[Trainer] Fix data skip logic for IterableDataset when resuming from checkpoint

Author: ZouKexin-522

paddleformers/trainer/trainer.py

@@ -1865,10 +1865,10 @@ def _inner_training_loop(
                 assert (
                     paddle.sum(paddle.stack(global_step_list) - global_step_list[0]) == 0
                 ), f"Error, get different global step, please check! step list: {[x.item() for x in global_step_list]}"
-
-            epochs_trained = self.state.global_step // num_update_steps_per_epoch
+            _num_update_steps_per_epoch_for_skip = (self.state.max_steps if len_dataloader is None else num_update_steps_per_epoch)
+            epochs_trained = self.state.global_step // _num_update_steps_per_epoch_for_skip
             if not args.ignore_data_skip:
-                steps_trained_in_current_epoch = self.state.global_step % (num_update_steps_per_epoch)
+                steps_trained_in_current_epoch = self.state.global_step % _num_update_steps_per_epoch_for_skip
                 steps_trained_in_current_epoch *= args.gradient_accumulation_steps
             else:
                 steps_trained_in_current_epoch = 0