fix the dp bug

Author: zty-king

paddlenlp/trainer/trainer.py

@@ -2816,7 +2816,12 @@ def _save_checkpoint(self, model, metrics=None):
                             or "remove_master_weight" not in self.args.unified_checkpoint_config
                         ):
                             paddle.save(global_rank, os.path.join(signal_dir, f".master_weight.done.{global_rank}"))
-            if self.args.should_save or self.args.use_expert_parallel:
+
+            if (
+                self.args.should_save
+                or self.args.use_expert_parallel
+                or (self.args.data_parallel_degree > 1 and not self.args.use_hybrid_parallel)
+            ):
                 if not self.args.use_hybrid_parallel:
                     logger.info("Saving optimizer files.")
                     if self.args.unified_checkpoint:

paddlenlp/trainer/training_args.py

@@ -1807,6 +1807,9 @@ def is_segment_parallel_supported():
                         # DP use hybrid group
                         strategy = fleet.DistributedStrategy()
                         fleet.init(is_collective=True, strategy=strategy)
+                    elif self.using_flex_checkpoint:
+                        strategy = fleet.DistributedStrategy()
+                        fleet.init(is_collective=True, strategy=strategy)
                     else:
                         paddle.distributed.init_parallel_env()