avoid contention in model loading

ptuning/README.md

@@ -8,7 +8,7 @@
 ## 软件依赖
 运行微调需要4.27.1版本的`transformers`。除 ChatGLM-6B 的依赖之外，还需要安装以下依赖
 ```
-pip install rouge_chinese nltk jieba datasets
+pip install rouge_chinese nltk jieba datasets filelock
 ```
 ## 使用方法
 

ptuning/main.py

@@ -45,6 +45,8 @@ from trainer_seq2seq import Seq2SeqTrainer
 
 from arguments import ModelArguments, DataTrainingArguments
 
+from filelock import FileLock
+
 logger = logging.getLogger(__name__)
 
 def main():
@@ -122,7 +124,8 @@ def main():
                 new_prefix_state_dict[k[len("transformer.prefix_encoder."):]] = v
         model.transformer.prefix_encoder.load_state_dict(new_prefix_state_dict)
     else:
-        model = AutoModel.from_pretrained(model_args.model_name_or_path, config=config, trust_remote_code=True)
+        with FileLock("model.lock"):
+            model = AutoModel.from_pretrained(model_args.model_name_or_path, config=config, trust_remote_code=True)
 
     if model_args.quantization_bit is not None:
         print(f"Quantized to {model_args.quantization_bit} bit")