@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 6da0078523

Fork 0

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

[Feat] Allow custom naming of vLLM processes (#21445)

Signed-off-by: chaunceyjiang <chaunceyjiang@gmail.com>

Chauncey committed Jul 24, 2025 at 18:15 UTC 6da0078523d42575ff9a369dbf3546f8989ceb52

10 files changed +46 -17

requirements/common.txt

@@ -48,3 +48,4 @@ scipy # Required for phi-4-multimodal-instruct

48 ninja # Required for xgrammar, rocm, tpu, xpu

49 pybase64 # fast base64 implementation

50 cbor2 # Required for cross-language serialization of hashable objects

51 +setproctitle # Used to set process names for better debugging and monitoring

requirements/docs.txt

@@ -22,6 +22,7 @@ pillow

22 psutil

23 pybase64

24 pydantic

25 +setproctitle

26 torch

27 transformers

28 zmq

vllm/entrypoints/cli/serve.py

+2 -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, get_tcp_uri

24 +from vllm.utils import FlexibleArgumentParser, bind_process_name, 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):

78 if args.api_server_count > 1:

79 raise ValueError("api_server_count can't be set in headless mode")

80 -

80 + bind_process_name("APIServer_Headless")

81 # Create the EngineConfig.

82 engine_args = vllm.AsyncEngineArgs.from_cli_args(args)

83 usage_context = UsageContext.OPENAI_API_SERVER

vllm/entrypoints/openai/api_server.py

+4 -3

@@ -101,8 +101,9 @@ from vllm.transformers_utils.config import (

101 maybe_register_config_serialize_by_value)

102 from vllm.transformers_utils.tokenizer import MistralTokenizer

103 from vllm.usage.usage_lib import UsageContext

104 -from vllm.utils import (Device, FlexibleArgumentParser, get_open_zmq_ipc_path,

105 - is_valid_ipv6_address, set_ulimit)

104 +from vllm.utils import (Device, FlexibleArgumentParser, bind_process_name,

105 + get_open_zmq_ipc_path, is_valid_ipv6_address,

106 + set_ulimit)

107 from vllm.v1.metrics.prometheus import get_prometheus_registry

108 from vllm.version import __version__ as VLLM_VERSION

109

@@ -1804,7 +1805,7 @@ async def run_server_worker(listen_address,

1805 ToolParserManager.import_tool_parser(args.tool_parser_plugin)

1806

1807 server_index = client_config.get("client_index", 0) if client_config else 0

1807 -

1808 + bind_process_name("APIServer", str(server_index))

1809 # Load logging config for uvicorn if specified

1810 log_config = load_log_config(args.log_config_file)

1811 if log_config is not None:

vllm/envs.py

@@ -985,6 +985,12 @@ environment_variables: dict[str, Callable[[], Any]] = {

985 # Used to force set up loopback IP

986 "VLLM_LOOPBACK_IP":

987 lambda: os.getenv("VLLM_LOOPBACK_IP", ""),

988 +

989 + # Used to set the process name prefix for vLLM processes.

990 + # This is useful for debugging and monitoring purposes.

991 + # The default value is "VLLM".

992 + "VLLM_PROCESS_NAME_PREFIX":

993 + lambda: os.getenv("VLLM_PROCESS_NAME_PREFIX", "VLLM"),

994 }

995

996 # --8<-- [end:env-vars-definition]

vllm/utils/__init__.py

+14

@@ -58,6 +58,7 @@ import numpy as np

58 import numpy.typing as npt

59 import psutil

60 import regex as re

61 +import setproctitle

62 import torch

63 import torch.types

64 import yaml

@@ -3278,3 +3279,16 @@ def has_deep_gemm() -> bool:

3279 """Whether the optional `deep_gemm` package is available."""

3280

3281 return _has_module("deep_gemm")

3282 +

3283 +

3284 +def bind_process_name(name: str, suffix: str = "") -> None:

3285 + """Bind the process name to a specific name with an optional suffix.

3286 +

3287 + Args:

3288 + name: The base name to bind the process to.

3289 + suffix: An optional suffix to append to the base name.

3290 + """

3291 + name = f"{envs.VLLM_PROCESS_NAME_PREFIX}::{name}"

3292 + if suffix:

3293 + name = f"{name}_{suffix}"

3294 + setproctitle.setproctitle(name)

vllm/v1/engine/coordinator.py

+6 -5

@@ -13,7 +13,8 @@ from vllm.logger import init_logger

13 from vllm.utils import get_mp_context, make_zmq_socket

14 from vllm.v1.engine import EngineCoreOutputs, EngineCoreRequestType

15 from vllm.v1.serial_utils import MsgpackDecoder

16 -from vllm.v1.utils import get_engine_client_zmq_addr, shutdown

16 +from vllm.v1.utils import (bind_process_name, get_engine_client_zmq_addr,

17 + shutdown)

19 logger = init_logger(__name__)

@@ -79,7 +80,7 @@ class DPCoordinator:

81 context = get_mp_context()

82 self.proc: multiprocessing.Process = context.Process(

82 - target=CoordinatorProc.run_coordinator,

83 + target=DPCoordinatorProc.run_coordinator,

84 name="VLLM_DP_Coordinator",

85 kwargs={

86 "engine_count": parallel_config.data_parallel_size,

@@ -113,12 +114,12 @@ class EngineState:

114 self.request_counts = [0, 0] # [waiting, running]

115

116

116 -class CoordinatorProc:

117 +class DPCoordinatorProc:

118

119 def __init__(self,

120 engine_count: int,

121 min_stats_update_interval_ms: int = 100):

121 -

122 + bind_process_name(self.__class__.__name__)

123 self.ctx = zmq.Context()

124

125 self.engines = [EngineState() for _ in range(engine_count)]

@@ -137,7 +138,7 @@ class CoordinatorProc:

138 back_publish_address: str,

139 min_stats_update_interval_ms: int = 100,

140 ):

140 - coordinator = CoordinatorProc(

141 + coordinator = DPCoordinatorProc(

142 engine_count=engine_count,

143 min_stats_update_interval_ms=min_stats_update_interval_ms)

144 try:

vllm/v1/engine/core.py

+3 -1

@@ -25,7 +25,8 @@ from vllm.logging_utils.dump_input import dump_engine_exception

25 from vllm.lora.request import LoRARequest

26 from vllm.transformers_utils.config import (

27 maybe_register_config_serialize_by_value)

28 -from vllm.utils import make_zmq_socket, resolve_obj_by_qualname

28 +from vllm.utils import (bind_process_name, make_zmq_socket,

29 + resolve_obj_by_qualname)

30 from vllm.v1.core.kv_cache_utils import (get_kv_cache_config,

31 unify_kv_cache_configs)

32 from vllm.v1.core.sched.interface import SchedulerInterface

@@ -411,6 +412,7 @@ class EngineCoreProc(EngineCore):

412 client_handshake_address: Optional[str] = None,

413 engine_index: int = 0,

414 ):

415 + bind_process_name(self.__class__.__name__, f"{engine_index}")

416 self.input_queue = queue.Queue[tuple[EngineCoreRequestType, Any]]()

417 self.output_queue = queue.Queue[Union[tuple[int, EngineCoreOutputs],

418 bytes]]()

vllm/v1/executor/multiproc_executor.py

+6 -3

@@ -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 (get_distributed_init_method, get_loopback_ip,

34 - get_mp_context, get_open_port)

33 +from vllm.utils import (bind_process_name, get_distributed_init_method,

34 + get_loopback_ip, get_mp_context, get_open_port)

35 from vllm.v1.executor.abstract import Executor, FailureCallback

36 from vllm.v1.outputs import ModelRunnerOutput

37 from vllm.worker.worker_base import WorkerWrapperBase

@@ -365,7 +365,10 @@ class WorkerProc:

365 }

366 wrapper.init_worker(all_kwargs)

367 self.worker = wrapper

368 -

368 + bind_process_name(

369 + self.worker.worker.__class__.__name__,

370 + f"TP{self.rank}_DP{vllm_config.parallel_config.data_parallel_rank}"

371 + )

372 pid = os.getpid()

373 _add_prefix(sys.stdout, f"VllmWorker rank={rank}", pid)

374 _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 (get_open_port, get_open_zmq_ipc_path, get_tcp_uri,

19 - kill_process_tree)

18 +from vllm.utils import (bind_process_name, get_open_port,

19 + get_open_zmq_ipc_path, get_tcp_uri, kill_process_tree)

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 -

147 + bind_process_name(self.__class__.__name__)

148 # Start API servers

149 spawn_context = multiprocessing.get_context("spawn")

150 self.processes: list[BaseProcess] = []