A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 23:45 UTC
[Misc] Rework process titles (#21780)
Signed-off-by: Nick Hill <nhill@redhat.com>
Nick Hill committed
Jul 29, 2025 at 06:14 UTC
7234fe26858f2c621901494c307c90e65fe35340
7 files changed
+48
-26
vllm/entrypoints/cli/serve.py
+4
-2
@@ -21,7 +21,7 @@ from vllm.entrypoints.utils import (VLLM_SUBCMD_PARSER_EPILOG,
21
from vllm.executor.multiproc_worker_utils import _add_prefix
22
from vllm.logger import init_logger
23
from vllm.usage.usage_lib import UsageContext
24
-from vllm.utils import FlexibleArgumentParser, bind_process_name, get_tcp_uri
24
+from vllm.utils import FlexibleArgumentParser, get_tcp_uri
25
from vllm.v1.engine.core import EngineCoreProc
26
from vllm.v1.engine.utils import CoreEngineProcManager, launch_core_engines
27
from vllm.v1.executor.abstract import Executor
@@ -77,7 +77,7 @@ def run_headless(args: argparse.Namespace):
77
78
if args.api_server_count > 1:
79
raise ValueError("api_server_count can't be set in headless mode")
80
- bind_process_name("APIServer_Headless")
80
+ # set_process_title("Headless_ProcManager")
81
# Create the EngineConfig.
82
engine_args = vllm.AsyncEngineArgs.from_cli_args(args)
83
usage_context = UsageContext.OPENAI_API_SERVER
@@ -140,6 +140,8 @@ def run_multi_api_server(args: argparse.Namespace):
140
num_api_servers = args.api_server_count
141
assert num_api_servers > 0
142
143
+ # set_process_title("ProcManager")
144
+
145
if num_api_servers > 1:
146
setup_multiprocess_prometheus()
147
vllm/entrypoints/openai/api_server.py
+12
-4
@@ -11,6 +11,7 @@ import multiprocessing
11
import os
12
import signal
13
import socket
14
+import sys
15
import tempfile
16
import uuid
17
from argparse import Namespace
@@ -94,15 +95,15 @@ from vllm.entrypoints.openai.serving_transcription import (
95
from vllm.entrypoints.openai.tool_parsers import ToolParserManager
96
from vllm.entrypoints.utils import (cli_env_setup, load_aware_call,
97
log_non_default_args, with_cancellation)
98
+from vllm.executor.multiproc_worker_utils import _add_prefix
99
from vllm.logger import init_logger
100
from vllm.reasoning import ReasoningParserManager
101
from vllm.transformers_utils.config import (
102
maybe_register_config_serialize_by_value)
103
from vllm.transformers_utils.tokenizer import MistralTokenizer
104
from vllm.usage.usage_lib import UsageContext
103
-from vllm.utils import (Device, FlexibleArgumentParser, bind_process_name,
104
- get_open_zmq_ipc_path, is_valid_ipv6_address,
105
- set_ulimit)
105
+from vllm.utils import (Device, FlexibleArgumentParser, get_open_zmq_ipc_path,
106
+ is_valid_ipv6_address, set_process_title, set_ulimit)
107
from vllm.v1.metrics.prometheus import get_prometheus_registry
108
from vllm.version import __version__ as VLLM_VERSION
109
@@ -1805,6 +1806,13 @@ def setup_server(args):
1806
1807
async def run_server(args, **uvicorn_kwargs) -> None:
1808
"""Run a single-worker API server."""
1809
+
1810
+ # Add process-specific prefix to stdout and stderr.
1811
+ process_name = "APIServer"
1812
+ pid = os.getpid()
1813
+ _add_prefix(sys.stdout, process_name, pid)
1814
+ _add_prefix(sys.stderr, process_name, pid)
1815
+
1816
listen_address, sock = setup_server(args)
1817
await run_server_worker(listen_address, sock, args, **uvicorn_kwargs)
1818
@@ -1820,7 +1828,7 @@ async def run_server_worker(listen_address,
1828
ToolParserManager.import_tool_parser(args.tool_parser_plugin)
1829
1830
server_index = client_config.get("client_index", 0) if client_config else 0
1823
- bind_process_name("APIServer", str(server_index))
1831
+ set_process_title("APIServer", str(server_index))
1832
# Load logging config for uvicorn if specified
1833
log_config = load_log_config(args.log_config_file)
1834
if log_config is not None:
vllm/utils/__init__.py
+12
-4
@@ -3282,14 +3282,22 @@ def has_deep_gemm() -> bool:
3282
return _has_module("deep_gemm")
3283
3284
3285
-def bind_process_name(name: str, suffix: str = "") -> None:
3286
- """Bind the process name to a specific name with an optional suffix.
3285
+def set_process_title(name: str,
3286
+ suffix: str = "",
3287
+ append: bool = False) -> None:
3288
+ """
3289
+ Set the current process title to a specific name with an
3290
+ optional suffix.
3291
3292
Args:
3289
- name: The base name to bind the process to.
3293
+ name: The title to assign to the current process.
3294
suffix: An optional suffix to append to the base name.
3295
+ append: Whether to append to the existing process title.
3296
"""
3292
- name = f"{envs.VLLM_PROCESS_NAME_PREFIX}::{name}"
3297
if suffix:
3298
name = f"{name}_{suffix}"
3299
+ if append:
3300
+ name = f"{setproctitle.getproctitle()}_{name}"
3301
+ else:
3302
+ name = f"{envs.VLLM_PROCESS_NAME_PREFIX}::{name}"
3303
setproctitle.setproctitle(name)
vllm/v1/engine/coordinator.py
+3
-4
@@ -10,11 +10,10 @@ import zmq
10
11
from vllm.config import ParallelConfig
12
from vllm.logger import init_logger
13
-from vllm.utils import get_mp_context, make_zmq_socket
13
+from vllm.utils import get_mp_context, make_zmq_socket, set_process_title
14
from vllm.v1.engine import EngineCoreOutputs, EngineCoreRequestType
15
from vllm.v1.serial_utils import MsgpackDecoder
16
-from vllm.v1.utils import (bind_process_name, get_engine_client_zmq_addr,
17
- shutdown)
16
+from vllm.v1.utils import get_engine_client_zmq_addr, shutdown
17
18
logger = init_logger(__name__)
19
@@ -119,7 +118,7 @@ class DPCoordinatorProc:
118
def __init__(self,
119
engine_count: int,
120
min_stats_update_interval_ms: int = 100):
122
- bind_process_name(self.__class__.__name__)
121
+ set_process_title("DPCoordinator")
122
self.ctx = zmq.Context()
123
124
self.engines = [EngineState() for _ in range(engine_count)]
vllm/v1/engine/core.py
+4
-3
@@ -26,8 +26,8 @@ from vllm.lora.request import LoRARequest
26
from vllm.tasks import POOLING_TASKS, SupportedTask
27
from vllm.transformers_utils.config import (
28
maybe_register_config_serialize_by_value)
29
-from vllm.utils import (bind_process_name, make_zmq_socket,
30
- resolve_obj_by_qualname)
29
+from vllm.utils import (make_zmq_socket, resolve_obj_by_qualname,
30
+ set_process_title)
31
from vllm.v1.core.kv_cache_utils import (get_kv_cache_config,
32
unify_kv_cache_configs)
33
from vllm.v1.core.sched.interface import SchedulerInterface
@@ -425,7 +425,6 @@ class EngineCoreProc(EngineCore):
425
client_handshake_address: Optional[str] = None,
426
engine_index: int = 0,
427
):
428
- bind_process_name(self.__class__.__name__, f"{engine_index}")
428
self.input_queue = queue.Queue[tuple[EngineCoreRequestType, Any]]()
429
self.output_queue = queue.Queue[Union[tuple[int, EngineCoreOutputs],
430
bytes]]()
@@ -630,11 +629,13 @@ class EngineCoreProc(EngineCore):
629
parallel_config: ParallelConfig = kwargs[
630
"vllm_config"].parallel_config
631
if parallel_config.data_parallel_size > 1 or dp_rank > 0:
632
+ set_process_title("DPEngineCore", str(dp_rank))
633
# Set data parallel rank for this engine process.
634
parallel_config.data_parallel_rank = dp_rank
635
parallel_config.data_parallel_rank_local = local_dp_rank
636
engine_core = DPEngineCoreProc(*args, **kwargs)
637
else:
638
+ set_process_title("EngineCore")
639
engine_core = EngineCoreProc(*args, **kwargs)
640
641
engine_core.run_busy_loop()
vllm/v1/executor/multiproc_executor.py
+10
-6
@@ -30,8 +30,8 @@ from vllm.distributed.kv_transfer.kv_connector.utils import KVOutputAggregator
30
from vllm.executor.multiproc_worker_utils import (
31
_add_prefix, set_multiprocessing_worker_envs)
32
from vllm.logger import init_logger
33
-from vllm.utils import (bind_process_name, get_distributed_init_method,
34
- get_loopback_ip, get_mp_context, get_open_port)
33
+from vllm.utils import (get_distributed_init_method, get_loopback_ip,
34
+ get_mp_context, get_open_port, set_process_title)
35
from vllm.v1.executor.abstract import Executor, FailureCallback
36
from vllm.v1.outputs import ModelRunnerOutput
37
from vllm.worker.worker_base import WorkerWrapperBase
@@ -376,10 +376,14 @@ class WorkerProc:
376
}
377
wrapper.init_worker(all_kwargs)
378
self.worker = wrapper
379
- bind_process_name(
380
- self.worker.worker.__class__.__name__,
381
- f"TP{self.rank}_DP{vllm_config.parallel_config.data_parallel_rank}"
382
- )
379
+
380
+ pp_size = vllm_config.parallel_config.pipeline_parallel_size
381
+ tp_size = vllm_config.parallel_config.tensor_parallel_size
382
+ pp_str = f"PP{rank // tp_size}" if pp_size > 1 else ""
383
+ tp_str = f"TP{rank % tp_size}" if tp_size > 1 else ""
384
+ suffix = f"{pp_str}{'_' if pp_str and tp_str else ''}{tp_str}"
385
+ if suffix:
386
+ set_process_title(suffix, append=True)
387
pid = os.getpid()
388
_add_prefix(sys.stdout, f"VllmWorker rank={rank}", pid)
389
_add_prefix(sys.stderr, f"VllmWorker rank={rank}", pid)
vllm/v1/utils.py
+3
-3
@@ -15,8 +15,8 @@ import torch
15
from vllm.logger import init_logger
16
from vllm.usage.usage_lib import (UsageContext, is_usage_stats_enabled,
17
usage_message)
18
-from vllm.utils import (bind_process_name, get_open_port,
19
- get_open_zmq_ipc_path, get_tcp_uri, kill_process_tree)
18
+from vllm.utils import (get_open_port, get_open_zmq_ipc_path, get_tcp_uri,
19
+ kill_process_tree)
20
21
if TYPE_CHECKING:
22
from vllm.v1.engine.coordinator import DPCoordinator
@@ -144,7 +144,7 @@ class APIServerProcessManager:
144
self.listen_address = listen_address
145
self.sock = sock
146
self.args = args
147
- bind_process_name(self.__class__.__name__)
147
+
148
# Start API servers
149
spawn_context = multiprocessing.get_context("spawn")
150
self.processes: list[BaseProcess] = []