Training resumed from a previous checkpoint will always lead to a saving operation after one training step, how to avoid this?