vllm.config.multimodal ¶

@dataclass(config=ConfigDict(extra="forbid"))
class AudioDummyOptions(BaseDummyOptions):
    """Options for generating dummy audio data during profiling."""

    length: int | None = Field(None, gt=0)

@dataclass
class BaseDummyOptions:
    """Base options for generating dummy data during profiling."""

    count: int = Field(999, ge=0)

@dataclass(config=ConfigDict(extra="forbid"))
class ImageDummyOptions(BaseDummyOptions):
    """Options for generating dummy image data during profiling."""

    width: int | None = Field(None, gt=0)
    height: int | None = Field(None, gt=0)

@config
class MultiModalConfig:
    """Controls the behavior of multimodal models."""

    language_model_only: bool = False
    """If True, disables all multimodal inputs by setting all modality limits to 0.
    Equivalent to setting `--limit-mm-per-prompt` to 0 for every modality."""
    limit_per_prompt: MMDummyOptions = Field(default_factory=dict)
    """The maximum number of input items and options allowed per
    prompt for each modality.

    Defaults to 999 for each modality.

    Legacy format (count only):
        {"image": 16, "video": 2}

    Configurable format (with options):
        {"video": {"count": 1, "num_frames": 32, "width": 512, "height": 512},
        "image": {"count": 5, "width": 512, "height": 512}}

    Mixed format (combining both):
        {"image": 16, "video": {"count": 1, "num_frames": 32, "width": 512,
        "height": 512}}
    """
    enable_mm_embeds: bool = False
    """If `True`, enables passing multimodal embeddings:
    for `LLM` class, this refers to tensor inputs under `multi_modal_data`;
    for the OpenAI-compatible server, this refers to chat messages with content
    `"type": "*_embeds"`.

    When enabled with `--limit-mm-per-prompt` set to 0 for a modality,
    precomputed embeddings skip count validation for that modality, 
    saving memory by not loading encoder modules while still enabling 
    embeddings as an input. Limits greater than 0 still apply to embeddings.

    WARNING: The vLLM engine may crash if incorrect shape of embeddings is passed.
    Only enable this flag for trusted users!"""
    media_io_kwargs: dict[str, dict[str, Any]] = Field(default_factory=dict)
    """Additional args passed to process media inputs, keyed by modalities.
    For example, to set num_frames for video, set
    `--media-io-kwargs '{"video": {"num_frames": 40} }'`"""
    mm_processor_kwargs: dict[str, object] | None = None
    """Arguments to be forwarded to the model's processor for multi-modal data,
    e.g., image processor. Overrides for the multi-modal processor obtained
    from `transformers.AutoProcessor.from_pretrained`.

    The available overrides depend on the model that is being run.

    For example, for Phi-3-Vision:
    `{"num_crops": 4}`."""
    mm_device_do_normalize: bool | None = True
    """
    Move the do_normalize computation in the mm preprocessing to before the ViT, 
    and let the device do it, so that CPU computation can be saved.
    """
    mm_processor_cache_gb: float = Field(default=4, ge=0)
    """The size (in GiB) of the multi-modal processor cache, which is used to
    avoid re-processing past multi-modal inputs.

    This cache is duplicated for each API process and engine core process,
    resulting in a total memory usage of
    `mm_processor_cache_gb * (api_server_count + data_parallel_size)`.

    A single processed item larger than this budget is served uncached
    (with a warning) instead of failing. Raise this value to cache such items.

    Set to `0` to disable this cache completely (not recommended)."""
    mm_processor_cache_type: MMCacheType = "lru"
    """Type of cache to use for the multi-modal preprocessor/mapper. If `shm`,
    use shared memory FIFO cache. If `lru`, use mirrored LRU cache."""
    mm_hasher_algorithm: MMHasherAlgorithm = Field(
        default_factory=_get_mm_hasher_algorithm
    )
    """Hash algorithm to use for multi-modal input caching. Use `"sha256"` or
    `"sha512"` for FIPS-compliant deployments."""
    mm_shm_cache_max_object_size_mb: int = Field(default=128, ge=0)
    """Size limit (in MiB) for each object stored in the multi-modal processor
    shared memory cache. Only effective when `mm_processor_cache_type` is
    `"shm"`."""
    mm_encoder_only: bool = False
    """
    When enabled, skips the language component of the model.

    This is usually only valid in disaggregated Encoder process.
    """
    mm_encoder_tp_mode: MMEncoderTPMode = "weights"
    """Indicates how to optimize multi-modal encoder inference using tensor
    parallelism (TP).

    - `"weights"`: Within the same vLLM engine, split the weights of
      each layer across TP ranks. (default TP behavior)
    - `"data"`: Within the same vLLM engine, split the batched input data
      across TP ranks to process the data in parallel, while hosting
      the full weights on each TP rank.
      This batch-level DP is not to be confused with API request-level
      DP (which is controlled by `--data-parallel-size`).
      This is only supported on a per-model basis and falls back to
      `"weights"` if the encoder does not support DP."""
    mm_encoder_attn_backend: AttentionBackendEnum | None = None
    """Optional override for the multi-modal encoder attention backend when
    using vision transformers. Accepts any value from
    `vllm.v1.attention.backends.registry.AttentionBackendEnum` (e.g. `FLASH_ATTN`)."""
    mm_encoder_attn_dtype: Literal["fp8"] | None = None
    """Optional dtype override for ViT encoder attention. Set to `"fp8"` to
    enable FP8 quantization via the FlashInfer cuDNN backend. When set to
    `"fp8"` without a scale file, dynamic scaling is used automatically.
    See docs/features/quantization/fp8_vit_attn.md for details."""
    mm_encoder_fp8_scale_path: str | None = None
    """Path to a JSON file containing per-layer FP8 Q/K/V scales for ViT
    encoder attention. When provided (with `mm_encoder_attn_dtype="fp8"`),
    static scaling is used. When omitted, dynamic scaling is used."""
    mm_encoder_fp8_scale_save_path: str | None = None
    """When set with dynamic FP8 scaling (`mm_encoder_attn_dtype="fp8"`
    and no `mm_encoder_fp8_scale_path`), saves the calibrated scales to
    this file after the amax history buffer is full. The saved file can
    then be used as `mm_encoder_fp8_scale_path` in subsequent runs."""
    mm_encoder_fp8_scale_save_margin: float = Field(default=1.5, gt=0.0)
    """Safety margin multiplied onto scales when auto-saving. A value > 1
    leaves headroom so that inputs with larger activations than the
    calibration set do not overflow FP8 range. Default 1.5."""
    interleave_mm_strings: bool = False
    """Enable fully interleaved support for multimodal prompts, while using
    --chat-template-content-format=string."""
    skip_mm_profiling: bool = False
    """When enabled, skips multimodal memory profiling and only profiles with
    language backbone model during engine initialization.

    This reduces engine startup time but shifts the responsibility to users for
    estimating the peak memory usage of the activation of multimodal encoder and
    embedding cache."""
    video_pruning_rate: float | None = Field(default=None, ge=0.0, lt=1.0)
    """Fraction of video tokens to prune from each video. Value sits in range
    [0;1); pruning is enabled when it is greater than 0. The pruning algorithm
    is selected by `video_pruning_method`.
    """
    video_pruning_method: VideoPruningMethod = "evs"
    """Video token pruning algorithm applied when `video_pruning_rate` > 0:
    - "evs": Efficient Video Sampling.
    - "vidcom2": Video Compression Commander.
    """
    mm_tensor_ipc: MMTensorIPC = "direct_rpc"
    """IPC (inter-process communication) method for multimodal tensors.
    - "direct_rpc": Use msgspec serialization via RPC
    - "torch_shm": Use torch.multiprocessing shared memory for zero-copy IPC
    Defaults to "direct_rpc". """
    allow_missing_mm_embeddings: bool = False
    """Whether a pre-computed-embedding input may omit the `*_embeds` tensor.

    In an encode/prefill/decode (EPD) deployment the encoder instance publishes
    embeddings through the EC connector. An EC consumer loads those embeddings
    from the connector, while a KV consumer receives the resulting prompt KV
    cache. Their requests only need the grid/size metadata that sizes the
    placeholder range.

    Derived, not user-settable: `VllmConfig.__post_init__` sets this to True
    on EC and KV consumers. Everywhere else it stays False so that a request
    which forgets its embeddings still fails fast in the frontend, with a clear
    error, rather than deep inside the model."""

    mm_ipc_gpu_memory_gb: float = Field(default=0, ge=0)
    """Amount of GPU memory (in GiB) sequestered on the engine's device for
    GPU-side multimodal work in the API-server (frontend) process, such as
    hardware video decoding.

    This budget is carved out of the engine's KV-cache memory so the headroom
    physically exists, and frontend GPU decode paths acquire from a blocking
    byte-counting semaphore of this size before allocating on the device.

    Set to `0` (default) to disable frontend GPU multimodal memory gating."""

    @field_validator("limit_per_prompt", mode="before")
    @classmethod
    def _validate_limit_per_prompt(
        cls,
        value: dict[str, int | dict[str, int]],
    ) -> MMDummyOptions:
        out: MMDummyOptions = {}

        for k, v in value.items():
            # Handle legacy format where only count is specified
            if isinstance(v, int):
                v = {"count": v}

            # Convert to the appropriate DummyOptions subclass
            if k == "video":
                out[k] = VideoDummyOptions(**v)
            elif k == "image":
                out[k] = ImageDummyOptions(**v)
            elif k == "audio":
                out[k] = AudioDummyOptions(**v)
            else:
                out[k] = BaseDummyOptions(**v)

        return out

    @field_validator("mm_encoder_attn_backend", mode="before")
    @classmethod
    def _validate_mm_encoder_attn_backend(
        cls, value: str | AttentionBackendEnum | None
    ) -> AttentionBackendEnum | None:
        if isinstance(value, str) and value.upper() == "XFORMERS":
            raise ValueError(
                "Attention backend 'XFORMERS' has been removed (See PR #29262 for "
                "details). Please select a supported attention backend."
            )

        if value is None or isinstance(value, AttentionBackendEnum):
            return value

        assert isinstance(value, str), (
            "mm_encoder_attn_backend must be a string or an AttentionBackendEnum."
        )
        return AttentionBackendEnum[value.upper()]

    @model_validator(mode="after")
    def _validate_multimodal_config(self):
        if self.mm_processor_cache_type != "shm" and (
            self.mm_shm_cache_max_object_size_mb
            != MultiModalConfig.mm_shm_cache_max_object_size_mb
        ):
            raise ValueError(
                "'mm_shm_cache_max_object_size_mb' should only be set when "
                "'mm_processor_cache_type' is 'shm'."
            )
        # Validate FP8 scale path combinations.
        if self.mm_encoder_attn_dtype != "fp8" and (
            self.mm_encoder_fp8_scale_path is not None
            or self.mm_encoder_fp8_scale_save_path is not None
        ):
            raise ValueError(
                "'mm_encoder_fp8_scale_path' and "
                "'mm_encoder_fp8_scale_save_path' require "
                "'mm_encoder_attn_dtype' to be 'fp8'."
            )
        if (
            self.mm_encoder_fp8_scale_path is not None
            and self.mm_encoder_fp8_scale_save_path is not None
        ):
            raise ValueError(
                "'mm_encoder_fp8_scale_save_path' cannot be used with "
                "'mm_encoder_fp8_scale_path' (saving requires dynamic scaling)."
            )

        # Validate file paths exist.
        if self.mm_encoder_fp8_scale_path is not None:
            scale_path = Path(self.mm_encoder_fp8_scale_path)
            if not scale_path.is_file():
                raise FileNotFoundError(f"FP8 scale file not found: {scale_path}")
        if self.mm_encoder_fp8_scale_save_path is not None:
            save_parent = Path(self.mm_encoder_fp8_scale_save_path).parent
            if not save_parent.is_dir():
                raise FileNotFoundError(
                    f"Parent directory for FP8 scale save path not found: {save_parent}"
                )
        return self

    @staticmethod
    def fold_mm_processor_device(
        mm_processor_kwargs: dict[str, Any] | None,
        mm_processor_device: MMProcessorDevice | None,
    ) -> dict[str, Any] | None:
        """Fold the `mm_processor_device` convenience flag into the kwargs.

        The flag keeps no state of its own: `mm_processor_kwargs["device"]` is
        the only representation of where the processor runs, so an explicit
        `device` there always wins and `"auto"` stays unresolved for
        `VllmConfig`, which is where the EC role needed to resolve it lives.

        Args:
            mm_processor_kwargs: The kwargs as given, or None.
            mm_processor_device: The flag's value, or None when unset.

        Returns:
            The kwargs to build the config with, unchanged unless the flag adds
            a `device`.
        """
        if mm_processor_device in (None, "auto"):
            return mm_processor_kwargs
        if (mm_processor_kwargs or {}).get("device") is not None:
            return mm_processor_kwargs

        from vllm.platforms import current_platform

        # Any explicit value other than "cpu" means "the accelerator", so a
        # programmatically-set "cuda" still works on a platform whose device type
        # is named differently ("xpu"), and degrades to CPU where there is none.
        device = (
            "cpu"
            if mm_processor_device == "cpu"
            else (current_platform.device_type or "cpu")
        )
        return {**(mm_processor_kwargs or {}), "device": device}

    def get_mm_processor_device_type(self) -> str | None:
        """The torch device type `mm_processor_kwargs["device"]` names.

        `mm_processor_kwargs` is untyped, so `device` may be any form torch
        accepts -- `"cuda"`, `"cuda:1"`, `torch.device(...)`, or a bare index.
        Normalising through torch rather than parsing the string keeps the
        non-string forms from slipping past a caller's comparison.

        Returns:
            The device type, or None when no device is requested.

        Raises:
            ValueError: If `device` is not something `torch.device` accepts.
                `validate_mm_processor_device` is what surfaces this during
                startup, so the value is only parsed once.
        """
        device = (self.mm_processor_kwargs or {}).get("device")
        if device is None:
            return None
        try:
            return torch.device(device).type  # type: ignore[arg-type]
        except (RuntimeError, TypeError, ValueError):
            raise ValueError(
                f'Invalid "device" in mm_processor_kwargs: {device!r}. Expected a '
                'torch device such as "cpu", "cuda" or "cuda:0".'
            ) from None

    def validate_mm_processor_device(self, ec_config: ECTransferConfig | None) -> None:
        """Check `mm_processor_kwargs["device"]` for this deployment.

        The only place the requested device is validated, so it runs even on a
        CPU-only platform: the value is parsed before any early return.

        Args:
            ec_config: The deployment's EC config, or None when it is not an
                encode/prefill/decode deployment. Passed in because it is not
                reachable from here, and because a field assigned after
                construction would not re-trigger this config's validators.

        Raises:
            ValueError: If the requested device is not a torch device, or if it
                is the accelerator on an instance that also runs the language
                model.
        """
        from vllm.platforms import current_platform

        device_type = self.get_mm_processor_device_type()
        accelerator = current_platform.device_type
        if device_type is None or accelerator in ("", "cpu"):
            return
        if device_type != accelerator:
            return

        if ec_config is None or not ec_config.is_encode_only:
            raise ValueError(
                f"Cannot run the multi-modal processor on {device_type!r}: this "
                "instance also runs the language model. The processor would "
                "share the device with the model's forward pass, so its "
                "transform kernels contend with that compute, and because it "
                "runs in the API-server process its allocations are outside the "
                "memory the engine profiled for its KV cache -- risking OOM or "
                "a silently shrunken cache.\n"
                "Accelerator preprocessing is only supported on an encode-only "
                "instance of an encode/prefill/decode deployment (an EC "
                "producer that is not also a consumer), which runs no forward "
                "pass and allocates no KV cache.\n"
                'Use --mm-processor-device=cpu, or drop "device" from '
                "--mm-processor-kwargs."
            )

        logger.info_once(
            "Running the multi-modal processor on %s. Override with "
            "--mm-processor-device=cpu.",
            device_type,
        )

    def compute_hash(self) -> str:
        """
        WARNING: Whenever a new field is added to this config,
        ensure that it is included in the factors list if
        it affects the computation graph.

        Provide a hash that uniquely identifies all the configs
        that affect the structure of the computation
        graph from input ids/embeddings to the final hidden states,
        excluding anything before input ids/embeddings and after
        the final hidden states.
        """
        factors: list[Any] = [
            self.mm_encoder_attn_backend.name
            if self.mm_encoder_attn_backend is not None
            else None,
            self.mm_encoder_tp_mode,
            self.mm_encoder_attn_dtype,
            self.mm_encoder_fp8_scale_path,
            self.mm_device_do_normalize,
        ]
        hash_str = safe_hash(str(factors).encode(), usedforsecurity=False).hexdigest()
        return hash_str

    def get_limit_per_prompt(self, modality: str) -> int:
        """
        Get the maximum number of input items allowed per prompt
        for the given modality (backward compatible).
        """
        if self.language_model_only:
            return 0

        limit_data = self.limit_per_prompt.get(modality)

        if limit_data is None:
            # Unspecified modality is set to 999 by default
            return 999

        return limit_data.count

    def merge_mm_processor_kwargs(
        self,
        inference_kwargs: Mapping[str, object],
    ) -> dict[str, object]:
        """
        Get the keyword arguments to pass to the multi-modal processor
        according to the extra arguments passed during inference.
        """
        kwargs = self.mm_processor_kwargs or {}
        if self.mm_device_do_normalize:
            kwargs["do_normalize"] = False
            kwargs["do_rescale"] = False
        return kwargs | dict(inference_kwargs)

    def use_gpu_video_backend(self) -> bool:
        """Return whether the configured video loader or codec uses the GPU."""
        from vllm.multimodal.video import VIDEO_LOADER_REGISTRY

        video_kwargs = self.media_io_kwargs.get("video", {})
        video_loader_backend = (
            video_kwargs.get("video_backend") or envs.VLLM_VIDEO_LOADER_BACKEND
        )
        codec_backend = video_kwargs.get("backend")
        return VIDEO_LOADER_REGISTRY.backend_requires_gpu(video_loader_backend) or (
            codec_backend is not None
            and VIDEO_LOADER_REGISTRY.backend_requires_gpu(codec_backend)
        )

    def is_multimodal_pruning_enabled(self):
        return self.get_video_pruning_spec() is not None

    def get_video_pruning_spec(self) -> tuple[VideoPruningMethod, float] | None:
        """Return `(method, rate)` when video pruning is enabled, else None.
        `rate` is the fraction of video tokens to prune."""
        if self.video_pruning_rate is not None and self.video_pruning_rate > 0:
            return (self.video_pruning_method, float(self.video_pruning_rate))
        return None

@final
class MultiModalDummyOptionsBuiltins(TypedDict, total=False):
    """Type annotations for modality types predefined by vLLM."""

    image: ImageDummyOptions
    """Options for dummy images."""

    video: VideoDummyOptions
    """Options for dummy videos."""

    audio: AudioDummyOptions
    """Options for dummy audios."""

@dataclass(config=ConfigDict(extra="forbid"))
class VideoDummyOptions(BaseDummyOptions):
    """Options for generating dummy video data during profiling."""

    num_frames: int | None = Field(None, gt=0)
    width: int | None = Field(None, gt=0)
    height: int | None = Field(None, gt=0)