Does PyTorch verify the consistency of training functionality between previous and current versions through loss alignment?