Add setproctitle (#1024)

zhhangBian · web-flow · commit f9a3fe28ebad · 2025-09-02T14:46:37.000+08:00
diff --git a/lightllm/server/api_http.py b/lightllm/server/api_http.py
@@ -25,6 +25,7 @@
 import os
 from io import BytesIO
 import pickle
+import setproctitle
 
 asyncio.set_event_loop_policy(uvloop.EventLoopPolicy())
 import ujson as json
@@ -83,6 +84,8 @@ def set_args(self, args):
             self.g_generate_func = lightllm_generate
             self.g_generate_stream_func = lightllm_generate_stream
 
+        setproctitle.setproctitle(f"lightllm::{get_unique_server_name()}::api_server")
+
         if args.run_mode == "pd_master":
             self.metric_client = MetricClient(args.metric_port)
             self.httpserver_manager = HttpServerManagerForPDMaster(
diff --git a/lightllm/server/api_start.py b/lightllm/server/api_start.py
@@ -94,7 +94,7 @@ def normal_or_p_d_start(args):
 
     if args.graph_max_len_in_batch == 0:
         args.graph_max_len_in_batch = args.max_req_total_len
-    
+
     # mode setting check.
     if args.output_constraint_mode != "none":
         assert args.disable_dynamic_prompt_cache is False
diff --git a/lightllm/server/audioserver/manager.py b/lightllm/server/audioserver/manager.py
@@ -5,6 +5,7 @@
 import uvloop
 import rpyc
 import inspect
+import setproctitle
 from typing import List
 
 asyncio.set_event_loop_policy(uvloop.EventLoopPolicy())
@@ -14,6 +15,7 @@
 from lightllm.server.multimodal_params import AudioItem
 from .model_infer.model_rpc import start_model_process, AudioModelRpcClient
 from lightllm.utils.graceful_utils import graceful_registry
+from lightllm.utils.envs_utils import get_unique_server_name
 from rpyc.utils.classic import obtain
 
 logger = init_logger(__name__)
@@ -140,6 +142,7 @@ def clean_up(self):
 def start_audio_process(args, router_port, audio_port, cache_port, pipe_writer):
     # 注册graceful 退出的处理
     graceful_registry(inspect.currentframe().f_code.co_name)
+    setproctitle.setproctitle(f"lightllm::{get_unique_server_name()}::audio_server")
 
     try:
         audioserver = AudioManager(args, router_port, audio_port, cache_port)
diff --git a/lightllm/server/config_server/api_http.py b/lightllm/server/config_server/api_http.py
@@ -2,6 +2,7 @@
 import asyncio
 import base64
 import pickle
+import setproctitle
 import multiprocessing as mp
 from fastapi import FastAPI, WebSocket, WebSocketDisconnect, Request, Query
 from threading import Lock
@@ -10,7 +11,7 @@
 from lightllm.utils.log_utils import init_logger
 from ..pd_io_struct import PD_Master_Obj
 from .nccl_tcp_store import start_tcp_store_server
-from lightllm.utils.envs_utils import get_env_start_args
+from lightllm.utils.envs_utils import get_env_start_args, get_unique_server_name
 from lightllm.utils.process_check import start_parent_check_thread
 
 
@@ -202,5 +203,6 @@ async def http_start_tcp_store_server(
         return {"status": "ok"}
 
 
+setproctitle.setproctitle(f"lightllm::{get_unique_server_name()}::config_server")
 logger.info("config server start_parent_check_thread...")
 start_parent_check_thread()
diff --git a/lightllm/server/config_server/nccl_tcp_store.py b/lightllm/server/config_server/nccl_tcp_store.py
@@ -1,9 +1,11 @@
 import psutil
 import time
+import setproctitle
 import torch.distributed as dist
 import torch.multiprocessing as mp
 from lightllm.utils.log_utils import init_logger
 from lightllm.utils.process_check import start_parent_check_thread
+from lightllm.utils.envs_utils import get_unique_server_name
 
 logger = init_logger(__name__)
 
@@ -26,6 +28,7 @@ def _start_tcp_store_server(nccl_store_host, nccl_store_port):
     start a TCPStore server.
     """
     start_parent_check_thread()
+    setproctitle.setproctitle(f"lightllm::{get_unique_server_name()}::nccl_tcp_store")
 
     try:
         from torch._C._distributed_c10d import _DEFAULT_PG_NCCL_TIMEOUT
diff --git a/lightllm/server/detokenization/manager.py b/lightllm/server/detokenization/manager.py
@@ -1,5 +1,6 @@
 import uvloop
 import asyncio
+import setproctitle
 
 asyncio.set_event_loop_policy(uvloop.EventLoopPolicy())
 import zmq
@@ -15,6 +16,7 @@
 import pickle
 import time
 from lightllm.utils.log_utils import init_logger
+from lightllm.utils.envs_utils import get_unique_server_name
 
 logger = init_logger(__name__)
 
@@ -173,6 +175,7 @@ def remove_finished_reqs(self):
 def start_detokenization_process(args, detokenization_port, detokenization_pub_port, pipe_writer):
     # 注册graceful 退出的处理
     graceful_registry(inspect.currentframe().f_code.co_name)
+    setproctitle.setproctitle(f"lightllm::{get_unique_server_name()}::detokenization_server")
 
     try:
         manager = DeTokenizationManager(
diff --git a/lightllm/server/embed_cache/manager.py b/lightllm/server/embed_cache/manager.py
@@ -1,10 +1,12 @@
 import rpyc
 import uuid
 import inspect
+import setproctitle
 from typing import Union, Optional
 from lightllm.utils.graceful_utils import graceful_registry
 from lightllm.server.embed_cache.impl.naive_memory_cache import InMemoryCache
 from rpyc.utils.classic import obtain
+from lightllm.utils.envs_utils import get_unique_server_name
 
 
 class CacheServer(rpyc.Service):
@@ -53,6 +55,7 @@ def start_cache_manager(port: int, args, pipe_writer):
     # 注册graceful 退出的处理
     graceful_registry(inspect.currentframe().f_code.co_name)
 
+    setproctitle.setproctitle(f"lightllm::{get_unique_server_name()}::cache_manager")
     manager = InMemoryCache(args)
     service = CacheServer(manager)
     from rpyc.utils.server import ThreadedServer
diff --git a/lightllm/server/health_monitor/manager.py b/lightllm/server/health_monitor/manager.py
@@ -1,6 +1,7 @@
 import sys
 import uvloop
 import asyncio
+import setproctitle
 
 asyncio.set_event_loop_policy(uvloop.EventLoopPolicy())
 import os
@@ -10,6 +11,7 @@
 
 from lightllm.utils.log_utils import init_logger
 from lightllm.utils.graceful_utils import graceful_registry
+from lightllm.utils.envs_utils import get_unique_server_name
 
 logger = init_logger(__name__)
 
@@ -88,6 +90,7 @@ def get_all_cared_pids():
 def start_health_check_process(args, pipe_writer):
     # 注册graceful 退出的处理
     graceful_registry(inspect.currentframe().f_code.co_name)
+    setproctitle.setproctitle(f"lightllm::{get_unique_server_name()}::health_monitor")
     pipe_writer.send("init ok")
 
     all_process_ids = get_all_cared_pids()
diff --git a/lightllm/server/metrics/manager.py b/lightllm/server/metrics/manager.py
@@ -5,11 +5,13 @@
 import inspect
 import functools
 import queue
+import setproctitle
 from .metrics import Monitor
 from prometheus_client import generate_latest
 from rpyc import SocketStream
 from lightllm.utils.log_utils import init_logger
 from lightllm.utils.graceful_utils import graceful_registry
+from lightllm.utils.envs_utils import get_unique_server_name
 
 logger = init_logger(__name__)
 
@@ -136,6 +138,7 @@ def run(self):
 def start_metric_manager(port: int, args, pipe_writer):
     # 注册graceful 退出的处理
     graceful_registry(inspect.currentframe().f_code.co_name)
+    setproctitle.setproctitle(f"lightllm::{get_unique_server_name()}::metric_manager")
 
     service = MetricServer(args)
     if args.metric_gateway is not None:
diff --git a/lightllm/server/router/manager.py b/lightllm/server/router/manager.py
@@ -4,6 +4,7 @@
 import torch
 import pickle
 import inspect
+import setproctitle
 
 asyncio.set_event_loop_policy(uvloop.EventLoopPolicy())
 import zmq
@@ -507,6 +508,7 @@ def clean_up(self):
 def start_router_process(args, router_port, detokenization_port, metric_port, pipe_writer):
     # 注册 graceful 退出的处理
     graceful_registry(inspect.currentframe().f_code.co_name)
+    setproctitle.setproctitle(f"lightllm::{get_unique_server_name()}::router_server")
     start_parent_check_thread()
 
     def handle_exception(loop, context):
diff --git a/lightllm/server/router/model_infer/mode_backend/continues_batch/pd_mode/decode_node_impl/decode_kv_move_manager.py b/lightllm/server/router/model_infer/mode_backend/continues_batch/pd_mode/decode_node_impl/decode_kv_move_manager.py
@@ -8,6 +8,7 @@
 import psutil
 import threading
 import inspect
+import setproctitle
 from rpyc.utils.classic import obtain
 from dataclasses import dataclass
 from typing import List, Dict, Optional, Tuple, Union
@@ -366,6 +367,7 @@ def _init_env(args, info_queue: mp.Queue, mem_queues: List[mp.Queue], event: mp.
 
     # 注册graceful 退出的处理
     graceful_registry(inspect.currentframe().f_code.co_name)
+    setproctitle.setproctitle(f"lightllm::{get_unique_server_name()}::decode_kv_move_manager")
 
     manager = DecodeKVMoveManager(args, info_queue, mem_queues)
     t = ThreadedServer(manager, port=args.pd_decode_rpyc_port, protocol_config={"allow_pickle": True})
diff --git a/lightllm/server/router/model_infer/mode_backend/continues_batch/pd_mode/decode_node_impl/decode_trans_process.py b/lightllm/server/router/model_infer/mode_backend/continues_batch/pd_mode/decode_node_impl/decode_trans_process.py
@@ -3,6 +3,7 @@
 import sys
 import inspect
 import threading
+import setproctitle
 import torch.multiprocessing as mp
 from torch.distributed import TCPStore
 from datetime import timedelta
@@ -13,6 +14,7 @@
 from lightllm.utils.device_utils import kv_trans_use_p2p
 from lightllm.utils.graceful_utils import graceful_registry
 from lightllm.distributed.pynccl import PyNcclCommunicator, StatelessP2PProcessGroup
+from lightllm.utils.envs_utils import get_unique_server_name
 
 logger = init_logger(__name__)
 
@@ -100,6 +102,10 @@ def _init_env(args, device_id: int, task_in_queue: mp.Queue, task_out_queue: mp.
 
     dp_size_in_node = max(1, args.dp // args.nnodes)
 
+    setproctitle.setproctitle(
+        f"lightllm::{get_unique_server_name()}::decode_trans:Device{device_id}_DpSizeInNode{dp_size_in_node}"
+    )
+
     try:
         torch.cuda.set_device(device_id)
         graceful_registry(inspect.currentframe().f_code.co_name)
diff --git a/lightllm/server/router/model_infer/mode_backend/continues_batch/pd_mode/decode_node_impl/up_status.py b/lightllm/server/router/model_infer/mode_backend/continues_batch/pd_mode/decode_node_impl/up_status.py
@@ -4,6 +4,7 @@
 import threading
 import websockets
 import inspect
+import setproctitle
 
 from typing import Dict
 from dataclasses import asdict
@@ -12,6 +13,7 @@
 from lightllm.utils.graceful_utils import graceful_registry
 from lightllm.server.pd_io_struct import PD_Master_Obj
 import torch.multiprocessing as mp
+from lightllm.utils.envs_utils import get_unique_server_name
 
 logger = init_logger(__name__)
 
@@ -108,6 +110,7 @@ async def up_kv_status_task(self, pd_master_obj: PD_Master_Obj):
 
 def _init_env(args, task_in_queue: mp.Queue, task_out_queue: mp.Queue):
     graceful_registry(inspect.currentframe().f_code.co_name)
+    setproctitle.setproctitle(f"lightllm::{get_unique_server_name()}::up_kv_status")
     up_kv_manager = UpStatusManager(args, task_in_queue, task_out_queue)
     logger.info(f"up kv manager {str(up_kv_manager)} start ok")
     while True:
diff --git a/lightllm/server/router/model_infer/mode_backend/continues_batch/pd_mode/prefill_node_impl/prefill_kv_move_manager.py b/lightllm/server/router/model_infer/mode_backend/continues_batch/pd_mode/prefill_node_impl/prefill_kv_move_manager.py
@@ -11,6 +11,7 @@
 import threading
 import inspect
 import collections
+import setproctitle
 from typing import List, Dict, Union
 from lightllm.utils.log_utils import init_logger
 from .prefill_infer_rpyc import PDPrefillInferRpcServer
@@ -227,6 +228,7 @@ def _init_env(args, info_queue: mp.Queue, mem_queues: List[mp.Queue], event: mp.
 
     # 注册graceful 退出的处理
     graceful_registry(inspect.currentframe().f_code.co_name)
+    setproctitle.setproctitle(f"lightllm::{get_unique_server_name()}::prefill_kv_move_manager")
 
     manager = PrefillKVMoveManager(args, info_queue, mem_queues)
     kv_trans_process_check = threading.Thread(target=manager.check_trans_process_loop, daemon=True)
diff --git a/lightllm/server/router/model_infer/mode_backend/continues_batch/pd_mode/prefill_node_impl/prefill_trans_process.py b/lightllm/server/router/model_infer/mode_backend/continues_batch/pd_mode/prefill_node_impl/prefill_trans_process.py
@@ -3,6 +3,7 @@
 import sys
 import inspect
 import threading
+import setproctitle
 import torch.multiprocessing as mp
 from torch.distributed import TCPStore
 from datetime import timedelta
@@ -13,6 +14,7 @@
 from lightllm.utils.device_utils import kv_trans_use_p2p
 from lightllm.utils.graceful_utils import graceful_registry
 from lightllm.distributed.pynccl import StatelessP2PProcessGroup, PyNcclCommunicator
+from lightllm.utils.envs_utils import get_unique_server_name
 
 
 logger = init_logger(__name__)
@@ -102,13 +104,17 @@ def _init_env(
     os.environ["NCCL_SOCKET_NTHREADS"] = "1"
     torch.backends.cudnn.enabled = False
 
+    dp_size_in_node = max(1, args.dp // args.nnodes)
+    setproctitle.setproctitle(
+        f"lightllm::{get_unique_server_name()}::prefill_trans:Device{device_id}_DpSizeInNode{dp_size_in_node}"
+    )
+
     try:
         torch.cuda.set_device(device_id)
         graceful_registry(inspect.currentframe().f_code.co_name)
         master_store = TCPStore(
             host_name=store_ip, port=store_port, is_master=True, use_libuv=True, timeout=timedelta(seconds=30)
         )
-        dp_size_in_node = max(1, args.dp // args.nnodes)
         task_out_queue.put("proc_start")
         mem_managers: List[MemoryManager] = [mem_queue.get(timeout=60) for mem_queue in mem_queues]
         task_out_queue.put("get_mem_managers_ok")
diff --git a/lightllm/server/router/model_infer/model_rpc.py b/lightllm/server/router/model_infer/model_rpc.py
@@ -4,6 +4,7 @@
 import multiprocessing
 import threading
 import inspect
+import setproctitle
 from datetime import timedelta
 from typing import Dict, List, Tuple
 from lightllm.server.router.model_infer.mode_backend import (
@@ -27,6 +28,7 @@
 from lightllm.utils.log_utils import init_logger
 from lightllm.utils.graceful_utils import graceful_registry
 from lightllm.utils.process_check import start_parent_check_thread
+from lightllm.utils.envs_utils import get_unique_server_name
 
 logger = init_logger(__name__)
 
@@ -207,6 +209,7 @@ def _init_env(
 
     # 注册graceful 退出的处理
     graceful_registry(inspect.currentframe().f_code.co_name)
+    setproctitle.setproctitle(f"lightllm::{get_unique_server_name()}::model_infer:RANK{rank}")
     start_parent_check_thread()
 
     # 将调度锁注册到全局的共享变量中
diff --git a/lightllm/server/visualserver/manager.py b/lightllm/server/visualserver/manager.py
@@ -5,6 +5,7 @@
 import rpyc
 import pickle
 import inspect
+import setproctitle
 from typing import List
 from lightllm.server.core.objs.io_objs.group_req import GroupReqIndexes
 from lightllm.server.core.objs import ShmReqManager
@@ -15,6 +16,7 @@
 from lightllm.utils.log_utils import init_logger
 from lightllm.utils.graceful_utils import graceful_registry
 from lightllm.utils.process_check import start_parent_check_thread
+from lightllm.utils.envs_utils import get_unique_server_name
 from rpyc.utils.classic import obtain
 
 
@@ -178,6 +180,7 @@ def clean_up(self):
 def start_visual_process(args, next_module_port, visual_port, cache_port, model_rpc_ports, pipe_writer):
     # 注册graceful 退出的处理
     graceful_registry(inspect.currentframe().f_code.co_name)
+    setproctitle.setproctitle(f"lightllm::{get_unique_server_name()}::visual_server")
     start_parent_check_thread()
     try:
         visualserver = VisualManager(args, next_module_port, visual_port, cache_port, model_rpc_ports)
diff --git a/requirements.txt b/requirements.txt
@@ -85,4 +85,5 @@ sgl-kernel==0.2.6
 httpx==0.28.1
 librosa==0.11.0
 cuda_bindings==12.9.0
-orjson==3.11.2
+orjson==3.11.2
+setproctitle==1.3.6
