[Distributed] enable tensor_parallel_output for finetuning

PaddlePaddle · May 7, 2024 · 88b1da4 · 88b1da4
1 parent fdcabf8
commit 88b1da4
Showing 1 changed file with 3 additions and 3 deletions.
diff --git a/llm/finetune_generation.py b/llm/finetune_generation.py
@@ -140,7 +140,7 @@ def main():
         if not training_args.autotuner_benchmark:
             model = AutoModelForCausalLMPipe.from_pretrained(
                 model_args.model_name_or_path,
-                tensor_parallel_output=False,
+                tensor_parallel_output=True,
                 tensor_parallel_degree=training_args.tensor_parallel_degree,
                 tensor_parallel_rank=training_args.tensor_parallel_rank,
                 use_flash_attention=model_args.use_flash_attention,
@@ -152,7 +152,7 @@ def main():
             # NOTE(gongenlei): new add autotuner_benchmark
             model_config = AutoConfig.from_pretrained(
                 model_args.model_name_or_path,
-                tensor_parallel_output=False,
+                tensor_parallel_output=True,
                 tensor_parallel_degree=training_args.tensor_parallel_degree,
                 tensor_parallel_rank=training_args.tensor_parallel_rank,
                 dtype=dtype,
@@ -163,7 +163,7 @@ def main():
     else:
         model_config = AutoConfig.from_pretrained(
             model_args.model_name_or_path,
-            tensor_parallel_output=False,
+            tensor_parallel_output=True,
             tensor_parallel_degree=training_args.tensor_parallel_degree,
             tensor_parallel_rank=training_args.tensor_parallel_rank,
             dtype=dtype,