Amazon is submitting code upstream to PyTorch to make the framework more hardware-neutral by removing Nvidia-specific assumptions.