infrastructure-torch-cuda-initializ-importing-a-buggy-nightly-pytorch-build-initialize-13ce2444

THEN Before using vLLM with Ray, verify that the PyTorch version does not pre-initialize CUDA on import. Use the `cuDeviceGetCount` call from `libcuda.so.1` to check: if the error code is 0, the torch version is buggy and should be replaced with one that returns CUDA_ERROR_NOT_INITIALIZED (error code 3) on import.

import ctypes
x = ctypes.c_int(-1)
ans = ctypes.CDLL('libcuda.so.1').cuDeviceGetCount(ctypes.byref(x))
if ans == 0:
    print('Buggy torch version detected – CUDA context initialized on import.')
    # Recommend installing a fixed version: pip install torch==2.2.0 or later stable