This code

import torch

def foo():
  print torch
  import torch.cuda
foo()

crashes on the print statement with UnboundLocalError: local variable 'torch' referenced before assignment.

All of the following permutations work without a problem:

import torch

def foo():
  import torch.cuda
  print torch
foo()

import torch
print torch
import torch.cuda

import torch
import torch.cuda
print torch

My use case for this pattern is to only import torch.cuda when command-flags request the usage of a GPU; in Torch7 I've used this pattern because users without CUDA installed were unable to install cutorch or cunn, so they had to be conditionally imported. Is this still a concern in PyTorch?