| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | from abc import ABC, abstractmethod |
| 4 | from collections.abc import Mapping |
| 5 | from typing import Generic, TypeVar |
| 6 | |
| 7 | import numpy as np |
| 8 | import numpy.typing as npt |
| 9 | from PIL import Image |
| 10 | |
| 11 | from vllm.config.multimodal import ( |
| 12 | AudioDummyOptions, |
| 13 | ImageDummyOptions, |
| 14 | MultiModalDummyOptions, |
| 15 | VideoDummyOptions, |
| 16 | ) |
| 17 | from vllm.inputs import MultiModalDataDict |
| 18 | from vllm.logger import init_logger |
| 19 | |
| 20 | from .context import BaseProcessingInfo |
| 21 | |
| 22 | _I = TypeVar("_I", bound=BaseProcessingInfo) |
| 23 | |
| 24 | logger = init_logger(__name__) |
| 25 | |
| 26 | |
| 27 | class BaseDummyInputsBuilder(ABC, Generic[_I]): |
| 28 | """Abstract base class that constructs the dummy data to profile |
| 29 | multi-modal models. |
| 30 | """ |
| 31 | |
| 32 | def __init__(self, info: _I) -> None: |
| 33 | super().__init__() |
| 34 | |
| 35 | self.info = info |
| 36 | |
| 37 | @abstractmethod |
| 38 | def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str: |
| 39 | """Build the text input corresponding to `mm_counts`.""" |
| 40 | raise NotImplementedError |
| 41 | |
| 42 | @abstractmethod |
| 43 | def get_dummy_mm_data( |
| 44 | self, |
| 45 | seq_len: int, |
| 46 | mm_counts: Mapping[str, int], |
| 47 | mm_options: MultiModalDummyOptions, |
| 48 | ) -> MultiModalDataDict: |
| 49 | """Build the multimodal input which, after processing, results in |
| 50 | the maximum possible number of placeholder tokens. |
| 51 | |
| 52 | Args: |
| 53 | seq_len: Sequence length |
| 54 | mm_counts: Count of items per modality |
| 55 | mm_options: Configurable options per modality (optional). |
| 56 | If None, use model defaults for backward compatibility. |
| 57 | If provided, models can use these to customize dummy |
| 58 | data generation. |
| 59 | |
| 60 | """ |
| 61 | raise NotImplementedError |
| 62 | |
| 63 | def _get_dummy_audios( |
| 64 | self, |
| 65 | *, |
| 66 | length: int, |
| 67 | num_audios: int, |
| 68 | overrides: AudioDummyOptions | None = None, |
| 69 | ) -> list[npt.NDArray]: |
| 70 | if num_audios == 0: |
| 71 | return [] |
| 72 | if overrides and overrides.length: |
| 73 | if overrides.length > length: |
| 74 | logger.warning( |
| 75 | "audio.length override (%d) exceeds model's " |
| 76 | "maximum length (%d), will be ignored", |
| 77 | overrides.length, |
| 78 | length, |
| 79 | ) |
| 80 | length = min(length, overrides.length) |
| 81 | audio = np.zeros((length,)) |
| 82 | return [audio] * num_audios |
| 83 | |
| 84 | def _get_dummy_images( |
| 85 | self, |
| 86 | *, |
| 87 | width: int, |
| 88 | height: int, |
| 89 | num_images: int, |
| 90 | overrides: ImageDummyOptions | None = None, |
| 91 | ) -> list[Image.Image]: |
| 92 | if num_images == 0: |
| 93 | return [] |
| 94 | if overrides: |
| 95 | if overrides.width: |
| 96 | if overrides.width > width: |
| 97 | logger.warning( |
| 98 | "image.width override (%d) exceeds model's " |
| 99 | "maximum width (%d), will be ignored", |
| 100 | overrides.width, |
| 101 | width, |
| 102 | ) |
| 103 | width = min(width, overrides.width) |
| 104 | if overrides.height: |
| 105 | if overrides.height > height: |
| 106 | logger.warning( |
| 107 | "image.height override (%d) exceeds model's " |
| 108 | "maximum height (%d), will be ignored", |
| 109 | overrides.height, |
| 110 | height, |
| 111 | ) |
| 112 | height = min(height, overrides.height) |
| 113 | image = Image.new("RGB", (width, height), color=255) |
| 114 | return [image] * num_images |
| 115 | |
| 116 | def _get_dummy_videos( |
| 117 | self, |
| 118 | *, |
| 119 | width: int, |
| 120 | height: int, |
| 121 | num_frames: int, |
| 122 | num_videos: int, |
| 123 | overrides: VideoDummyOptions | None = None, |
| 124 | ) -> list[npt.NDArray]: |
| 125 | if num_videos == 0: |
| 126 | return [] |
| 127 | if overrides: |
| 128 | if overrides.num_frames: |
| 129 | if overrides.num_frames > num_frames: |
| 130 | logger.warning( |
| 131 | "video.num_frames override (%d) exceeds model's " |
| 132 | "maximum number of frames (%d), will be ignored", |
| 133 | overrides.num_frames, |
| 134 | num_frames, |
| 135 | ) |
| 136 | num_frames = min(num_frames, overrides.num_frames) |
| 137 | if overrides.width: |
| 138 | if overrides.width > width: |
| 139 | logger.warning( |
| 140 | "video.width override (%d) exceeds model's " |
| 141 | "maximum width (%d), will be ignored", |
| 142 | overrides.width, |
| 143 | width, |
| 144 | ) |
| 145 | width = min(width, overrides.width) |
| 146 | if overrides.height: |
| 147 | if overrides.height > height: |
| 148 | logger.warning( |
| 149 | "video.height override (%d) exceeds model's " |
| 150 | "maximum height (%d), will be ignored", |
| 151 | overrides.height, |
| 152 | height, |
| 153 | ) |
| 154 | height = min(height, overrides.height) |
| 155 | video = np.full((num_frames, height, width, 3), 255, dtype=np.uint8) |
| 156 | return [video] * num_videos |