Skip to content

vllm.multimodal.processing.dummy_inputs

Classes:

BaseDummyInputsBuilder

Bases: ABC, Generic[_I]

Abstract base class that constructs the dummy data to profile multi-modal models.

Methods:

Source code in vllm/multimodal/processing/dummy_inputs.py 
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
class BaseDummyInputsBuilder(ABC, Generic[_I]):
    """Abstract base class that constructs the dummy data to profile
    multi-modal models.
    """

    def __init__(self, info: _I) -> None:
        super().__init__()

        self.info = info

    @abstractmethod
    def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
        """Build the text input corresponding to `mm_counts`."""
        raise NotImplementedError

    @abstractmethod
    def get_dummy_mm_data(
        self,
        seq_len: int,
        mm_counts: Mapping[str, int],
        mm_options: MultiModalDummyOptions,
    ) -> MultiModalDataDict:
        """Build the multimodal input which, after processing, results in
        the maximum possible number of placeholder tokens.

        Args:
            seq_len: Sequence length
            mm_counts: Count of items per modality
            mm_options: Configurable options per modality (optional).
                       If None, use model defaults for backward compatibility.
                       If provided, models can use these to customize dummy
                       data generation.

        """
        raise NotImplementedError

    def _get_dummy_audios(
        self,
        *,
        length: int,
        num_audios: int,
        overrides: AudioDummyOptions | None = None,
    ) -> list[npt.NDArray]:
        if num_audios == 0:
            return []
        if overrides and overrides.length:
            if overrides.length > length:
                logger.warning(
                    "audio.length override (%d) exceeds model's "
                    "maximum length (%d), will be ignored",
                    overrides.length,
                    length,
                )
            length = min(length, overrides.length)
        audio = np.zeros((length,))
        return [audio] * num_audios

    def _get_dummy_images(
        self,
        *,
        width: int,
        height: int,
        num_images: int,
        overrides: ImageDummyOptions | None = None,
    ) -> list[Image.Image]:
        if num_images == 0:
            return []
        if overrides:
            if overrides.width:
                if overrides.width > width:
                    logger.warning(
                        "image.width override (%d) exceeds model's "
                        "maximum width (%d), will be ignored",
                        overrides.width,
                        width,
                    )
                width = min(width, overrides.width)
            if overrides.height:
                if overrides.height > height:
                    logger.warning(
                        "image.height override (%d) exceeds model's "
                        "maximum height (%d), will be ignored",
                        overrides.height,
                        height,
                    )
                height = min(height, overrides.height)
        image = Image.new("RGB", (width, height), color=255)
        return [image] * num_images

    def _get_dummy_videos(
        self,
        *,
        width: int,
        height: int,
        num_frames: int,
        num_videos: int,
        overrides: VideoDummyOptions | None = None,
    ) -> list[npt.NDArray]:
        if num_videos == 0:
            return []
        if overrides:
            if overrides.num_frames:
                if overrides.num_frames > num_frames:
                    logger.warning(
                        "video.num_frames override (%d) exceeds model's "
                        "maximum number of frames (%d), will be ignored",
                        overrides.num_frames,
                        num_frames,
                    )
                num_frames = min(num_frames, overrides.num_frames)
            if overrides.width:
                if overrides.width > width:
                    logger.warning(
                        "video.width override (%d) exceeds model's "
                        "maximum width (%d), will be ignored",
                        overrides.width,
                        width,
                    )
                width = min(width, overrides.width)
            if overrides.height:
                if overrides.height > height:
                    logger.warning(
                        "video.height override (%d) exceeds model's "
                        "maximum height (%d), will be ignored",
                        overrides.height,
                        height,
                    )
                height = min(height, overrides.height)
        video = np.full((num_frames, height, width, 3), 255, dtype=np.uint8)
        return [video] * num_videos

get_dummy_mm_data(seq_len, mm_counts, mm_options) abstractmethod

Build the multimodal input which, after processing, results in the maximum possible number of placeholder tokens.

Parameters:

  • seq_len

    (int) –

    Sequence length

  • mm_counts

    (Mapping[str, int]) –

    Count of items per modality

  • mm_options

    (MultiModalDummyOptions) –

    Configurable options per modality (optional). If None, use model defaults for backward compatibility. If provided, models can use these to customize dummy data generation.

Source code in vllm/multimodal/processing/dummy_inputs.py 
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
@abstractmethod
def get_dummy_mm_data(
    self,
    seq_len: int,
    mm_counts: Mapping[str, int],
    mm_options: MultiModalDummyOptions,
) -> MultiModalDataDict:
    """Build the multimodal input which, after processing, results in
    the maximum possible number of placeholder tokens.

    Args:
        seq_len: Sequence length
        mm_counts: Count of items per modality
        mm_options: Configurable options per modality (optional).
                   If None, use model defaults for backward compatibility.
                   If provided, models can use these to customize dummy
                   data generation.

    """
    raise NotImplementedError

get_dummy_text(mm_counts) abstractmethod

Build the text input corresponding to mm_counts.

Source code in vllm/multimodal/processing/dummy_inputs.py 
37
38
39
40
@abstractmethod
def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
    """Build the text input corresponding to `mm_counts`."""
    raise NotImplementedError