[LLM] relocate tensor_parallel_output to avoid conflict

Author: SylarTiaNII

llm/finetune_generation.py

@@ -16,6 +16,7 @@
 import sys
 from dataclasses import dataclass, field
 from functools import partial
+from typing import Optional
 
 import paddle
 from argument import (
@@ -66,6 +67,10 @@ class FinetuneArguments(TrainingArguments):
         default=0,
         metadata={"help": "The steps use to control the learing rate."},
     )
+    tensor_parallel_output: Optional[bool] = field(
+        default=False,
+        metadata={"help": "whether to output logits in distributed status"},
+    )
 
 
 def read_local_dataset(path):

llm/utils.py

@@ -212,7 +212,7 @@ def prediction_step(
             if isinstance(logits, (list, tuple)):
                 logits = logits[0]
             # all gather logits when enabling tensor_parallel_output
-            if self.args.tensor_parallel_degree > 1 and self.args.tensor_parallel_output:
+            if self.args.tensor_parallel_degree > 1 and getattr(self.args, "tensor_parallel_output", False):
                 hcg = fleet.get_hybrid_communicate_group()
                 model_parallel_group = hcg.get_model_parallel_group()
                 gathered_logits = []

paddlenlp/trainer/training_args.py

@@ -803,10 +803,6 @@ class TrainingArguments:
         default=False,
         metadata={"help": "whether to run distributed training in auto parallel mode"},
     )
-    tensor_parallel_output: Optional[bool] = field(
-        default=False,
-        metadata={"help": "whether to output logits in distributed status"},
-    )
 
     def __post_init__(self):
         env_local_rank = int(os.environ.get("PADDLE_RANK_IN_NODE", -1))