infrastructure-torch-version-check-certain-nightly-pytorch-builds-torch-2-2-0-dev2023-91c3c005

THEN Add a runtime check before distributed initialization: use `ctypes` to call `cuDeviceGetCount`. If the error code is 0 (CUDA_SUCCESS) instead of 3 (CUDA_ERROR_NOT_INITIALIZED), the torch version is buggy and should be upgraded or avoided.

import ctypes
import torch
try:
    libcuda = ctypes.CDLL('libcuda.so.1')
    x = ctypes.c_int(-1)
    ans = libcuda.cuDeviceGetCount(ctypes.byref(x))
    if ans == 0:
        print('Warning: Buggy torch version detected; upgrade recommended.')
except OSError:
    pass