PyTorch multiprocessing with CUDA sets tensors to 0

distributed
12

I have same issue. Torch.multiprocessing on CUDA turns tensors to zeros