[Trainer] Update clear_grad for release/2.8

paddlenlp/trainer/trainer.py

@@ -1062,11 +1062,12 @@ def fused_allreduce_gradients_no_sync(paramlist, hcg):
                     if optimizer_was_run:
                         self.lr_scheduler.step()
 
-                    if enable_release_grads and args.pipeline_parallel_degree > 1:
+                    if args.release_grads or enable_release_grads:
                         self.optimizer.clear_grad(set_to_zero=False)
-                        for _, buffers in model._chunk_2_comm_buffers.items():
-                            for buffer in buffers:
-                                buffer._clear_grad_storage()
+                        if args.pipeline_parallel_degree > 1:
+                            for _, buffers in model._chunk_2_comm_buffers.items():
+                                for buffer in buffers:
+                                    buffer._clear_grad_storage()
                     else:
                         self.optimizer.clear_grad()
 

paddlenlp/trainer/training_args.py

@@ -344,6 +344,8 @@ class TrainingArguments:
             Whether skip profile timer, timer will record time usage of forward/ backward/ step, etc.
         distributed_dataloader (`bool`, *optional*):
             Whether to use distributed dataloader. Default is `False`.
+        release_grads (`bool`, *optional*):
+            Whether to release gradients during training. Default is `False`.
     """
 
     output_dir: str = field(
@@ -791,6 +793,9 @@ class TrainingArguments:
         default=False,
         metadata={"help": "Enable MoE (Mixture of Experts) expert parallel training"},
     )
+    release_grads: Optional[bool] = field(
+        default=False, metadata={"help": "Whether to release gradients during training. Default is `False`."}
+    )
 
     def __post_init__(self):
         env_local_rank = int(os.environ.get("PADDLE_RANK_IN_NODE", -1))