vllm.model_executor.models.minimax_vl_01

MiniMaxVL01ImageInputs module-attribute

MiniMaxVL01ImageInputs = Union[
    MiniMaxVL01ImagePixelInputs,
    MiniMaxVL01ImageEmbeddingInputs,
]

MiniMaxVL01DummyInputsBuilder

Bases: LlavaDummyInputsBuilder

Source code in vllm/model_executor/models/minimax_vl_01.py

class MiniMaxVL01DummyInputsBuilder(LlavaDummyInputsBuilder):
    pass

MiniMaxVL01ForConditionalGeneration

Bases: Module, SupportsMultiModal, SupportsPP

Source code in vllm/model_executor/models/minimax_vl_01.py

@MULTIMODAL_REGISTRY.register_processor(
    MiniMaxVL01MultiModalProcessor,
    info=MiniMaxVL01ProcessingInfo,
    dummy_inputs=MiniMaxVL01DummyInputsBuilder)
class MiniMaxVL01ForConditionalGeneration(nn.Module, SupportsMultiModal,
                                          SupportsPP):

    packed_modules_mapping = {
        "qkv_proj": ["q_proj", "k_proj", "v_proj"],
        "gate_up_proj": ["gate_proj", "up_proj"]
    }

    @classmethod
    def get_placeholder_str(cls, modality: str, i: int) -> Optional[str]:
        if modality.startswith("image"):
            return "<image>"

        raise ValueError("Only image modality is supported")

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None:
        super().__init__()

        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
        multimodal_config = vllm_config.model_config.multimodal_config

        self.config = config
        self.multimodal_config = multimodal_config

        # TODO: Optionally initializes this for supporting embeddings.
        self.vision_tower = init_vision_tower_for_llava(
            config,
            quant_config,
            require_post_norm=False,
            prefix=maybe_prefix(prefix, "vision_tower"))
        self.multi_modal_projector = MiniMaxVL01MultiModalProjector(
            vision_hidden_size=config.vision_config.hidden_size,
            text_hidden_size=config.text_config.hidden_size,
            projector_hidden_act=config.projector_hidden_act,
            multimodal_projector_bias=True,
            quant_config=quant_config,
            prefix=maybe_prefix(prefix, "multi_modal_projector"))
        self.image_newline = nn.Parameter(
            torch.empty(config.text_config.hidden_size))
        self.language_model = init_vllm_registered_model(
            vllm_config=vllm_config,
            hf_config=config.text_config,
            prefix=maybe_prefix(prefix, "language_model"),
        )
        self.vision_feature_layer = config.vision_feature_layer
        self.vocab_size = config.text_config.vocab_size
        self.pad_token_id = -1
        if self.config.pad_token_id is not None:
            self.pad_token_id = self.config.pad_token_id

        self.make_empty_intermediate_tensors = (
            self.language_model.make_empty_intermediate_tensors)

    def get_input_embeddings(
        self,
        input_ids: torch.Tensor,
        multimodal_embeddings: Optional[MultiModalEmbeddings] = None,
    ) -> torch.Tensor:
        inputs_embeds = self.language_model.get_input_embeddings(input_ids)
        if multimodal_embeddings is not None \
            and len(multimodal_embeddings) != 0:
            inputs_embeds = merge_multimodal_embeddings(
                input_ids,
                inputs_embeds,
                multimodal_embeddings,
                self.config.image_token_index,
            )
        return inputs_embeds

    def get_language_model(self) -> torch.nn.Module:
        return self.language_model

    def _select_image_features(self, image_features: torch.Tensor, *,
                               strategy: str) -> torch.Tensor:
        if strategy == "default":
            return image_features[:, 1:]
        elif strategy == "full":
            return image_features

        raise ValueError(f"Unexpected select feature strategy: {strategy}")

    def _image_pixels_to_features(
        self,
        vision_tower: Union[CLIPVisionModel, SiglipVisionModel,
                            PixtralHFVisionModel],
        pixel_values: Union[torch.Tensor, list[torch.Tensor]],
    ) -> Union[torch.Tensor, tuple[torch.Tensor, ...]]:
        # NOTE: we skip the step to select the vision feature layer since
        # this is already done inside the vision tower
        image_features = vision_tower(pixel_values)

        def select_features(leaf: torch.Tensor):
            return self._select_image_features(
                leaf,
                strategy=self.config.vision_feature_select_strategy,
            )

        return cast(
            Union[torch.Tensor, tuple[torch.Tensor, ...]],
            json_map_leaves(select_features, image_features),
        )

    def _process_image_pixels(
        self,
        inputs: MiniMaxVL01ImagePixelInputs,
    ) -> Union[torch.Tensor, tuple[torch.Tensor, ...]]:
        assert self.vision_tower is not None

        pixel_values = inputs["pixel_values"]

        return self._image_pixels_to_features(self.vision_tower, pixel_values)

    def _process_image_input(
        self,
        image_input: MiniMaxVL01ImageInputs,
    ) -> Union[torch.Tensor, tuple[torch.Tensor, ...]]:
        if image_input["type"] == "image_embeds":
            return image_input["data"]

        assert self.vision_tower is not None
        image_features = self._process_image_pixels(image_input)

        if isinstance(image_features, torch.Tensor):
            return self.multi_modal_projector(image_features)

        feature_sizes = [
            image_feature.shape[0] for image_feature in image_features
        ]

        image_embeds = self.multi_modal_projector(torch.cat(image_features))
        image_embeds = torch.split(image_embeds, feature_sizes)
        return image_embeds

    def _validate_pixel_values(self, data: torch.Tensor) -> torch.Tensor:
        h = w = self.config.vision_config.image_size
        expected_dims = (3, h, w)
        actual_dims = tuple(data.shape[1:])

        if actual_dims != expected_dims:
            expected_expr = ("batch_size", *map(str, expected_dims))
            raise ValueError(
                f"The expected shape of pixel values is {expected_expr}. "
                f"You supplied {tuple(data.shape)}.")

        return data

    def _parse_and_validate_image_input(
            self, **kwargs: object) -> Optional[MiniMaxVL01ImageInputs]:
        pixel_values = kwargs.pop("pixel_values", None)
        image_embeds = kwargs.pop("image_embeds", None)

        if pixel_values is None and image_embeds is None:
            return None

        if pixel_values is not None:
            if not isinstance(pixel_values, (torch.Tensor, list)):
                raise ValueError("Incorrect type of pixel values. "
                                 f"Got type: {type(pixel_values)}")

            return MiniMaxVL01ImagePixelInputs(
                type="pixel_values",
                pixel_values=self._validate_pixel_values(
                    flatten_bn(pixel_values, concat=True)),
            )

        if image_embeds is not None:
            if not isinstance(image_embeds, (torch.Tensor, list)):
                raise ValueError("Incorrect type of image embeddings. "
                                 f"Got type: {type(image_embeds)}")

            return MiniMaxVL01ImageEmbeddingInputs(
                type="image_embeds",
                data=flatten_bn(image_embeds, concat=True),
            )

        raise AssertionError("This line should be unreachable.")

    def get_multimodal_embeddings(self,
                                  **kwargs: object) -> MultiModalEmbeddings:
        image_input = self._parse_and_validate_image_input(**kwargs)
        if image_input is None:
            return []

        return self._process_image_input(image_input)

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        intermediate_tensors: Optional[IntermediateTensors] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
        **kwargs: object,
    ) -> Union[torch.Tensor, IntermediateTensors]:

        if intermediate_tensors is not None:
            inputs_embeds = None
        elif inputs_embeds is None:
            vision_embeddings = self.get_multimodal_embeddings(**kwargs)
            inputs_embeds = self.get_input_embeddings(input_ids,
                                                      vision_embeddings)
            input_ids = None

        hidden_states = self.language_model.model(input_ids,
                                                  positions,
                                                  intermediate_tensors,
                                                  inputs_embeds=inputs_embeds)

        return hidden_states

    def compute_logits(
        self,
        hidden_states: torch.Tensor,
        sampling_metadata: SamplingMetadata,
    ) -> Optional[torch.Tensor]:
        return self.language_model.compute_logits(hidden_states,
                                                  sampling_metadata)

    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
        loader = AutoWeightsLoader(self)
        return loader.load_weights(weights)

config instance-attribute

config = config

image_newline instance-attribute

image_newline = Parameter(empty(hidden_size))

language_model instance-attribute

language_model = init_vllm_registered_model(
    vllm_config=vllm_config,
    hf_config=text_config,
    prefix=maybe_prefix(prefix, "language_model"),
)

make_empty_intermediate_tensors instance-attribute

make_empty_intermediate_tensors = (
    make_empty_intermediate_tensors
)

multi_modal_projector instance-attribute

multi_modal_projector = MiniMaxVL01MultiModalProjector(
    vision_hidden_size=hidden_size,
    text_hidden_size=hidden_size,
    projector_hidden_act=projector_hidden_act,
    multimodal_projector_bias=True,
    quant_config=quant_config,
    prefix=maybe_prefix(prefix, "multi_modal_projector"),
)

multimodal_config instance-attribute

multimodal_config = multimodal_config

packed_modules_mapping class-attribute instance-attribute

packed_modules_mapping = {
    "qkv_proj": ["q_proj", "k_proj", "v_proj"],
    "gate_up_proj": ["gate_proj", "up_proj"],
}

pad_token_id instance-attribute

pad_token_id = -1

vision_feature_layer instance-attribute

vision_feature_layer = vision_feature_layer

vision_tower instance-attribute

vision_tower = init_vision_tower_for_llava(
    config,
    quant_config,
    require_post_norm=False,
    prefix=maybe_prefix(prefix, "vision_tower"),
)

vocab_size instance-attribute

vocab_size = vocab_size

__init__

__init__(
    *, vllm_config: VllmConfig, prefix: str = ""
) -> None

Source code in vllm/model_executor/models/minimax_vl_01.py

def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None:
    super().__init__()

    config = vllm_config.model_config.hf_config
    quant_config = vllm_config.quant_config
    multimodal_config = vllm_config.model_config.multimodal_config

    self.config = config
    self.multimodal_config = multimodal_config

    # TODO: Optionally initializes this for supporting embeddings.
    self.vision_tower = init_vision_tower_for_llava(
        config,
        quant_config,
        require_post_norm=False,
        prefix=maybe_prefix(prefix, "vision_tower"))
    self.multi_modal_projector = MiniMaxVL01MultiModalProjector(
        vision_hidden_size=config.vision_config.hidden_size,
        text_hidden_size=config.text_config.hidden_size,
        projector_hidden_act=config.projector_hidden_act,
        multimodal_projector_bias=True,
        quant_config=quant_config,
        prefix=maybe_prefix(prefix, "multi_modal_projector"))
    self.image_newline = nn.Parameter(
        torch.empty(config.text_config.hidden_size))
    self.language_model = init_vllm_registered_model(
        vllm_config=vllm_config,
        hf_config=config.text_config,
        prefix=maybe_prefix(prefix, "language_model"),
    )
    self.vision_feature_layer = config.vision_feature_layer
    self.vocab_size = config.text_config.vocab_size
    self.pad_token_id = -1
    if self.config.pad_token_id is not None:
        self.pad_token_id = self.config.pad_token_id

    self.make_empty_intermediate_tensors = (
        self.language_model.make_empty_intermediate_tensors)

_image_pixels_to_features

_image_pixels_to_features(
    vision_tower: Union[
        CLIPVisionModel,
        SiglipVisionModel,
        PixtralHFVisionModel,
    ],
    pixel_values: Union[Tensor, list[Tensor]],
) -> Union[Tensor, tuple[Tensor, ...]]

Source code in vllm/model_executor/models/minimax_vl_01.py

def _image_pixels_to_features(
    self,
    vision_tower: Union[CLIPVisionModel, SiglipVisionModel,
                        PixtralHFVisionModel],
    pixel_values: Union[torch.Tensor, list[torch.Tensor]],
) -> Union[torch.Tensor, tuple[torch.Tensor, ...]]:
    # NOTE: we skip the step to select the vision feature layer since
    # this is already done inside the vision tower
    image_features = vision_tower(pixel_values)

    def select_features(leaf: torch.Tensor):
        return self._select_image_features(
            leaf,
            strategy=self.config.vision_feature_select_strategy,
        )

    return cast(
        Union[torch.Tensor, tuple[torch.Tensor, ...]],
        json_map_leaves(select_features, image_features),
    )

_parse_and_validate_image_input

_parse_and_validate_image_input(
    **kwargs: object,
) -> Optional[MiniMaxVL01ImageInputs]

Source code in vllm/model_executor/models/minimax_vl_01.py

def _parse_and_validate_image_input(
        self, **kwargs: object) -> Optional[MiniMaxVL01ImageInputs]:
    pixel_values = kwargs.pop("pixel_values", None)
    image_embeds = kwargs.pop("image_embeds", None)

    if pixel_values is None and image_embeds is None:
        return None

    if pixel_values is not None:
        if not isinstance(pixel_values, (torch.Tensor, list)):
            raise ValueError("Incorrect type of pixel values. "
                             f"Got type: {type(pixel_values)}")

        return MiniMaxVL01ImagePixelInputs(
            type="pixel_values",
            pixel_values=self._validate_pixel_values(
                flatten_bn(pixel_values, concat=True)),
        )

    if image_embeds is not None:
        if not isinstance(image_embeds, (torch.Tensor, list)):
            raise ValueError("Incorrect type of image embeddings. "
                             f"Got type: {type(image_embeds)}")

        return MiniMaxVL01ImageEmbeddingInputs(
            type="image_embeds",
            data=flatten_bn(image_embeds, concat=True),
        )

    raise AssertionError("This line should be unreachable.")

_process_image_input

_process_image_input(
    image_input: MiniMaxVL01ImageInputs,
) -> Union[Tensor, tuple[Tensor, ...]]

Source code in vllm/model_executor/models/minimax_vl_01.py

def _process_image_input(
    self,
    image_input: MiniMaxVL01ImageInputs,
) -> Union[torch.Tensor, tuple[torch.Tensor, ...]]:
    if image_input["type"] == "image_embeds":
        return image_input["data"]

    assert self.vision_tower is not None
    image_features = self._process_image_pixels(image_input)

    if isinstance(image_features, torch.Tensor):
        return self.multi_modal_projector(image_features)

    feature_sizes = [
        image_feature.shape[0] for image_feature in image_features
    ]

    image_embeds = self.multi_modal_projector(torch.cat(image_features))
    image_embeds = torch.split(image_embeds, feature_sizes)
    return image_embeds

_process_image_pixels

_process_image_pixels(
    inputs: MiniMaxVL01ImagePixelInputs,
) -> Union[Tensor, tuple[Tensor, ...]]

Source code in vllm/model_executor/models/minimax_vl_01.py

def _process_image_pixels(
    self,
    inputs: MiniMaxVL01ImagePixelInputs,
) -> Union[torch.Tensor, tuple[torch.Tensor, ...]]:
    assert self.vision_tower is not None

    pixel_values = inputs["pixel_values"]

    return self._image_pixels_to_features(self.vision_tower, pixel_values)

_select_image_features

_select_image_features(
    image_features: Tensor, *, strategy: str
) -> Tensor

Source code in vllm/model_executor/models/minimax_vl_01.py

def _select_image_features(self, image_features: torch.Tensor, *,
                           strategy: str) -> torch.Tensor:
    if strategy == "default":
        return image_features[:, 1:]
    elif strategy == "full":
        return image_features

    raise ValueError(f"Unexpected select feature strategy: {strategy}")

_validate_pixel_values

_validate_pixel_values(data: Tensor) -> Tensor

Source code in vllm/model_executor/models/minimax_vl_01.py

def _validate_pixel_values(self, data: torch.Tensor) -> torch.Tensor:
    h = w = self.config.vision_config.image_size
    expected_dims = (3, h, w)
    actual_dims = tuple(data.shape[1:])

    if actual_dims != expected_dims:
        expected_expr = ("batch_size", *map(str, expected_dims))
        raise ValueError(
            f"The expected shape of pixel values is {expected_expr}. "
            f"You supplied {tuple(data.shape)}.")

    return data

compute_logits

compute_logits(
    hidden_states: Tensor,
    sampling_metadata: SamplingMetadata,
) -> Optional[Tensor]

Source code in vllm/model_executor/models/minimax_vl_01.py

def compute_logits(
    self,
    hidden_states: torch.Tensor,
    sampling_metadata: SamplingMetadata,
) -> Optional[torch.Tensor]:
    return self.language_model.compute_logits(hidden_states,
                                              sampling_metadata)

forward

forward(
    input_ids: Tensor,
    positions: Tensor,
    intermediate_tensors: Optional[
        IntermediateTensors
    ] = None,
    inputs_embeds: Optional[Tensor] = None,
    **kwargs: object,
) -> Union[Tensor, IntermediateTensors]

Source code in vllm/model_executor/models/minimax_vl_01.py

def forward(
    self,
    input_ids: torch.Tensor,
    positions: torch.Tensor,
    intermediate_tensors: Optional[IntermediateTensors] = None,
    inputs_embeds: Optional[torch.Tensor] = None,
    **kwargs: object,
) -> Union[torch.Tensor, IntermediateTensors]:

    if intermediate_tensors is not None:
        inputs_embeds = None
    elif inputs_embeds is None:
        vision_embeddings = self.get_multimodal_embeddings(**kwargs)
        inputs_embeds = self.get_input_embeddings(input_ids,
                                                  vision_embeddings)
        input_ids = None

    hidden_states = self.language_model.model(input_ids,
                                              positions,
                                              intermediate_tensors,
                                              inputs_embeds=inputs_embeds)

    return hidden_states

get_input_embeddings

get_input_embeddings(
    input_ids: Tensor,
    multimodal_embeddings: Optional[
        MultiModalEmbeddings
    ] = None,
) -> Tensor

Source code in vllm/model_executor/models/minimax_vl_01.py

def get_input_embeddings(
    self,
    input_ids: torch.Tensor,
    multimodal_embeddings: Optional[MultiModalEmbeddings] = None,
) -> torch.Tensor:
    inputs_embeds = self.language_model.get_input_embeddings(input_ids)
    if multimodal_embeddings is not None \
        and len(multimodal_embeddings) != 0:
        inputs_embeds = merge_multimodal_embeddings(
            input_ids,
            inputs_embeds,
            multimodal_embeddings,
            self.config.image_token_index,
        )
    return inputs_embeds

get_language_model

get_language_model() -> Module

Source code in vllm/model_executor/models/minimax_vl_01.py

def get_language_model(self) -> torch.nn.Module:
    return self.language_model

get_multimodal_embeddings

get_multimodal_embeddings(
    **kwargs: object,
) -> MultiModalEmbeddings

Source code in vllm/model_executor/models/minimax_vl_01.py

def get_multimodal_embeddings(self,
                              **kwargs: object) -> MultiModalEmbeddings:
    image_input = self._parse_and_validate_image_input(**kwargs)
    if image_input is None:
        return []

    return self._process_image_input(image_input)

get_placeholder_str classmethod

get_placeholder_str(modality: str, i: int) -> Optional[str]

Source code in vllm/model_executor/models/minimax_vl_01.py

@classmethod
def get_placeholder_str(cls, modality: str, i: int) -> Optional[str]:
    if modality.startswith("image"):
        return "<image>"

    raise ValueError("Only image modality is supported")

load_weights

load_weights(
    weights: Iterable[tuple[str, Tensor]],
) -> set[str]

Source code in vllm/model_executor/models/minimax_vl_01.py

def load_weights(self, weights: Iterable[tuple[str,
                                               torch.Tensor]]) -> set[str]:
    loader = AutoWeightsLoader(self)
    return loader.load_weights(weights)

MiniMaxVL01ImageEmbeddingInputs

Bases: TypedDict

Source code in vllm/model_executor/models/minimax_vl_01.py

class MiniMaxVL01ImageEmbeddingInputs(TypedDict):
    type: Literal["image_embeds"]
    data: torch.Tensor
    """Shape: `(batch_size * num_images, image_feature_size, hidden_size)`

    `hidden_size` must match the hidden size of language model backbone.
    """

data instance-attribute

data: Tensor

Shape: (batch_size * num_images, image_feature_size, hidden_size)

hidden_size must match the hidden size of language model backbone.

type instance-attribute

type: Literal['image_embeds']

MiniMaxVL01ImagePixelInputs

Bases: TypedDict

Source code in vllm/model_executor/models/minimax_vl_01.py

class MiniMaxVL01ImagePixelInputs(TypedDict):
    type: Literal["pixel_values"]
    pixel_values: torch.Tensor
    """
    Shape: `(batch_size * num_images, num_channels, height, width)`

    Note that `height` or `width` may be different per batch and image,
    in which case the data is passed as a list instead of a batched tensor.
    """

pixel_values instance-attribute

pixel_values: Tensor

Shape: (batch_size * num_images, num_channels, height, width)

Note that height or width may be different per batch and image, in which case the data is passed as a list instead of a batched tensor.

type instance-attribute

type: Literal['pixel_values']

MiniMaxVL01MultiModalProcessor

Bases: BaseLlavaMultiModalProcessor[MiniMaxVL01ProcessingInfo]

Source code in vllm/model_executor/models/minimax_vl_01.py

class MiniMaxVL01MultiModalProcessor(
        BaseLlavaMultiModalProcessor[MiniMaxVL01ProcessingInfo]):

    def _call_hf_processor(
        self,
        prompt: str,
        mm_data: Mapping[str, object],
        mm_kwargs: Mapping[str, object],
        tok_kwargs: Mapping[str, object],
    ) -> BatchFeature:
        processed_outputs = super()._call_hf_processor(
            prompt=prompt,
            mm_data=mm_data,
            mm_kwargs=mm_kwargs,
            tok_kwargs=tok_kwargs,
        )

        pixel_values = processed_outputs.get("pixel_values")
        if pixel_values is not None:
            # Avoid padding since we need the output for each image to be
            # independent of other images for the cache to work correctly
            image_sizes = processed_outputs["image_sizes"]
            assert len(pixel_values) == len(image_sizes)

            processed_outputs["pixel_values"] = [
                p[:, :h, :w] for p, (h, w) in zip(pixel_values, image_sizes)
            ]

        return processed_outputs

    def _get_mm_fields_config(
        self,
        hf_inputs: BatchFeature,
        hf_processor_mm_kwargs: Mapping[str, object],
    ) -> Mapping[str, MultiModalFieldConfig]:
        return {
            "pixel_values": MultiModalFieldConfig.batched("image"),
            "image_embeds": MultiModalFieldConfig.batched("image"),
        }

_call_hf_processor

_call_hf_processor(
    prompt: str,
    mm_data: Mapping[str, object],
    mm_kwargs: Mapping[str, object],
    tok_kwargs: Mapping[str, object],
) -> BatchFeature

Source code in vllm/model_executor/models/minimax_vl_01.py

def _call_hf_processor(
    self,
    prompt: str,
    mm_data: Mapping[str, object],
    mm_kwargs: Mapping[str, object],
    tok_kwargs: Mapping[str, object],
) -> BatchFeature:
    processed_outputs = super()._call_hf_processor(
        prompt=prompt,
        mm_data=mm_data,
        mm_kwargs=mm_kwargs,
        tok_kwargs=tok_kwargs,
    )

    pixel_values = processed_outputs.get("pixel_values")
    if pixel_values is not None:
        # Avoid padding since we need the output for each image to be
        # independent of other images for the cache to work correctly
        image_sizes = processed_outputs["image_sizes"]
        assert len(pixel_values) == len(image_sizes)

        processed_outputs["pixel_values"] = [
            p[:, :h, :w] for p, (h, w) in zip(pixel_values, image_sizes)
        ]

    return processed_outputs

_get_mm_fields_config

_get_mm_fields_config(
    hf_inputs: BatchFeature,
    hf_processor_mm_kwargs: Mapping[str, object],
) -> Mapping[str, MultiModalFieldConfig]

Source code in vllm/model_executor/models/minimax_vl_01.py

def _get_mm_fields_config(
    self,
    hf_inputs: BatchFeature,
    hf_processor_mm_kwargs: Mapping[str, object],
) -> Mapping[str, MultiModalFieldConfig]:
    return {
        "pixel_values": MultiModalFieldConfig.batched("image"),
        "image_embeds": MultiModalFieldConfig.batched("image"),
    }

MiniMaxVL01MultiModalProjector

Bases: Module

Source code in vllm/model_executor/models/minimax_vl_01.py

class MiniMaxVL01MultiModalProjector(nn.Module):

    def __init__(self,
                 vision_hidden_size: int,
                 text_hidden_size: int,
                 projector_hidden_act: str,
                 multimodal_projector_bias: bool,
                 quant_config: Optional[QuantizationConfig] = None,
                 prefix: str = ""):
        super().__init__()

        self.linear_1 = ColumnParallelLinear(vision_hidden_size,
                                             text_hidden_size,
                                             bias=multimodal_projector_bias,
                                             quant_config=quant_config,
                                             prefix=f"{prefix}.linear_1")
        self.act = get_act_fn(projector_hidden_act)
        self.linear_2 = RowParallelLinear(text_hidden_size,
                                          text_hidden_size,
                                          bias=multimodal_projector_bias,
                                          quant_config=quant_config,
                                          prefix=f"{prefix}.linear_2")

    def forward(self, image_features: torch.Tensor) -> torch.Tensor:
        hidden_states, _ = self.linear_1(image_features)
        hidden_states = self.act(hidden_states)
        hidden_states, _ = self.linear_2(hidden_states)
        return hidden_states

act instance-attribute

act = get_act_fn(projector_hidden_act)

linear_1 instance-attribute

linear_1 = ColumnParallelLinear(
    vision_hidden_size,
    text_hidden_size,
    bias=multimodal_projector_bias,
    quant_config=quant_config,
    prefix=f"{prefix}.linear_1",
)

linear_2 instance-attribute

linear_2 = RowParallelLinear(
    text_hidden_size,
    text_hidden_size,
    bias=multimodal_projector_bias,
    quant_config=quant_config,
    prefix=f"{prefix}.linear_2",
)

__init__

__init__(
    vision_hidden_size: int,
    text_hidden_size: int,
    projector_hidden_act: str,
    multimodal_projector_bias: bool,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
)

Source code in vllm/model_executor/models/minimax_vl_01.py

def __init__(self,
             vision_hidden_size: int,
             text_hidden_size: int,
             projector_hidden_act: str,
             multimodal_projector_bias: bool,
             quant_config: Optional[QuantizationConfig] = None,
             prefix: str = ""):
    super().__init__()

    self.linear_1 = ColumnParallelLinear(vision_hidden_size,
                                         text_hidden_size,
                                         bias=multimodal_projector_bias,
                                         quant_config=quant_config,
                                         prefix=f"{prefix}.linear_1")
    self.act = get_act_fn(projector_hidden_act)
    self.linear_2 = RowParallelLinear(text_hidden_size,
                                      text_hidden_size,
                                      bias=multimodal_projector_bias,
                                      quant_config=quant_config,
                                      prefix=f"{prefix}.linear_2")

forward

forward(image_features: Tensor) -> Tensor

Source code in vllm/model_executor/models/minimax_vl_01.py

def forward(self, image_features: torch.Tensor) -> torch.Tensor:
    hidden_states, _ = self.linear_1(image_features)
    hidden_states = self.act(hidden_states)
    hidden_states, _ = self.linear_2(hidden_states)
    return hidden_states

MiniMaxVL01ProcessingInfo

Bases: LlavaNextProcessingInfo

Source code in vllm/model_executor/models/minimax_vl_01.py

class MiniMaxVL01ProcessingInfo(LlavaNextProcessingInfo):

    def get_hf_config(self):
        return self.ctx.get_hf_config(MiniMaxVL01Config)

    def get_hf_processor(self, **kwargs: object):
        hf_processor = self.ctx.get_hf_processor(**kwargs)
        image_processor = hf_processor.image_processor
        image_processor.anyres_preprocess = (
            image_processor.anyres_for_vllm_preprocess)

        return hf_processor

    def get_supported_mm_limits(self) -> Mapping[str, Optional[int]]:
        return {"image": None}

get_hf_config

get_hf_config()

Source code in vllm/model_executor/models/minimax_vl_01.py

def get_hf_config(self):
    return self.ctx.get_hf_config(MiniMaxVL01Config)

get_hf_processor

get_hf_processor(**kwargs: object)

Source code in vllm/model_executor/models/minimax_vl_01.py

def get_hf_processor(self, **kwargs: object):
    hf_processor = self.ctx.get_hf_processor(**kwargs)
    image_processor = hf_processor.image_processor
    image_processor.anyres_preprocess = (
        image_processor.anyres_for_vllm_preprocess)

    return hf_processor

get_supported_mm_limits

get_supported_mm_limits() -> Mapping[str, Optional[int]]

Source code in vllm/model_executor/models/minimax_vl_01.py

def get_supported_mm_limits(self) -> Mapping[str, Optional[int]]:
    return {"image": None}