add env var

Author: ckl117

llm/predict/predictor.py

@@ -1469,6 +1469,7 @@ def predict_dy_insert(
         # )
 
         if self.config.output_via_mq:
+            os.environ["INFERENCE_TRUNCATED_RETURN_EOS"] = "1"
             result_queue = mp.Queue()
             task_queue = mp.Queue()
             done_event = mp.Event()
@@ -1488,6 +1489,8 @@ def predict_dy_insert(
             if flag_current_rank_run:
                 read_res_process.start()
                 done_event.wait()
+        else:
+            os.environ["INFERENCE_TRUNCATED_RETURN_EOS"] = "0"
 
         done_task_id_set = set()