fix

Author: ltd0924

fastdeploy/entrypoints/engine_client.py

@@ -92,7 +92,6 @@ def __init__(
             suffix=pid,
             create=False,
         )
-        self.semaphore = StatefulSemaphore((envs.FD_SUPPORT_MAX_CONNECTIONS + workers - 1) // workers)
         self.connection_manager = DealerConnectionManager(
             pid, max_connections=int(os.getenv("FD_DEALER_CONNECTIONS", 50))
         )

fastdeploy/entrypoints/openai/serving_completion.py

@@ -244,7 +244,6 @@ async def completion_full_generator(
             self.engine_client.semaphore.release()
             if dealer is not None:
                 await self.engine_client.connection_manager.cleanup_request(request_id)
-                self.engine_client.semaphore.release()
 
     async def _echo_back_prompt(self, request, res, idx):
         if res["outputs"].get("send_idx", -1) == 0 and request.echo:
@@ -279,7 +278,6 @@ async def completion_stream_generator(
         try:
             await self._ensure_connection_manager()
             dealer, response_queue = await self.engine_client.connection_manager.get_connection(request_id)
-            dealer.write([b"", request_id.encode("utf-8")])
 
             for i in range(num_choices):
                 req_id = f"{request_id}-{i}"
@@ -440,7 +438,6 @@ async def completion_stream_generator(
             yield f"data: {ErrorResponse(message=str(e), code=400).model_dump_json(exclude_unset=True)}\n\n"
         finally:
             del request
-            self.engine_client.semaphore.release()
             if dealer is not None:
                 await self.engine_client.connection_manager.cleanup_request(request_id)
                 self.engine_client.semaphore.release()