add auto finish thread when use LLM object

vllm/engine/llm_engine.py

@@ -1318,7 +1318,7 @@ class LLMEngine:
                     seq.append_token_id(sample.output_token, sample.logprobs)
     
     def finish_thread(self):
-        if self.zero_overhead:
+        if self.zero_overhead and self.thread_running:
             self.thread_running = False
             self.sem_m2s.release()
     
@@ -1390,8 +1390,8 @@ class LLMEngine:
     def zero_overhead_step(self) -> List[Union[RequestOutput, PoolingRequestOutput]]:
         if not self.thread_running:
             self.zero_thread.join()
-            self.zero_thread = threading.Thread(target=self.thread_zero_overhead)
             self.thread_running = True
+            self.zero_thread = threading.Thread(target=self.thread_zero_overhead)
             self.zero_thread.start()
         self.sem_m2s.release()
         recode_output = self.q_recorder.get()

vllm/entrypoints/llm.py

@@ -1412,6 +1412,7 @@ class LLM:
         if use_tqdm:
             pbar.close()
 
+        self.llm_engine.finish_thread()
         # Sort the outputs by request ID.
         # This is necessary because some requests may be finished earlier than
         # its previous requests.