infrastructure-cuda-version-check-using-a-pytorch-nightly-build-torch-2-2-0-dev20231-8de49738

THEN Detect the buggy PyTorch version by checking if 'import torch' initializes CUDA without any explicit call. If the script below returns error code 0, the PyTorch version is buggy and should be upgraded to a stable release or a later nightly where the issue is fixed.

import torch
import ctypes
x = ctypes.c_int(-1)
ans = ctypes.CDLL('libcuda.so.1').cuDeviceGetCount(ctypes.byref(x))
if ans == 0:
    print('Buggy PyTorch version detected; consider upgrading.')
else:
    print('PyTorch version is safe.')