There is an option in torch.load
to map the parameters from one specified device to another.
As pointed out in the link you send, there is a way of forcing all GPU tensors to be in CPU while loading, which I copy here:
torch.load('my_file.pt', map_location=lambda storage, loc: storage)