fix health check for pd mode

Author: wangzaijun

lightllm/server/api_server.py

@@ -113,7 +113,7 @@ async def healthcheck(request: Request):
 
     from lightllm.utils.health_check import health_check
 
-    if await health_check(g_objs.httpserver_manager, request):
+    if await health_check(g_objs.args, g_objs.httpserver_manager, request):
         return JSONResponse({"message": "Ok"}, status_code=200)
     else:
         return JSONResponse({"message": "Error"}, status_code=404)

lightllm/utils/health_check.py

@@ -4,18 +4,25 @@
 from lightllm.server.multimodal_params import MultimodalParams
 from lightllm.server.httpserver.manager import HttpServerManager
 from fastapi import Request
+from lightllm.server.req_id_generator import ReqIDGenerator
 from lightllm.utils.log_utils import init_logger
 
 logger = init_logger(__name__)
 
 
-async def health_check(httpserver_manager: HttpServerManager, request: Request):
+_g_health_req_id_gen = ReqIDGenerator()
+_g_health_req_id_gen.generate_id()
+
+
+async def health_check(args, httpserver_manager: HttpServerManager, request: Request):
     try:
         request_dict = {"inputs": "你好！", "parameters": {"do_sample": True, "temperature": 0.8, "max_new_tokens": 2}}
         prompt = request_dict.pop("inputs")
         sample_params_dict = request_dict["parameters"]
         sampling_params = SamplingParams(**sample_params_dict)
         sampling_params.verify()
+        if args.run_mode in ["prefill", "decode"]:
+            sampling_params.group_request_id = -_g_health_req_id_gen.generate_id()  # health monitor 的 id 是负的
         multimodal_params_dict = request_dict.get("multimodal_params", {})
         multimodal_params = MultimodalParams(**multimodal_params_dict)