| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import logging |
| 5 | from collections.abc import Sequence |
| 6 | |
| 7 | import torch |
| 8 | |
| 9 | from vllm.entrypoints.pooling.typing import AnyPoolingRequest |
| 10 | from vllm.entrypoints.serve.engine.typing import AnyRequest |
| 11 | from vllm.logger import init_logger |
| 12 | from vllm.lora.request import LoRARequest |
| 13 | from vllm.pooling_params import PoolingParams |
| 14 | from vllm.sampling_params import BeamSearchParams, SamplingParams |
| 15 | |
| 16 | logger = init_logger(__name__) |
| 17 | |
| 18 | |
| 19 | class RequestLogger: |
| 20 | def __init__(self, *, max_log_len: int | None) -> None: |
| 21 | self.max_log_len = max_log_len |
| 22 | |
| 23 | if not logger.isEnabledFor(logging.INFO): |
| 24 | logger.warning_once( |
| 25 | "`--enable-log-requests` is set but " |
| 26 | "the minimum log level is higher than INFO. " |
| 27 | "No request information will be logged." |
| 28 | ) |
| 29 | elif not logger.isEnabledFor(logging.DEBUG): |
| 30 | logger.info_once( |
| 31 | "`--enable-log-requests` is set but " |
| 32 | "the minimum log level is higher than DEBUG. " |
| 33 | "Only limited information will be logged to minimize overhead. " |
| 34 | "To view more details, set `--log-level DEBUG`." |
| 35 | ) |
| 36 | |
| 37 | def log_inputs( |
| 38 | self, |
| 39 | request_id: str, |
| 40 | prompt: str | None, |
| 41 | prompt_token_ids: list[int] | None, |
| 42 | prompt_embeds: torch.Tensor | None, |
| 43 | params: SamplingParams | PoolingParams | BeamSearchParams | None, |
| 44 | lora_request: LoRARequest | None, |
| 45 | ) -> None: |
| 46 | if logger.isEnabledFor(logging.DEBUG): |
| 47 | max_log_len = self.max_log_len |
| 48 | if max_log_len is not None: |
| 49 | if prompt is not None: |
| 50 | prompt = prompt[:max_log_len] |
| 51 | |
| 52 | if prompt_token_ids is not None: |
| 53 | prompt_token_ids = prompt_token_ids[:max_log_len] |
| 54 | |
| 55 | logger.debug( |
| 56 | "Request %s details: prompt: %r, " |
| 57 | "prompt_token_ids: %s, " |
| 58 | "prompt_embeds shape: %s.", |
| 59 | request_id, |
| 60 | prompt, |
| 61 | prompt_token_ids, |
| 62 | prompt_embeds.shape if prompt_embeds is not None else None, |
| 63 | ) |
| 64 | |
| 65 | logger.info( |
| 66 | "Received request %s: params: %s, lora_request: %s.", |
| 67 | request_id, |
| 68 | params, |
| 69 | lora_request, |
| 70 | ) |
| 71 | |
| 72 | def log_request_body(self, request: AnyRequest | AnyPoolingRequest) -> None: |
| 73 | if logger.isEnabledFor(logging.DEBUG): |
| 74 | max_log_len = self.max_log_len if self.max_log_len is not None else -1 |
| 75 | logger.debug( |
| 76 | "Request %s JSON body: %s", |
| 77 | getattr(request, "request_id", "N/A"), |
| 78 | request.model_dump_json(exclude_unset=True)[:max_log_len], |
| 79 | ) |
| 80 | |
| 81 | def log_outputs( |
| 82 | self, |
| 83 | request_id: str, |
| 84 | outputs: str, |
| 85 | output_token_ids: Sequence[int] | None, |
| 86 | finish_reason: str | None = None, |
| 87 | is_streaming: bool = False, |
| 88 | delta: bool = False, |
| 89 | ) -> None: |
| 90 | max_log_len = self.max_log_len |
| 91 | if max_log_len is not None and outputs is not None: |
| 92 | outputs = outputs[:max_log_len] |
| 93 | |
| 94 | stream_info = "" |
| 95 | if is_streaming: |
| 96 | stream_info = " (streaming delta)" if delta else " (streaming complete)" |
| 97 | |
| 98 | if logger.isEnabledFor(logging.DEBUG): |
| 99 | if max_log_len is not None and output_token_ids is not None: |
| 100 | output_token_ids = list(output_token_ids)[:max_log_len] |
| 101 | |
| 102 | logger.debug( |
| 103 | "Generated response %s%s details: output_token_ids: %s", |
| 104 | request_id, |
| 105 | stream_info, |
| 106 | output_token_ids, |
| 107 | ) |
| 108 | |
| 109 | logger.info( |
| 110 | "Generated response %s%s: output: %r, finish_reason: %s", |
| 111 | request_id, |
| 112 | stream_info, |
| 113 | outputs, |
| 114 | finish_reason, |
| 115 | ) |