[fix] Fix client pool lock.Optimization task paused

Author: LittleMouse

api_server.py

@@ -139,8 +139,9 @@ async def format_stream():
                 except asyncio.CancelledError:
                     logger.warning("Client disconnected early, terminating inference...")
                     if backend and isinstance(backend, LlmClientBackend):
-                        for task in backend._active_tasks:
-                            task.cancel()
+                        current_task = asyncio.current_task()
+                        if current_task in backend._active_tasks:
+                            current_task.cancel()
                     raise
                 finally:
                     logger.debug("Stream connection closed")

backend/llm_client_backend.py

@@ -114,7 +114,8 @@ async def _get_client(self, request):
         except asyncio.TimeoutError:
             raise RuntimeError("Server busy, please try again later.")
         finally:
-            self._pool_lock.release()
+            if self._pool_lock.locked():
+                self._pool_lock.release()
 
     async def _release_client(self, client):
         async with self._pool_lock: