Depends on map_location arg in torch.load: might be loaded straight to GPU memory