fix

Author: hiworldwzj

lightllm/server/router/model_infer/mode_backend/pd_nixl/decode_node_impl/decode_kv_move_manager.py

@@ -1,5 +1,6 @@
 import inspect
 import pickle
+import setproctitle
 import torch.multiprocessing as mp
 import time
 from typing import List, Dict, Optional, Tuple, Union, Callable
@@ -10,6 +11,7 @@
 from ..trans_process_obj import KVTransProcess
 from ..base_kv_move_manager import BaseKVMoveManager
 from lightllm.utils.error_utils import log_exception
+from lightllm.utils.envs_utils import get_unique_server_name
 
 logger = init_logger(__name__)
 
@@ -29,6 +31,7 @@ def _init_env(args, info_queue: mp.Queue, event: mp.Event):
 
     # 注册graceful 退出的处理
     graceful_registry(inspect.currentframe().f_code.co_name)
+    setproctitle.setproctitle(f"lightllm::{get_unique_server_name()}::nixl_decode_kv_move_manager")
 
     from .up_status import start_up_kv_status_process
 

lightllm/server/router/model_infer/mode_backend/pd_nixl/decode_node_impl/decode_trans_process.py

@@ -2,6 +2,7 @@
 import time
 import inspect
 import threading
+import setproctitle
 import torch.multiprocessing as mp
 import collections
 import queue
@@ -22,6 +23,7 @@
 from lightllm.server.core.objs import StartArgs
 from ..nixl_kv_transporter import NixlKVTransporter
 from lightllm.utils.error_utils import log_exception
+from lightllm.utils.envs_utils import get_unique_server_name
 
 logger = init_logger(__name__)
 
@@ -48,6 +50,7 @@ def _init_env(
     up_status_in_queue: Optional[mp.SimpleQueue],
 ):
     torch.backends.cudnn.enabled = False
+    setproctitle.setproctitle(f"lightllm::{get_unique_server_name()}::nixl_decode_trans:Device{device_id}")
 
     try:
         torch.cuda.set_device(device_id)

lightllm/server/router/model_infer/mode_backend/pd_nixl/decode_node_impl/up_status.py

@@ -5,6 +5,7 @@
 import websockets
 import inspect
 import pickle
+import setproctitle
 
 from typing import Dict, Union
 from dataclasses import asdict
@@ -13,6 +14,7 @@
 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.SimpleQueue):
     graceful_registry(inspect.currentframe().f_code.co_name)
+    setproctitle.setproctitle(f"lightllm::{get_unique_server_name()}::nixl_up_kv_status")
     up_kv_manager = UpStatusManager(args, task_in_queue)
     logger.info(f"up kv manager {str(up_kv_manager)} start ok")
     while True:

lightllm/server/router/model_infer/mode_backend/pd_nixl/prefill_node_impl/prefill_kv_move_manager.py

@@ -1,4 +1,5 @@
 import inspect
+import setproctitle
 import torch.multiprocessing as mp
 import time
 from typing import List, Dict, Union, Callable
@@ -9,6 +10,7 @@
 from ..trans_process_obj import KVTransProcess
 from ..base_kv_move_manager import BaseKVMoveManager
 from lightllm.utils.error_utils import log_exception
+from lightllm.utils.envs_utils import get_unique_server_name
 
 logger = init_logger(__name__)
 
@@ -28,6 +30,7 @@ def _init_env(args, info_queue: mp.Queue, event: mp.Event):
 
     # 注册graceful 退出的处理
     graceful_registry(inspect.currentframe().f_code.co_name)
+    setproctitle.setproctitle(f"lightllm::{get_unique_server_name()}::nixl_prefill_kv_move_manager")
 
     from .prefill_trans_process import start_prefill_trans_process
 

lightllm/server/router/model_infer/mode_backend/pd_nixl/prefill_node_impl/prefill_trans_process.py

@@ -2,6 +2,7 @@
 import time
 import inspect
 import threading
+import setproctitle
 import torch.multiprocessing as mp
 import collections
 import queue
@@ -15,6 +16,7 @@
 from lightllm.server.core.objs import StartArgs
 from ..nixl_kv_transporter import NixlKVTransporter
 from lightllm.utils.error_utils import log_exception
+from lightllm.utils.envs_utils import get_unique_server_name
 
 
 logger = init_logger(__name__)
@@ -41,6 +43,7 @@ def _init_env(
     task_out_queue: mp.Queue,
 ):
     torch.backends.cudnn.enabled = False
+    setproctitle.setproctitle(f"lightllm::{get_unique_server_name()}::nixl_prefill_trans:Device{device_id}")
 
     try:
         torch.cuda.set_device(device_id)

lightllm/server/visualserver/model_infer/__init__.py

@@ -4,12 +4,13 @@
 import uuid
 import os
 import multiprocessing
+import setproctitle
 from lightllm.utils.retry_utils import retry
 from rpyc.utils.factory import unix_connect
 from rpyc.utils.classic import obtain
 from rpyc.utils.server import ThreadedServer
 from lightllm.utils.graceful_utils import graceful_registry
-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 .model_rpc_client import VisualModelRpcClient
 from .model_rpc import VisualModelRpcServer
 from ..objs import rpyc_config
@@ -18,6 +19,7 @@
 def _init_env(socket_path: str, success_event):
     # 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_model_infer")
 
     import lightllm.utils.rpyc_fix_utils as _
 

lightllm/server/visualserver/proxy_manager.py

@@ -211,7 +211,7 @@ def start_visual_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()}::visual_server")
+    setproctitle.setproctitle(f"lightllm::{get_unique_server_name()}::visual_proxy_server")
     start_parent_check_thread()
     try:
         visualserver = ProxyVisualManager(args=args)

lightllm/server/visualserver/visual_only_manager.py

@@ -177,7 +177,7 @@ def start_visual_process(args: StartArgs, 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")
+    setproctitle.setproctitle(f"lightllm::{get_unique_server_name()}::visual_only_server")
     start_parent_check_thread()
 
     try: