[FEA]: Pathfinder: get include directory for where nvcc is hosted

[capybara-club]

Is this a duplicate?

• I confirmed there appear to be no duplicate issues for this request and that I agree to the Code of Conduct

Area

cuda.pathfinder

Is your feature request related to a problem? Please describe.

My library uses cuda-python to compile custom cute kernels. I have a user that is reporting that my current get_include_dir_cuda() fails for him for some reason. I saw pathfinder is just now being worked on and I was wondering what I could use for now that would be robust to different user setups. My current code is:

def get_include_dir_cuda() -> Path:
    """Best-effort guess of the Toolkit’s <cuda>/include directory."""
    import os, shutil
    if os.getenv("CUDA_HOME"):
        return Path(os.environ["CUDA_HOME"]) / "include"
    # fall back to the directory that owns nvcc (works for most local installs)
    nvcc = shutil.which("nvcc")
    if nvcc:
        return Path(nvcc).parent.parent / "include"
    raise RuntimeError("Cannot find CUDA include directory")

I there anything I can replace this with that should work better? Will I eventually just replace this with a pathfinder function call?

Describe the solution you'd like

Something robust that will handoff finding cuda headers to Nvidia's package. The internals of cute/cutlass require a few of these headers.

Describe alternatives you've considered

No response

Additional context

Thanks!

[rwgk]

I already started working on finding CTK headers.

nvcc = shutil.which("nvcc")

That's something I haven't thought of before. I'll keep that in mind.

I'll link my CTK headers PR here when I open it (very soon I think).

[leofang]

Thank you for your interest in using pathfinder, @capybara-club! It seems you use NVCC only as a proxy to find headers, which is what you actually need, and you don't actually need NVCC?

I don't think NVCC is a robust proxy to search for headers, because it could be possible that NVCC does not exist depending on how CUDA components are installed. I believe Ralf's upcoming solution will be more robust.

[rwgk]

FYI — I just started the first round of CI testing for PR #956. Currently that PR includes the shutil.which("nvcc") fallback code. I'm keeping it in there for the moment to see if it works on all platforms. I'll probably remove the code later, as discussed earlier today in our CUDA Python weekly sync-up.

[capybara-club]

Thank you for your interest in using pathfinder, @capybara-club! It seems you use NVCC only as a proxy to find headers, which is what you actually need, and you don't actually need NVCC?

I don't think NVCC is a robust proxy to search for headers, because it could be possible that NVCC does not exist depending on how CUDA components are installed. I believe Ralf's upcoming solution will be more robust.

Yeah, I don't need nvcc, just the headers for what cute/cutlass's cpp code requires. I wonder if something like that is happening with the issue a user reported to me.

Thanks!