Fix dll dir in CUDA 13.x

[Hniii98]

#497

[inducer]

Thanks, this looks reasonable. I'll roll a new release, just in case.

[Hniii98]

Thanks for reviewing so fast! By the way, this is my first merged PR in the open-source world. Hope everything is done right.

[Hniii98]

Root Cause and Explanation

In Python 3.8, the behavior for DLL loading has changed. For more details, refer to: What's New in Python 3.8

Quote from Python 3.8 release notes:
"DLL dependencies for extension modules and DLLs loaded with ctypes on Windows are now resolved more securely. Only the system paths, the directory containing the DLL or PYD file, and directories added with add_dll_directory() are searched for load-time dependencies. Specifically, PATH and the current working directory are no longer used, and modifications to these will no longer have any effect on normal DLL resolution. If your application relies on these mechanisms, you should check for add_dll_directory() and if it exists, use it to add your DLLs directory while loading your library. Note that Windows 7 users will need to ensure that Windows Update KB2533623 has been installed (this is also verified by the installer). (Contributed by Steve Dower in bpo-36085.)"

CUDA and PyCUDA Compatibility

Before CUDA 13.0, all CUDA-related DLLs were located in the $CUDA_PATH/bin directory. This setup worked perfectly with the older DLL loading behavior in PyCUDA.

The code looked like this:

cuda_path = os.environ.get("CUDA_PATH")

if cuda_path is not None:
    os.add_dll_directory(join(cuda_path, "bin"))
    return

However, starting with CUDA 13.x (13.1.0/13.0.2/13.0.1/13.0.0), NVIDIA changed the directory structure, moving the DLL files to $CUDA_PATH/bin/x64.

This change, combined with the new DLL loading behavior introduced in Python 3.8+, means that even if you add $CUDA_PATH/bin/x64 to the $PATH environment variable, it will not solve the issue, as Python no longer uses $PATH to resolve DLLs.

Solution

To resolve the issue, explicitly add $CUDA_PATH/bin/x64 using os.add_dll_directory() if the directory exists.

Future Risks

If NVIDIA changes the DLL directory structure again in future CUDA releases, this issue may recur. PyCUDA might fail to locate the necessary DLLs until the new path is explicitly added using os.add_dll_directory().