Whether the .whl file works on your computer is NOT a matter of luck!

[dan64]

You wrote that the fact that the .whl file works on your computer is a matter of luck! This is not true, the problem is that the dlib initialization script requires the user to have the CUDA SDK and cuDNN installed of the same version expected by the downloaded dlib version, while only torch with same CUDA needs to be installed in the same environment.
Replace the init.py file in the dlib folder with the attached version.

Key Changes applied in the attached script:

- Removed Static CUDA Paths: Eliminated hardcoded references to the NVIDIA CUDA Toolkit.
- Dynamic Torch Path Detection: Used torch.__file__ to locate the torch package and constructed the path to its lib folder.
- Windows-Specific Logic: Wrapped the DLL path modification in a check for os.name == 'nt' to ensure it only runs on Windows.
- Robust Error Handling: Maintained try-except blocks to prevent failures if torch is not installed or paths are inaccessible.

Installation Requirement:

Ensure that the torch package with the same CUDA version used by dlib is installed in the same Python environment as dlib, so that the script can locate torch\lib correctly.

This modification allows dlib to dynamically use the CUDA DLLs bundled with torch, improving portability and reducing dependency on system-wide CUDA installations.

init.zip

[eddiehe99]

After uninstalling PyTorch, I installed dlib v20.0.1 and it successfully detected and used CUDA. I guess the trick is to move the NVIDIA files to specific paths and set the environment variables so the system can find CUDA and cuDNN.

[dan64]

In the case it could be useful. I recently I had to compile dlib for CUDA 13.0, but current release is unable to properly manage CUDA 13.0 due to changes in the paths introduced by NVIDIA with the SDK for CUDA 13.0. I had to modify the file FindCUDNN.cmake and CMakeLists.txt
(attached to this post).

Dan

CMakeLists.txt
FindCUDNN.cmake.txt