When PyTorch fails to load the CUDA runtime for any reason, it falls back to CPU, often silently, and becomes more than 20 times slower on CNN inference. Not sure if this system could avoid it. Debugging that remotely on a user’s system was fun.