image-match/python/py/Lib/site-packages/transformers/models/mistral3/modular_mistral3.py

# Copyright 2025 HuggingFace Inc. team. All rights reserved.
#
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import torch
from torch import nn

from ...activations import ACT2FN
from ...cache_utils import Cache
from ...modeling_outputs import BaseModelOutputWithPooling
from ...processing_utils import Unpack
from ...utils import auto_docstring, logging
from ...utils.generic import can_return_tuple, merge_with_config_defaults
from ..llava.modeling_llava import (
    LlavaCausalLMOutputWithPast,
    LlavaForConditionalGeneration,
    LlavaModel,
    LlavaModelOutputWithPast,
    LlavaPreTrainedModel,
    TransformersKwargs,
)
from ..mistral.modeling_mistral import MistralRMSNorm
from .configuration_mistral3 import Mistral3Config


logger = logging.get_logger(__name__)


class Mistral3RMSNorm(MistralRMSNorm):
    pass


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

    def __init__(self, config: Mistral3Config):
        super().__init__()
        self.config = config

        hidden_size = config.vision_config.hidden_size
        self.spatial_merge_size = config.spatial_merge_size
        self.patch_size = self.config.vision_config.patch_size
        self.merging_layer = nn.Linear(hidden_size * self.spatial_merge_size**2, 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, config: Mistral3Config):
        super().__init__()
        self.norm = Mistral3RMSNorm(config.vision_config.hidden_size, eps=config.text_config.rms_norm_eps)
        self.patch_merger = Mistral3PatchMerger(config)
        # We have hidden_size * the number of vision feature layers
        self.num_feature_layers = (
            1 if isinstance(config.vision_feature_layer, int) else len(config.vision_feature_layer)
        )
        self.linear_1 = nn.Linear(
            config.vision_config.hidden_size * self.num_feature_layers,
            config.text_config.hidden_size,
            bias=config.multimodal_projector_bias,
        )
        self.act = ACT2FN[config.projector_hidden_act]
        self.linear_2 = nn.Linear(
            config.text_config.hidden_size, config.text_config.hidden_size, bias=config.multimodal_projector_bias
        )

    def forward(self, image_features: torch.Tensor, image_sizes: 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 Mistral3CausalLMOutputWithPast(LlavaCausalLMOutputWithPast):
    pass


class Mistral3ModelOutputWithPast(LlavaModelOutputWithPast):
    pass


class Mistral3PreTrainedModel(LlavaPreTrainedModel):
    pass


class Mistral3Model(LlavaModel):
    @merge_with_config_defaults
    @can_return_tuple
    @auto_docstring(
        custom_intro="Obtains image last hidden states from the vision tower and apply multimodal projection."
    )
    def get_image_features(
        self,
        pixel_values: torch.FloatTensor,
        image_sizes: torch.Tensor,
        vision_feature_layer: int | list[int] | list[int] | None = None,
        output_hidden_states: bool | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | BaseModelOutputWithPooling:
        kwargs = {k: v for k, v in kwargs.items() if v is not None}
        # this is not memory efficient at all (output_hidden_states=True) will save all the hidden states.
        image_outputs = self.vision_tower(
            pixel_values,
            image_sizes=image_sizes,
            output_hidden_states=True,  # Ignore arg on purpose
            return_dict=True,
            **kwargs,
        )
        # If we have one vision feature layer, return the corresponding hidden states,
        # otherwise, select the hidden states of each feature layer and concatenate them
        if isinstance(vision_feature_layer, int):
            selected_image_feature = image_outputs.hidden_states[vision_feature_layer]
        else:
            hs_pool = [image_outputs.hidden_states[layer_idx] for layer_idx in vision_feature_layer]
            selected_image_feature = torch.cat(hs_pool, dim=-1)

        image_features = self.multi_modal_projector(selected_image_feature.squeeze(0), image_sizes)
        downsample_ratio = self.vision_tower.patch_size * self.config.spatial_merge_size
        split_sizes = (
            (torch.as_tensor(image_sizes, device=image_features.device) // downsample_ratio).prod(dim=-1).tolist()
        )
        image_features = torch.split(image_features.squeeze(0), split_sizes)
        image_outputs.pooler_output = image_features

        return image_outputs

    @merge_with_config_defaults
    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        pixel_values: torch.FloatTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        vision_feature_layer: int | list[int] | list[int] | None = None,
        use_cache: bool | None = None,
        image_sizes: torch.Tensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | Mistral3ModelOutputWithPast:
        if (input_ids is None) ^ (inputs_embeds is not None):
            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")

        if inputs_embeds is None:
            inputs_embeds = self.get_input_embeddings()(input_ids)

        if pixel_values is not None:
            image_features = self.get_image_features(
                pixel_values=pixel_values,
                vision_feature_layer=vision_feature_layer,
                image_sizes=image_sizes,
                return_dict=True,
            ).pooler_output
            image_features = torch.cat(image_features, dim=0).to(inputs_embeds.device, inputs_embeds.dtype)
            special_image_mask = self.get_placeholder_mask(
                input_ids, inputs_embeds=inputs_embeds, image_features=image_features
            )
            inputs_embeds = inputs_embeds.masked_scatter(special_image_mask, image_features)

        outputs = self.language_model(
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            **kwargs,
        )

        return Mistral3ModelOutputWithPast(
            last_hidden_state=outputs.last_hidden_state,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
            image_hidden_states=image_features if pixel_values is not None else None,
        )


class Mistral3ForConditionalGeneration(LlavaForConditionalGeneration):
    @merge_with_config_defaults
    @can_return_tuple
    @auto_docstring
    def get_image_features(
        self,
        pixel_values: torch.FloatTensor,
        image_sizes: torch.Tensor,
        vision_feature_layer: int | list[int] | list[int] | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | BaseModelOutputWithPooling:
        return self.model.get_image_features(
            pixel_values=pixel_values,
            image_sizes=image_sizes,
            vision_feature_layer=vision_feature_layer,
            **kwargs,
        )

    @merge_with_config_defaults
    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        pixel_values: torch.FloatTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        labels: torch.LongTensor | None = None,
        use_cache: bool | None = None,
        logits_to_keep: int | torch.Tensor = 0,
        image_sizes: torch.Tensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | Mistral3CausalLMOutputWithPast:
        r"""
        Example:

        ```python
        >>> from PIL import Image
        >>> import httpx
        >>> from io import BytesIO
        >>> from transformers import AutoProcessor, Mistral3ForConditionalGeneration

        >>> model = Mistral3ForConditionalGeneration.from_pretrained("mistralai/Mistral-Small-3.1-24B-Instruct-2503")
        >>> processor = AutoProcessor.from_pretrained("mistralai/Mistral-Small-3.1-24B-Instruct-2503")

        >>> prompt = "<s>[INST][IMG]What is the image?[/INST]"
        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
        >>> with httpx.stream("GET", url) as response:
        ...     image = Image.open(BytesIO(response.read()))

        >>> inputs = processor(images=image, text=prompt, return_tensors="pt")

        >>> # Generate
        >>> generate_ids = model.generate(**inputs, max_new_tokens=15)
        >>> processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
        "What is the image?The image depicts two cats lying on a pink blanket."
        ```"""
        outputs = self.model(
            input_ids=input_ids,
            pixel_values=pixel_values,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            image_sizes=image_sizes,
            **kwargs,
        )

        hidden_states = outputs[0]
        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
        logits = self.lm_head(hidden_states[:, slice_indices, :])

        loss = None
        if labels is not None:
            loss = self.loss_function(
                logits=logits, labels=labels, vocab_size=self.config.text_config.vocab_size, **kwargs
            )

        return Mistral3CausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
            image_hidden_states=outputs.image_hidden_states,
        )


__all__ = [
    "Mistral3Model",
    "Mistral3PreTrainedModel",
    "Mistral3ForConditionalGeneration",
]