diff --git a/vllm/v1/engine/async_llm.py b/vllm/v1/engine/async_llm.py index 0a7263ee2f..4329d9d076 100644 --- a/vllm/v1/engine/async_llm.py +++ b/vllm/v1/engine/async_llm.py @@ -7,9 +7,7 @@ import time import warnings from collections.abc import AsyncGenerator, Iterable, Mapping from copy import copy -from typing import Any - -import torch +from typing import Any, Optional import vllm.envs as envs from vllm import TokensPrompt @@ -34,6 +32,7 @@ from vllm.lora.request import LoRARequest from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalRegistry from vllm.outputs import STREAM_FINISHED, PoolingRequestOutput, RequestOutput from vllm.pooling_params import PoolingParams +from vllm.profiler.wrapper import TorchProfilerWrapper from vllm.renderers import renderer_from_config from vllm.renderers.inputs.preprocess import extract_prompt_components from vllm.sampling_params import RequestOutputKind, SamplingParams @@ -92,6 +91,7 @@ class AsyncLLM(EngineClient): client_addresses: dict[str, Any] | None = None, client_count: int = 1, client_index: int = 0, + profiler: Optional[TorchProfilerWrapper] = None, # type: ignore # noqa ) -> None: """ Create an AsyncLLM. @@ -195,19 +195,12 @@ class AsyncLLM(EngineClient): profiler_dir, ) worker_name = f"{socket.gethostname()}_{os.getpid()}.async_llm" - self.profiler = torch.profiler.profile( - activities=[ - torch.profiler.ProfilerActivity.CPU, - ], - with_stack=vllm_config.profiler_config.torch_profiler_with_stack, - on_trace_ready=torch.profiler.tensorboard_trace_handler( - profiler_dir, - worker_name=worker_name, - use_gzip=vllm_config.profiler_config.torch_profiler_use_gzip, - ), + self.profiler = TorchProfilerWrapper( + vllm_config.profiler_config, + worker_name=worker_name, + local_rank=0, + activities=["CPU"], ) - else: - self.profiler = None @classmethod def from_vllm_config(