mistral3.py

# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project

from collections.abc import Iterable, Mapping, Sequence
from typing import Annotated, Literal

import torch
import torch.nn as nn
from transformers import BatchFeature, Mistral3Config, PixtralVisionConfig
from transformers.models.pixtral import PixtralProcessor

from vllm.config import VllmConfig
from vllm.config.multimodal import BaseDummyOptions
from vllm.inputs import MultiModalDataDict
from vllm.model_executor.layers.activation import get_act_fn
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import ColumnParallelLinear, RowParallelLinear
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.models.module_mapping import MultiModelKeys
from vllm.multimodal import MULTIMODAL_REGISTRY
from vllm.multimodal.inputs import (
    MultiModalFieldConfig,
    MultiModalKwargsItems,
)
from vllm.multimodal.parse import ImageProcessorItems, ImageSize, MultiModalDataItems
from vllm.multimodal.processing import (
    BaseDummyInputsBuilder,
    BaseMultiModalProcessor,
    BaseProcessingInfo,
    PromptReplacement,
    PromptUpdate,
    PromptUpdateDetails,
)
from vllm.sequence import IntermediateTensors
from vllm.utils.tensor_schema import TensorSchema, TensorShape

from .interfaces import (
    MultiModalEmbeddings,
    SupportsEagle,
    SupportsEagle3,
    SupportsLoRA,
    SupportsMultiModal,
    SupportsPP,
)
from .pixtral import PixtralHFEncoderInfo, PixtralHFVisionModel
from .utils import (
    AutoWeightsLoader,
    WeightsMapper,
    get_layer_index,
    init_vllm_registered_model,
    maybe_prefix,
)
from .vision import get_vision_encoder_info


class Mistral3ImagePixelInputs(TensorSchema):
    """
    Dimensions:
        - bn: Batch size * number of images
        - c: Number of channels (3)
        - h: Height of each image
        - w: Width of each image
    """

    type: Literal["pixel_values_pixtral"] = "pixel_values_pixtral"

    # 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: Annotated[
        torch.Tensor | list[torch.Tensor],
        TensorShape("bn", 3, "h", "w", dynamic_dims={"h", "w"}),
    ]


class Mistral3PatchMerger(nn.Module):
    """
    Learned merging of spatial_merge_size ** 2 patches
    """

    def __init__(
        self, vision_hidden_size: int, spatial_merge_size: int, patch_size: int
    ):
        super().__init__()

        self.vision_hidden_size = vision_hidden_size
        self.spatial_merge_size = spatial_merge_size
        self.patch_size = patch_size
        self.merging_layer = nn.Linear(
            vision_hidden_size * self.spatial_merge_size**2,
            vision_hidden_size,
            bias=False,
        )

    def forward(
        self, image_features: torch.Tensor, image_sizes: torch.Tensor
    ) -> torch.Tensor:
        image_sizes = [
            (image_size[0] // self.patch_size, image_size[1] // self.patch_size)
            for image_size in image_sizes
        ]

        tokens_per_image = [h * w for h, w in image_sizes]
        d = image_features.shape[-1]

        permuted_tensor = []
        for image_index, image_tokens in enumerate(
            image_features.split(tokens_per_image)
        ):
            # Reshape image_tokens into a 2D grid
            h, w = image_sizes[image_index]
            image_grid = image_tokens.view(h, w, d).permute(2, 0, 1).unsqueeze(0)
            grid = torch.nn.functional.unfold(
                image_grid,
                kernel_size=self.spatial_merge_size,
                stride=self.spatial_merge_size,
            )
            grid = grid.view(d * self.spatial_merge_size**2, -1).t()
            permuted_tensor.append(grid)

        image_features = torch.cat(permuted_tensor, dim=0)
        image_features = self.merging_layer(image_features)
        return image_features


class Mistral3MultiModalProjector(nn.Module):
    def __init__(
        self,
        vision_hidden_size: int,
        text_hidden_size: int,
        spatial_merge_size: int,
        patch_size: int,
        projector_hidden_act: str,
        multimodal_projector_bias: bool,
        quant_config: QuantizationConfig | None = None,
        prefix: str = "",
    ):
        super().__init__()

        self.norm = RMSNorm(vision_hidden_size, eps=1e-5)
        self.patch_merger = Mistral3PatchMerger(
            vision_hidden_size=vision_hidden_size,
            spatial_merge_size=spatial_merge_size,
            patch_size=patch_size,
        )

        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, image_sizes: torch.Tensor
    ) -> torch.Tensor:
        image_features = self.norm(image_features)
        image_features = self.patch_merger(image_features, image_sizes)
        hidden_states, _ = self.linear_1(image_features)
        hidden_states = self.act(hidden_states)
        hidden_states, _ = self.linear_2(hidden_states)
        return hidden_states


class Mistral3ProcessingInfo(BaseProcessingInfo):
    def get_hf_config(self) -> Mistral3Config:
        return self.ctx.get_hf_config(Mistral3Config)

    def get_vision_encoder_info(self):
        return get_vision_encoder_info(self.get_hf_config())

    def get_hf_processor(self, **kwargs: object):
        return self.ctx.get_hf_processor(PixtralProcessor, **kwargs)

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

    def get_num_image_tokens(
        self,
        *,
        image_width: int,
        image_height: int,
    ) -> int:
        vision_encoder_info = self.get_vision_encoder_info()
        return vision_encoder_info.get_num_image_tokens(
            image_width=image_width,
            image_height=image_height,
        )

    def get_image_size_with_most_features(self) -> ImageSize:
        vision_encoder_info = self.get_vision_encoder_info()
        width = height = vision_encoder_info.get_image_size()
        return ImageSize(width=width, height=height)


class Mistral3DummyInputsBuilder(BaseDummyInputsBuilder[Mistral3ProcessingInfo]):
    def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
        num_images = mm_counts.get("image", 0)

        processor = self.info.get_hf_processor()
        image_token = processor.image_token

        return image_token * num_images

    def get_dummy_mm_data(
        self,
        seq_len: int,
        mm_counts: Mapping[str, int],
        mm_options: Mapping[str, BaseDummyOptions],
    ) -> MultiModalDataDict:
        num_images = mm_counts.get("image", 0)

        target_width, target_height = self.info.get_image_size_with_most_features()

        image_overrides = mm_options.get("image")

        return {
            "image": self._get_dummy_images(
                width=target_width,
                height=target_height,
                num_images=num_images,
                overrides=image_overrides,
            )
        }


class Mistral3MultiModalProcessor(BaseMultiModalProcessor[Mistral3ProcessingInfo]):
    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 dict(
            pixel_values=MultiModalFieldConfig.batched("image"),
            image_embeds=MultiModalFieldConfig.batched("image"),
        )

    def _get_prompt_updates(
        self,
        mm_items: MultiModalDataItems,
        hf_processor_mm_kwargs: Mapping[str, object],
        out_mm_kwargs: MultiModalKwargsItems,
    ) -> Sequence[PromptUpdate]:
        processor = self.info.get_hf_processor(**hf_processor_mm_kwargs)
        hf_config = self.info.get_hf_config()
        tokenizer = self.info.get_tokenizer()
        vocab = tokenizer.get_vocab()

        image_break_id = vocab[processor.image_break_token]
        image_token_id = hf_config.image_token_index
        image_end_id = vocab[processor.image_end_token]

        assert isinstance(hf_config.vision_config, PixtralVisionConfig)
        encoder_info = PixtralHFEncoderInfo(hf_config)

        def get_replacement(item_idx: int):
            images = mm_items.get_items("image", ImageProcessorItems)
            image_size = images.get_image_size(item_idx)

            ncols, nrows = encoder_info.get_patch_grid_size(
                image_width=image_size.width,
                image_height=image_size.height,
            )

            tokens = ([image_token_id] * ncols + [image_break_id]) * nrows
            tokens[-1] = image_end_id

            return PromptUpdateDetails.select_token_id(tokens, image_token_id)

        return [
            PromptReplacement(
                modality="image",
                target=[image_token_id],
                replacement=get_replacement,
            ),
        ]


def _get_num_hidden_layers(hf_config: Mistral3Config) -> int:
    """Determine the number of hidden layers to initialize up to in the
    visual encoder.

    Args:
        hf_config: Model config with vision feature layer(s).
    """
    feature_layers = hf_config.vision_feature_layer
    num_hidden_layers = hf_config.vision_config.num_hidden_layers
    # If we have one feature layer, initialize up to that layer
    if isinstance(feature_layers, int):
        return get_layer_index(feature_layers, num_hidden_layers)
    # If we have multiple feature layers, initialize up to the deepest one
    elif isinstance(feature_layers, (list, tuple)):
        return max(get_layer_index(idx, num_hidden_layers) for idx in feature_layers)
    raise TypeError(
        f"vision_layer_feature type: {type(feature_layers)} is not supported"
    )


def init_vision_tower_for_mistral3(
    hf_config: Mistral3Config,
    quant_config: QuantizationConfig | None,
    *,
    require_post_norm: bool | None = None,
    prefix: str = "",
) -> PixtralHFVisionModel:
    vision_config = hf_config.vision_config

    # Initialize the vision tower only up to the deepest required feature layer
    num_hidden_layers = _get_num_hidden_layers(hf_config)

    assert isinstance(vision_config, PixtralVisionConfig)

    return PixtralHFVisionModel(
        vision_config,
        quant_config=quant_config,
        num_hidden_layers_override=num_hidden_layers,
        require_post_norm=require_post_norm,
        prefix=prefix,
    )


@MULTIMODAL_REGISTRY.register_processor(
    Mistral3MultiModalProcessor,
    info=Mistral3ProcessingInfo,
    dummy_inputs=Mistral3DummyInputsBuilder,
)
class Mistral3ForConditionalGeneration(
    nn.Module,
    SupportsLoRA,
    SupportsMultiModal,
    SupportsPP,
    SupportsEagle,
    SupportsEagle3,
):
    packed_modules_mapping = {
        "qkv_proj": ["q_proj", "k_proj", "v_proj"],
        "gate_up_proj": ["gate_proj", "up_proj"],
    }

    hf_to_vllm_mapper = WeightsMapper(
        orig_to_new_prefix={
            # mapping for new names in checkpoint saved after transformers v4.52
            "model.language_model.": "language_model.model.",
            "model.vision_tower.": "vision_tower.",
            "model.multi_modal_projector.": "multi_modal_projector.",
            "lm_head.": "language_model.lm_head.",
            # Some PEFT LoRAs are trained against the text submodule directly
            # and produce names like `base_model.model.model.layers.*`.
            "model.": "language_model.model.",
        },
        orig_to_new_suffix={
            # FP8 quantized HF checkpoints use "activation_scale" and
            # "weight_scale_inv" but vLLM's FP8 linear layers register
            # them as "input_scale" and "weight_scale"
            ".activation_scale": ".input_scale",
            ".weight_scale_inv": ".weight_scale",
        },
    )

    @classmethod
    def get_placeholder_str(cls, modality: str, i: int) -> str | None:
        if modality.startswith("image"):
            return None

        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

        # NOTE: This is a special case for Pixtral-12B in the HF-format
        # https://huggingface.co/mistral-community/pixtral-12b/blob/main/config.json  # noqa
        if (
            config.projector_hidden_act is None
            and config.vision_config.hidden_act == "gelu"
        ):
            config.projector_hidden_act = "gelu"

        with self._mark_tower_model(vllm_config, "image"):
            self.vision_tower = init_vision_tower_for_mistral3(
                config,
                quant_config=quant_config,
                require_post_norm=False,
                prefix=maybe_prefix(prefix, "vision_tower"),
            )
            self.multi_modal_projector = Mistral3MultiModalProjector(
                vision_hidden_size=config.vision_config.hidden_size,
                text_hidden_size=config.text_config.hidden_size,
                projector_hidden_act=config.projector_hidden_act,
                spatial_merge_size=config.spatial_merge_size,
                patch_size=config.vision_config.patch_size,
                multimodal_projector_bias=config.multimodal_projector_bias,
                quant_config=quant_config,
                prefix=maybe_prefix(prefix, "multi_modal_projector"),
            )

        with self._mark_language_model(vllm_config):
            self.language_model = init_vllm_registered_model(
                vllm_config=vllm_config,
                hf_config=config.text_config,
                prefix=maybe_prefix(prefix, "language_model"),
            )

        self.make_empty_intermediate_tensors = (
            self.language_model.make_empty_intermediate_tensors
        )

    def _parse_and_validate_image_input(
        self, **kwargs: object
    ) -> Mistral3ImagePixelInputs | None:
        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

        return Mistral3ImagePixelInputs(
            type="pixel_values_pixtral",
            pixel_values=pixel_values,
        )

    def _process_image_input(
        self,
        image_input: Mistral3ImagePixelInputs,
    ) -> torch.Tensor | tuple[torch.Tensor, ...]:
        if image_input["type"] == "image_embeds":
            return image_input["data"]

        image_sizes = [
            (img.shape[-2], img.shape[-1]) for img in image_input["pixel_values"]
        ]

        image_features = self.vision_tower(image_input["pixel_values"])

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

        feature_sizes = [
            image_feature.shape[0] // self.config.spatial_merge_size**2
            for image_feature in image_features
        ]

        image_embeds = self.multi_modal_projector(
            torch.cat(image_features), image_sizes
        )
        if len(feature_sizes) > 1:
            image_embeds = torch.split(image_embeds, feature_sizes)
        else:
            image_embeds = (image_embeds,)
        return image_embeds

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

        vision_embeddings = self._process_image_input(image_input)

        return vision_embeddings

    def forward(
        self,
        input_ids: torch.Tensor | None,
        positions: torch.Tensor,
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
        **kwargs: object,
    ) -> torch.Tensor | IntermediateTensors:
        """Run forward pass for Mistral3.

        One key thing to understand is the `input_ids` already accounts for the
        positions of the to-be-inserted image embeddings.

        Concretely, consider a text prompt:
        `"USER: <image>\\nWhat's the content of the image?\\nASSISTANT:"`.

        Tokenizer outputs:
        `[1, 3148, 1001, 29901, 29871, 32000, 29871, 13, 5618, 29915, 29879,
        278, 2793, 310, 278, 1967, 29973, 13, 22933, 9047, 13566, 29901]`.

        To reserve space in KV cache, we have to insert placeholder tokens
        before they are inputted to the model, so the input processor prepends
        additional image tokens (denoted as `32000`), resulting in:
        `[1, 3148, 1001, 29901, 29871, 32000, ..., 32000, 29871, 13, 5618,
        29915, 29879, 278, 2793, 310, 278, 1967, 29973, 13, 22933, 9047, 13566,
        29901]`.

        We insert 575 tokens so that including the original image token in the
        input, there are a total of 576 (24 * 24) image tokens, which
        corresponds to the number of image tokens inputted to the language
        model, i.e. the number of image tokens outputted by the visual encoder.

        This way, the `positions` and `attn_metadata` are consistent
        with the `input_ids`.

        Args:
            input_ids: Flattened (concatenated) input_ids corresponding to a
                batch.
            positions: Position indices for the input tokens.
            intermediate_tensors: Intermediate tensors from prior forward pass.
            inputs_embeds: Optional tensor of input embeddings.

        Info:
            [`Mistral3ImagePixelInputs`][vllm.model_executor.models.mistral3.Mistral3ImagePixelInputs]
        """
        if intermediate_tensors is not None:
            inputs_embeds = 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,
    ) -> torch.Tensor | None:
        return self.language_model.compute_logits(hidden_states)

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

    def get_mm_mapping(self) -> MultiModelKeys:
        """
        Get the module prefix in multimodal models
        """
        return MultiModelKeys.from_string_field(
            language_model="language_model",
            connector="multi_modal_projector",
            tower_model="vision_tower",
        )