both cudnn and dataparallel pushes do: import torch.cuda On OSX, test_nn.py crashes. Also, dataparallel should smoothly load and work even when you have no CUDA, so this should be fixed.