@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 0cdd213641

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, 05:30 UTC

[Misc] Improve Worker process title and logging prefix (#22205)

Signed-off-by: 22quinn <33176974+22quinn@users.noreply.github.com>

22quinn committed Sep 8, 2025 at 21:43 UTC 0cdd213641d7adcf7ad0e9cf79867344ea1291d9

4 files changed +37 -23

vllm/utils/__init__.py

+3 -7

@@ -3359,7 +3359,7 @@ def has_triton_kernels() -> bool:

3359

3360 def set_process_title(name: str,

3361 suffix: str = "",

3362 - append: bool = False) -> None:

3362 + prefix: str = envs.VLLM_PROCESS_NAME_PREFIX) -> None:

3363 """

3364 Set the current process title to a specific name with an

3365 optional suffix.

@@ -3367,15 +3367,11 @@ def set_process_title(name: str,

3367 Args:

3368 name: The title to assign to the current process.

3369 suffix: An optional suffix to append to the base name.

3370 - append: Whether to append to the existing process title.

3370 + prefix: A prefix to prepend to the front separated by `::`.

3371 """

3372 if suffix:

3373 name = f"{name}_{suffix}"

3374 - if append:

3375 - name = f"{setproctitle.getproctitle()}_{name}"

3376 - else:

3377 - name = f"{envs.VLLM_PROCESS_NAME_PREFIX}::{name}"

3378 - setproctitle.setproctitle(name)

3374 + setproctitle.setproctitle(f"{prefix}::{name}")

3375

3376

3377 def _add_prefix(file: TextIO, worker_name: str, pid: int) -> None:

vllm/v1/engine/core.py

+3 -3

@@ -224,7 +224,7 @@ class EngineCore:

224

225 def add_request(self, request: Request, request_wave: int = 0):

226 """Add request to the scheduler.

227 -

227 +

228 `request_wave`: indicate which wave of requests this is expected to

229 belong to in DP case

230 """

@@ -433,7 +433,7 @@ class EngineCore:

433 def preprocess_add_request(

434 self, request: EngineCoreRequest) -> tuple[Request, int]:

435 """Preprocess the request.

436 -

436 +

437 This function could be directly used in input processing thread to allow

438 request initialization running in parallel with Model forward

439 """

@@ -697,7 +697,7 @@ class EngineCoreProc(EngineCore):

697 parallel_config: ParallelConfig = kwargs[

698 "vllm_config"].parallel_config

699 if parallel_config.data_parallel_size > 1 or dp_rank > 0:

700 - set_process_title("DPEngineCore", str(dp_rank))

700 + set_process_title("EngineCore", f"DP{dp_rank}")

701 decorate_logs()

702 # Set data parallel rank for this engine process.

703 parallel_config.data_parallel_rank = dp_rank

vllm/v1/engine/utils.py

+1 -1

@@ -116,7 +116,7 @@ class CoreEngineProcManager:

116 local_dp_ranks.append(local_index)

117 self.processes.append(

118 context.Process(target=target_fn,

119 - name=f"EngineCore_{global_index}",

119 + name=f"EngineCore_DP{global_index}",

120 kwargs=common_kwargs | {

121 "dp_rank": global_index,

122 "local_dp_rank": local_index,

vllm/v1/executor/multiproc_executor.py

+30 -12

@@ -26,6 +26,8 @@ from vllm.distributed import (destroy_distributed_environment,

26 from vllm.distributed.device_communicators.shm_broadcast import (Handle,

27 MessageQueue)

28 from vllm.distributed.kv_transfer.kv_connector.utils import KVOutputAggregator

29 +from vllm.distributed.parallel_state import (get_dp_group, get_ep_group,

30 + get_pp_group, get_tp_group)

31 from vllm.executor.multiproc_worker_utils import (

32 set_multiprocessing_worker_envs)

33 from vllm.logger import init_logger

@@ -397,17 +399,6 @@ class WorkerProc:

399 wrapper.init_worker(all_kwargs)

400 self.worker = wrapper

401

400 - pp_size = vllm_config.parallel_config.pipeline_parallel_size

401 - tp_size = vllm_config.parallel_config.tensor_parallel_size

402 - pp_str = f"PP{rank // tp_size}" if pp_size > 1 else ""

403 - tp_str = f"TP{rank % tp_size}" if tp_size > 1 else ""

404 - suffix = f"{pp_str}{'_' if pp_str and tp_str else ''}{tp_str}"

405 - process_name = "VllmWorker"

406 - if suffix:

407 - set_process_title(suffix, append=True)

408 - process_name = f"{process_name} {suffix}"

409 - decorate_logs(process_name)

410 -

402 # Initialize MessageQueue for receiving SchedulerOutput

403 self.rpc_broadcast_mq = MessageQueue.create_from_handle(

404 input_shm_handle, self.worker.rank)

@@ -425,8 +416,14 @@ class WorkerProc:

416 name="WorkerAsyncOutputCopy")

417 self.async_output_copy_thread.start()

418

428 - # Initialize device and loads weights

419 + # Initialize device

420 self.worker.init_device()

421 +

422 + # Set process title and log prefix

423 + self.setup_proc_title_and_log_prefix(

424 + enable_ep=vllm_config.parallel_config.enable_expert_parallel)

425 +

426 + # Load model

427 self.worker.load_model()

428

429 @staticmethod

@@ -663,3 +660,24 @@ class WorkerProc:

660

661 if output_rank is None or self.rank == output_rank:

662 self.handle_output(output)

663 +

664 + @staticmethod

665 + def setup_proc_title_and_log_prefix(enable_ep: bool) -> None:

666 + dp_size = get_dp_group().world_size

667 + dp_rank = get_dp_group().rank_in_group

668 + pp_size = get_pp_group().world_size

669 + pp_rank = get_pp_group().rank_in_group

670 + tp_size = get_tp_group().world_size

671 + tp_rank = get_tp_group().rank_in_group

672 + process_name = "Worker"

673 + if dp_size > 1:

674 + process_name += f"_DP{dp_rank}"

675 + if pp_size > 1:

676 + process_name += f"_PP{pp_rank}"

677 + if tp_size > 1:

678 + process_name += f"_TP{tp_rank}"

679 + if enable_ep:

680 + ep_rank = get_ep_group().rank_in_group

681 + process_name += f"_EP{ep_rank}"

682 + set_process_title(name=process_name)

683 + decorate_logs(process_name)