diff --git a/fastchat/serve/vllm_worker.py b/fastchat/serve/vllm_worker.py
index 0af680bb5..3614272ef 100644
--- a/fastchat/serve/vllm_worker.py
+++ b/fastchat/serve/vllm_worker.py
@@ -54,6 +54,7 @@ def __init__(
         logger.info(
             f"Loading the model {self.model_names} on worker {worker_id}, worker type: vLLM worker..."
         )
+        self.llm_engine = llm_engine
         self.tokenizer = llm_engine.engine.tokenizer
         # This is to support vllm >= 0.2.7 where TokenizerGroup was introduced
         # and llm_engine.engine.tokenizer was no longer a raw tokenizer
@@ -116,7 +117,7 @@ async def generate_stream(self, params):
             frequency_penalty=frequency_penalty,
             best_of=best_of,
         )
-        results_generator = engine.generate(context, sampling_params, request_id)
+        results_generator = self.llm_engine.generate(context, sampling_params, request_id)
 
         async for request_output in results_generator:
             prompt = request_output.prompt
@@ -135,7 +136,7 @@ async def generate_stream(self, params):
 
             aborted = False
             if request and await request.is_disconnected():
-                await engine.abort(request_id)
+                await self.llm_engine.abort(request_id)
                 request_output.finished = True
                 aborted = True
                 for output in request_output.outputs: