[RL] update reschedule finish reason (#2709)

Author: ltd0924

fastdeploy/entrypoints/openai/protocol.py

@@ -122,7 +122,7 @@ class ChatCompletionResponseChoice(BaseModel):
     """
     index: int
     message: ChatMessage
-    finish_reason: Optional[Literal["stop", "length", "tool_calls"]]
+    finish_reason: Optional[Literal["stop", "length", "tool_calls", "recover_stop"]]
 
 
 class ChatCompletionResponse(BaseModel):

fastdeploy/entrypoints/openai/serving_chat.py

@@ -222,7 +222,7 @@ async def chat_completion_stream_generator(
                         choice.finish_reason = "length"
 
                     if res.get("error_msg") is not None and "Recover" in res["error_msg"]:
-                        choice.finish_reason = "length"
+                        choice.finish_reason = "recover_stop"
 
                 if request.metadata is not None and request.metadata.get("training", False) and delta_text != "":
                     choice.delta.token_ids = output["token_ids"]
@@ -340,7 +340,7 @@ async def chat_completion_full_generator(
             choice.finish_reason = "length"
 
         if final_res.get("error_msg") is not None and "Recover" in final_res["error_msg"]:
-            choice.finish_reason = "length"
+            choice.finish_reason = "recover_stop"
         choices.append(choice)
 
         num_prompt_tokens = len(prompt_token_ids)