get lmdeploy_actor once

Author: YanhuiDua

xtuner/v1/ray/rollout/worker.py

@@ -101,6 +101,7 @@ def __init__(
         self.eos_token: List[int] = [eos_token] if isinstance(eos_token, int) else eos_token
         self.receive_abort_request = asyncio.Event()
         self.abort_timeout = 5.0
+        self.lmdeploy_actor_store = None
 
     def init_dist_port(self):
         """Initialize distributed communication ports.
@@ -613,6 +614,7 @@ async def _handle_non_stream_response(
                         assert finish_reason == "abort", (
                             f"routed_experts is None, but finish_reason is {finish_reason}, expected abort. response: {response}"
                         )
+                    del routed_experts
                 # NOTE: When set return_token_ids = True, the response must contain valid token_ids/logprobs.
                 # If not, we consider it as an invalid response and retry it.
                 # NOTE: !!! When finish_reason is abort, some queries may not return token_ids or logprobs. !!!