[bug] --skip_until_initial_step 使用时，--gradient_accumulation_steps > 1 时会过度地跳过重启位置，导致学习结束

Events

--resume restarts the learning state, and when --skip_until_initial_step is specified and --gradient_accumulation_steps is greater than 2, a significantly larger number of epochs are skipped than the number of steps that have been resumed. As a result, the internal number of steps decreases to a negative value, and the learning stops immediately without starting. The state is successfully loaded, but the number of epochs skipped after --skip_until_initial_step is executed is excessive. The training loop is not entered and terminates.