Add _get_cuda_version_from_cuda_h() in cuda_core/build_hooks.py

[rwgk]

WIP — Will work on it more asap.

[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

[rwgk]

Local testing with a venv that does not have cuda-bindings:

pip install -v . --group test

...
Building wheels for collected packages: cuda-core
Building wheel for cuda-core (pyproject.toml): started
Running command Building wheel for cuda-core (pyproject.toml)
CUDA paths: ['/usr/local/cuda']
CUDA_VERSION from /usr/local/cuda/include/cuda.h: 13000
...

$ pip list
Package Version
---------- -------
cuda-core 0.3.3a0
Cython 3.1.4
iniconfig 2.1.0
numpy 2.3.3
packaging 25.0
pip 25.2
pluggy 1.6.0
Pygments 2.19.2
pytest 8.4.2
setuptools 80.9.0

[github-actions]

Doc Preview CI
🚀 View preview at https://nvidia.github.io/cuda-python/pr-preview/pr-1085/
https://nvidia.github.io/cuda-python/pr-preview/pr-1085/cuda-core/
https://nvidia.github.io/cuda-python/pr-preview/pr-1085/cuda-bindings/
https://nvidia.github.io/cuda-python/pr-preview/pr-1085/cuda-pathfinder/
Preview will be ready when the GitHub Pages deployment is complete.

[leofang]

Calling nvidia-smi is better because it does exactly the same thing under the hood (opening up libcuda/nvcuda and getting the driver version), but it is cross-platform and so takes much less lines of code, and it is a good habit to keep the code surface in build_hook.py small because it is not easy to write tests for modules outside of the source tree.

[rwgk]

Calling nvidia-smi is better because it does exactly the same thing under the hood (opening up libcuda/nvcuda and getting the driver version), but it is cross-platform and so takes much less lines of code, and it is a good habit to keep the code surface in build_hook.py small because it is not easy to write tests for modules outside of the source tree.

The main reason I was leaning into it is that I think it'd be a useful addition to pathfinder.

It's more lines of code only because subprocess is being reused.

Assume that the code I added lives in pathfinder, you'd just reuse that and have a more secure, more robust, and faster (which doesn't matter here, but is useful in general) solution with even less lines of code.

[leofang]

The main reason I was leaning into it is that I think it'd be a useful addition to pathfinder.

Why would the pathfinder need to expose the driver version check as a public API? Or you only had the driver loading part in mind?

[rwgk]

The main reason I was leaning into it is that I think it'd be a useful addition to pathfinder.

Why would the pathfinder need to expose the driver version check as a public API?

It'd be useful in general, not just here. It's just a few lines of Python code that is very similar to other functionality in pathfinder, and does not introduce any new dependencies. Having the function there, would make what's needed here a no-brainer.

[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.

[leofang]

Note: I changed this to return None if no env var is set, but I think it is a mistake because we still need paths to the headers when creating Extension instances below. I think it is time to evaluate using pathfinder. It is already a build-time dependency anyway (cuda-bindings brought it in).

FWIW it seems #692 suddenly becomes within reach!

[leofang]

Since we must need CUDA headers at build time, this is no longer needed in terms of deciding cuda-bindings version, but it can be a useful check in case there is a local driver installation that cannot support the package being built. We can raise a nice warning in such cases via this check.

[rwgk]

I put this in the freezer for now, but the branch is still around: freezer/cuda_core_build_hooks_get_cuda_version_macro