Why might it be failing in the python code? (Presumably that's using WinDll from ctypes). Also, I have tried to import the cuda version of pytorch_cluster into the environment and I can see the code ...

Would be great to add current versions of torch (2.5, 2.6, 2.6) and maybe newer CUDA versions (12.6 (default for 2.7), 12.8) - maybe something like: ...