Compute how many gradient accumulation steps you need to reach a target effective batch size when a single device batch is limited by memory.
累积步数越多显存占用越低。