fix

Author: none

lightllm/server/core/objs/req.py

@@ -77,10 +77,8 @@ class Req(ctypes.Structure):
         ("prompt_cache_len", ctypes.c_int),  # 用于记录prompt cache 的命中长度，用于统计
         ("is_paused", ctypes.c_bool),  # 标记一个Req因为显存资源管理的原因被临时暂停了。
         ("finish_status", FinishStatus),
+        # 这个标记变量是http_server 写入，其他进程读取，用于标记该请求是否因为断网被aborted。
         ("is_aborted", ctypes.c_bool),
-        # 这个标记变量是router进程读取到is_aborted信息后，router 进程标记该请求已经被abort处理
-        # 等待推理进程处理，防止router进程反复给推理进程发送abort指令。
-        ("router_aborted", ctypes.c_bool),
         # 当FinishStatus 是正常结束状态时，finish_token_index 用于标识结束的
         # token 的index位置
         ("finish_token_index", ctypes.c_int),
@@ -100,7 +98,8 @@ class Req(ctypes.Structure):
         ("mtp_accepted_token_num", ctypes.c_int),
         # mtp_step 保存一个mtp使用的常量参数，用于快速访问，不会被外部输入初始化
         ("_mtp_step", ctypes.c_int),
-        # stop_str_matched用于判断停止字符串是否匹配成功
+        # stop_str_matched 用于判断停止字符串是否匹配成功,  detokenization 进程写入，router 进程读取
+        # 然后router发停止命令给推理进程，推理进程停止输出
         ("stop_str_matched", ctypes.c_bool),
     ]
 
@@ -129,7 +128,6 @@ def init(
         self.is_paused = False
         self.finish_status = FinishStatus()
         self.is_aborted = False
-        self.router_aborted = False
         self.shm_infer_released = False
         self.shm_cur_kv_len = 0
         self.shm_cur_output_len = 0

lightllm/server/router/manager.py

@@ -332,8 +332,8 @@ def _get_aborted_reqs_from_running_batch(self) -> List[Req]:
         if self.running_batch is None:
             return ans
         for req in self.running_batch.reqs:
-            if req.is_aborted and req.router_aborted is False:
-                req.router_aborted = True
+            if req.is_aborted and req._router_aborted is False:
+                req._router_aborted = True
                 ans.append(req)
         return ans
 
@@ -342,7 +342,8 @@ def _get_stop_str_reqs_from_running_batch(self) -> List[Req]:
         if self.running_batch is None:
             return ans
         for req in self.running_batch.reqs:
-            if req.stop_str_matched:
+            if req.stop_str_matched and req._router_stop_str_matched is False:
+                req._router_stop_str_matched = True
                 ans.append(req)
         return ans
 
@@ -382,6 +383,11 @@ def _add_req(self, group_req_indexes: GroupReqIndexes):
             req = self.shm_req_manager.get_req_obj_by_index(req_index)
             req.multimodal_params = group_req_indexes.multimodal_params
             req.start_time = group_req_indexes.time_mark
+            # 附加一个私有标记变量，标记请求是否已经被router发送过abort命令给推理进程，
+            # 防止反复发送abort命令给推理进程
+            req._router_aborted = False
+            # 作用同 _router_aborted 类似
+            req._router_stop_str_matched = False
             req_group.append(req)
 
             logger.info(f"router recive req id {req.request_id} cost time {time.time() - req.start_time} s")

lightllm/server/router/model_infer/infer_batch.py

@@ -392,9 +392,6 @@ def update_finish_status(self, eos_ids, output_len: int):
             self.finish_status.set_status(FinishStatus.FINISHED_LENGTH)
         return
 
-    def is_finished_or_aborted(self):
-        return self.finish_status.is_finished() or self.shm_req.router_aborted
-
     def _stop_sequences_matched(self, output_len: int):
         for stop_token_ids in self.stop_sequences:
             stop_len = len(stop_token_ids)