriate release)
• Assignee: triage owner

Signed-off-by: Omar Atie atiaomar1978-hub@users.noreply.github.com

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[atiaomar1978-hub]

CI run 29979997352

The failing job \PR has assignee, labels, and milestone\ is a metadata gate (missing \�ug\ label, assignee, and milestone). I cannot set these as an external contributor.

Updates in latest push

• Added regression coverage:
  • \ est_which_backend_falls_back_when_nvjitlink_too_old\
  • \ est_decide_nvjitlink_or_driver_falls_back_when_nvjitlink_too_old\
  • \ est_decide_nvjitlink_or_driver_selects_nvjitlink_when_version_symbol_present\
• Local tests run:
  • \cuda_pathfinder/tests/test_optional_cuda_import.py: 5 passed
  • \cuda_core/tests/test_optional_dependency_imports.py: blocked locally by missing
    vcuda.dll\ (no GPU on this machine); CI should cover these

Could a maintainer please add the \�ug\ label, assignee, and milestone so the metadata check passes?

[atiaomar1978-hub]

About me:
I’m Omar Atie, an open‑source contributor with experience building high‑performance integration systems, AI components, and modern data‑processing features in Apache Camel. My work spans distributed systems, API design, and component development, with a focus on reliability, performance, and clean architecture. I’m now expanding into CUDA open‑source development to contribute to GPU‑accelerated computing, improve developer experience, and help strengthen NVIDIA’s ecosystem through practical, high‑impact contributions.

Request:
I’d appreciate feedback from maintainers on how to best align this contribution with NVIDIA’s internal coding standards.
If workflow approvals, GitHub Actions permissions, or metadata updates are required, please guide me and I’ll make the necessary adjustments.
I’m also interested in becoming an active contributor in the CUDA ecosystem, so any guidance on how to integrate effectively with the team and contribution process would be greatly appreciated.

[atiaomar1978-hub]

Added release notes entry

Pushed a follow-up commit adding a cuda.core 1.2.0 release note under Fixes and enhancements documenting the #2408 regression fix (cuda_core/docs/source/release/1.2.0-notes.rst).

All commits on this branch are signed and show as verified on GitHub.

Current branch

Commit	Purpose	Verified
4ee6ae0	Fix: fall back to driver when nvJitLink < 12.3	✅
7f6a534	Tests: regression coverage	✅
91a1f4f	Docs: 1.2.0 release note	✅

Still blocked only by maintainer metadata (assignee, \�ug\ label, milestone) and the external-contributor workflow approval (\copy-pr-bot). Happy to adjust anything on request.