cuda.core: add tests for ObjectCode.from_library

[lijinf2]

Description

closes #663.

Adds tests for ObjectCode.from_library and extends the saxpy test-binary infrastructure from #2193.

• build_test_binaries.sh: package saxpy.o into saxpy.a for Linux / saxpy.lib for Windows via nvcc -lib
• CI (build-wheel.yml): upload library artifacts alongside object fixtures
• test_module.py: new get_saxpy_library fixture and three tests mirroring from_object coverage (bytes, file path, linker integration)

Checklist

• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[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.

[lijinf2]

/ok to test a0f1507

[github-actions]

Doc Preview CI
🚀 View preview at https://nvidia.github.io/cuda-python/pr-preview/pr-2281/
https://nvidia.github.io/cuda-python/pr-preview/pr-2281/cuda-core/
https://nvidia.github.io/cuda-python/pr-preview/pr-2281/cuda-bindings/
https://nvidia.github.io/cuda-python/pr-preview/pr-2281/cuda-pathfinder/
Preview will be ready when the GitHub Pages deployment is complete.

[leofang]

Thanks, Jinfeng! Technically we already tested from_library in the CI!

cuda-python/cuda_core/tests/graph/test_device_launch.py

Line 69 in dad6a42

cudadevrt = ObjectCode.from_library(cudadevrt_path)

But I wouldn't complain for completeness 🙂

Let me get back to review tomorrow...

[leofang]

Let's try to reduce the boilerplate code. I understand that it's a lot cheaper to just generate code using agents. But there is really no good reason to add same code over and over. pytest parametrization can go a long way.

[leofang]

Consolidate this fixture with get_saxpy_object above.

[lijinf2]

Done.

[leofang]

Consolidate the test with test_object_code_load_object_with_linker above.

[lijinf2]

Done

[lijinf2]

Let's try to reduce the boilerplate code. I understand that it's a lot cheaper to just generate code using agents. But there is really no good reason to add same code over and over. pytest parametrization can go a long way.

Thanks for the review!

I didn't realize test_device_launch.py already covers from_library. That path uses a toolkit cudadevrt library (file path) plus runtime Program compile. This PR adds the nvcc path (saxpy.o / saxpy.a) and covers both bytes and file-path inputs.

Agree on cutting boilerplate. I've carefully reviewed the AI-generated code line by line and updated the PR with pytest parametrization. Happy to iterate if anything else stands out.

[lijinf2]

/ok to test 5156984