How to efficiently copy PyTorch Tensor in CUDA device code?

Question

Context: I\'m implementing the backward pass for a PyTorch extension so I can inherit from torch.autograd.Function in Python and call the code