update oai

qwen_agent/llm/qwen_oai.py

@@ -416,24 +416,24 @@ def predict(
 class QwenChatAsOAI(BaseChatModel):
 
     def __init__(self, model: str, api_key: str, model_server: str):
-        checkpoint_path = model
+        self.checkpoint_path = copy.copy(model)
         super().__init__()
         tokenizer = AutoTokenizer.from_pretrained(
-            checkpoint_path,
+            self.checkpoint_path,
             trust_remote_code=True,
             resume_download=True,
         )
         device_map = "cpu"
         # device_map = "auto"
         model = AutoModelForCausalLM.from_pretrained(
-            checkpoint_path,
+            self.checkpoint_path,
             device_map=device_map,
             trust_remote_code=True,
             resume_download=True,
         ).eval()
 
         model.generation_config = GenerationConfig.from_pretrained(
-            checkpoint_path,
+            self.checkpoint_path,
             trust_remote_code=True,
             resume_download=True,
         )
@@ -444,7 +444,7 @@ class QwenChatAsOAI(BaseChatModel):
         messages: List[Dict],
         stop: Optional[List[str]] = None,
     ) -> Iterator[str]:
-        _request = ChatCompletionRequest(model=self.model,
+        _request = ChatCompletionRequest(model=self.checkpoint_path,
                                                 messages=messages,
                                                 stop=stop,
                                                 stream=True)
@@ -459,7 +459,7 @@ class QwenChatAsOAI(BaseChatModel):
         messages: List[Dict],
         stop: Optional[List[str]] = None,
     ) -> str:
-        _request = ChatCompletionRequest(model=self.model,
+        _request = ChatCompletionRequest(model=self.checkpoint_path,
                                                 messages=messages,
                                                 stop=stop,
                                                 stream=False)
@@ -471,12 +471,12 @@ class QwenChatAsOAI(BaseChatModel):
                             messages: List[Dict],
                             functions: Optional[List[Dict]] = None) -> Dict:
         if functions:
-            _request = ChatCompletionRequest(model=self.model,
+            _request = ChatCompletionRequest(model=self.checkpoint_path,
                                                     messages=messages,
                                                     functions=functions)
             response = create_chat_completion(_request)
         else:
-            _request = ChatCompletionRequest(model=self.model,
+            _request = ChatCompletionRequest(model=self.checkpoint_path,
                                              messages=messages)
             response = create_chat_completion(_request)
         # TODO: error handling

run_server.py

@@ -12,6 +12,7 @@ from qwen_agent.utils.utils import get_local_ip
 from qwen_server.schema import GlobalConfig
 os.environ["TRANSFORMERS_CACHE"] = ".cache/huggingface/"
 os.environ["HF_HOME"] = ".cache/huggingface/"
+os.environ["MPLCONFIGDIR"] = ".cache/matplotlib/"
 
 def parse_args():
     parser = argparse.ArgumentParser()