webapp/seed/Lib/site-packages/transformers/models/smolvlm/modeling_smolvlm.py

#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
#           This file was automatically generated from src/transformers/models/smolvlm/modular_smolvlm.py.
#               Do NOT edit this file manually as any edits will be overwritten by the generation of
#             the file from the modular. If any change should be done, please apply the change to the
#                          modular_smolvlm.py file directly. One of our CI enforces this.
#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
# Copyright 2025 the HuggingFace Inc. team. All rights reserved.
# Written by Orr Zohar
#
# 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.

from collections.abc import Callable
from dataclasses import dataclass

import torch
from torch import nn

from ...activations import ACT2FN
from ...cache_utils import Cache, DynamicCache
from ...generation import GenerationConfig, GenerationMixin
from ...masking_utils import create_bidirectional_mask
from ...modeling_flash_attention_utils import FlashAttentionKwargs
from ...modeling_layers import GradientCheckpointingLayer
from ...modeling_outputs import BaseModelOutput, BaseModelOutputWithPooling, ModelOutput
from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
from ...processing_utils import Unpack
from ...utils import (
    TransformersKwargs,
    auto_docstring,
    can_return_tuple,
    logging,
    torch_compilable_check,
)
from ...utils.generic import check_model_inputs
from ..auto import AutoModel
from .configuration_smolvlm import SmolVLMConfig, SmolVLMVisionConfig


logger = logging.get_logger(__name__)


@auto_docstring
class SmolVLMPreTrainedModel(PreTrainedModel):
    config: SmolVLMConfig
    base_model_prefix = "model"
    input_modalities = ("image", "text")
    supports_gradient_checkpointing = True
    _no_split_modules = ["SmolVLMVisionAttention", "SmolVLMDecoderLayer"]
    _skip_keys_device_placement = "past_key_values"
    _supports_flash_attn = True
    _supports_sdpa = True
    _supports_flex_attn = True
    _supports_attention_backend = True


class SmolVLMVisionEmbeddings(nn.Module):
    """
    This is a modified version of `siglip.modelign_siglip.SiglipVisionEmbeddings` to enable images of variable
    resolution.

    The modifications are adapted from [Patch n' Pack: NaViT, a Vision Transformer for any Aspect Ratio and Resolution](https://huggingface.co/papers/2307.06304)
    which allows treating images in their native aspect ratio and without the need to resize them to the same
    fixed size. In particular, we start from the original pre-trained SigLIP model
    (which uses images of fixed-size square images) and adapt it by training on images of variable resolutions.
    """

    def __init__(self, config: SmolVLMVisionConfig):
        super().__init__()
        self.embed_dim = config.hidden_size
        self.image_size = config.image_size
        self.patch_size = config.patch_size

        self.patch_embedding = nn.Conv2d(
            in_channels=config.num_channels,
            out_channels=self.embed_dim,
            kernel_size=self.patch_size,
            stride=self.patch_size,
            padding="valid",
        )

        self.num_patches_per_side = self.image_size // self.patch_size
        self.num_patches = self.num_patches_per_side**2
        self.num_positions = self.num_patches
        self.position_embedding = nn.Embedding(self.num_positions, self.embed_dim)

    def forward(self, pixel_values: torch.FloatTensor, patch_attention_mask: torch.BoolTensor) -> torch.Tensor:
        batch_size, _, max_im_h, max_im_w = pixel_values.shape

        patch_embeds = self.patch_embedding(pixel_values)
        embeddings = patch_embeds.flatten(2).transpose(1, 2)

        max_nb_patches_h, max_nb_patches_w = max_im_h // self.patch_size, max_im_w // self.patch_size
        boundaries = torch.arange(
            1 / self.num_patches_per_side, 1.0, 1 / self.num_patches_per_side, device=pixel_values.device
        )
        position_ids = torch.full(
            size=(batch_size, max_nb_patches_h * max_nb_patches_w), fill_value=0, device=pixel_values.device
        )

        nb_patches_h = patch_attention_mask[:, :, 0].sum(dim=1)  # (batch_size,)
        nb_patches_w = patch_attention_mask[:, 0, :].sum(dim=1)  # (batch_size,)

        step_h = 1.0 / nb_patches_h  # (batch_size,)
        step_w = 1.0 / nb_patches_w  # (batch_size,)

        max_patches_h = patch_attention_mask.size(1)
        max_patches_w = patch_attention_mask.size(2)
        h_indices = torch.arange(max_patches_h, device=position_ids.device, dtype=torch.float32)
        w_indices = torch.arange(max_patches_w, device=position_ids.device, dtype=torch.float32)

        fractional_coords_h = h_indices[None, :] * step_h[:, None]
        fractional_coords_w = w_indices[None, :] * step_w[:, None]

        fractional_coords_h = torch.clamp(fractional_coords_h, max=(1.0 - 1e-6))
        fractional_coords_w = torch.clamp(fractional_coords_w, max=(1.0 - 1e-6))

        fractional_coords_h = fractional_coords_h.to(pixel_values.dtype)
        fractional_coords_w = fractional_coords_w.to(pixel_values.dtype)

        bucket_coords_h = torch.bucketize(fractional_coords_h, boundaries, right=True)
        bucket_coords_w = torch.bucketize(fractional_coords_w, boundaries, right=True)

        pos_ids = bucket_coords_h[:, :, None] * self.num_patches_per_side + bucket_coords_w[:, None, :]
        pos_ids = pos_ids.reshape(batch_size, -1)

        position_ids[patch_attention_mask.view(batch_size, -1)] = pos_ids[patch_attention_mask.view(batch_size, -1)]

        embeddings = embeddings + self.position_embedding(position_ids)
        return embeddings


def eager_attention_forward(
    module: nn.Module,
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    attention_mask: torch.Tensor | None,
    scaling: float,
    dropout: float = 0.0,
    **kwargs,
):
    attn_weights = torch.matmul(query, key.transpose(-1, -2)) * scaling
    if attention_mask is not None:
        attn_weights = attn_weights + attention_mask

    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)

    attn_output = torch.matmul(attn_weights, value)
    attn_output = attn_output.transpose(1, 2).contiguous()

    return attn_output, attn_weights


class SmolVLMVisionAttention(nn.Module):
    """Multi-headed attention from 'Attention Is All You Need' paper"""

    def __init__(self, config):
        super().__init__()
        self.config = config
        self.embed_dim = config.hidden_size
        self.num_heads = config.num_attention_heads
        self.head_dim = self.embed_dim // self.num_heads
        if self.head_dim * self.num_heads != self.embed_dim:
            raise ValueError(
                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:"
                f" {self.num_heads})."
            )
        self.scale = self.head_dim**-0.5
        self.dropout = config.attention_dropout

        self.k_proj = nn.Linear(self.embed_dim, self.embed_dim)
        self.v_proj = nn.Linear(self.embed_dim, self.embed_dim)
        self.q_proj = nn.Linear(self.embed_dim, self.embed_dim)
        self.out_proj = nn.Linear(self.embed_dim, self.embed_dim)

        # Ignore copy
        self.is_causal = False

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.Tensor | None = None,
        **kwargs,
    ) -> tuple[torch.Tensor, torch.Tensor | None]:
        """Input shape: Batch x Time x Channel"""

        batch_size, seq_length, embed_dim = hidden_states.shape

        queries = self.q_proj(hidden_states)
        keys = self.k_proj(hidden_states)
        values = self.v_proj(hidden_states)

        queries = queries.view(batch_size, seq_length, self.num_heads, self.head_dim).transpose(1, 2)
        keys = keys.view(batch_size, seq_length, self.num_heads, self.head_dim).transpose(1, 2)
        values = values.view(batch_size, seq_length, self.num_heads, self.head_dim).transpose(1, 2)

        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
            self.config._attn_implementation, eager_attention_forward
        )

        attn_output, attn_weights = attention_interface(
            self,
            queries,
            keys,
            values,
            attention_mask,
            is_causal=self.is_causal,
            scaling=self.scale,
            dropout=0.0 if not self.training else self.dropout,
        )

        attn_output = attn_output.reshape(batch_size, seq_length, embed_dim).contiguous()
        attn_output = self.out_proj(attn_output)

        return attn_output, attn_weights


class SmolVLMVisionMLP(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.activation_fn = ACT2FN[config.hidden_act]
        self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
        self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        hidden_states = self.fc1(hidden_states)
        hidden_states = self.activation_fn(hidden_states)
        hidden_states = self.fc2(hidden_states)
        return hidden_states


class SmolVLMEncoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: SmolVLMVisionConfig):
        super().__init__()
        self.embed_dim = config.hidden_size
        self.self_attn = SmolVLMVisionAttention(config)
        self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
        self.mlp = SmolVLMVisionMLP(config)
        self.layer_norm2 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)

    @auto_docstring
    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.Tensor,
        **kwargs: Unpack[TransformersKwargs],
    ) -> torch.FloatTensor:
        residual = hidden_states

        hidden_states = self.layer_norm1(hidden_states)
        hidden_states, _ = self.self_attn(
            hidden_states=hidden_states,
            attention_mask=attention_mask,
            **kwargs,
        )
        hidden_states = residual + hidden_states

        residual = hidden_states
        hidden_states = self.layer_norm2(hidden_states)
        hidden_states = self.mlp(hidden_states)
        hidden_states = residual + hidden_states

        return hidden_states


class SmolVLMEncoder(nn.Module):
    """
    Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a
    [`SmolVLMEncoderLayer`].

    Args:
        config: SmolVLMConfig
    """

    def __init__(self, config: SmolVLMConfig):
        super().__init__()
        self.config = config
        self.layers = nn.ModuleList([SmolVLMEncoderLayer(config) for _ in range(config.num_hidden_layers)])
        self.gradient_checkpointing = False

    # Ignore copy
    @auto_docstring
    def forward(
        self,
        inputs_embeds,
        attention_mask: torch.Tensor | None = None,
    ) -> tuple | BaseModelOutput:
        hidden_states = inputs_embeds
        for encoder_layer in self.layers:
            layer_outputs = encoder_layer(
                hidden_states,
                attention_mask,
            )

            hidden_states = layer_outputs

        return BaseModelOutput(last_hidden_state=hidden_states)


@auto_docstring(
    custom_intro="""
    The SmolVLM Vision Transformer Model outputting raw image embedding.
    """
)
class SmolVLMVisionTransformer(SmolVLMPreTrainedModel):
    config: SmolVLMVisionConfig
    input_modalities = ("image",)
    _supports_sdpa = True
    _supports_flash_attn = True
    _supports_flex_attn = True
    _can_record_outputs = {
        "hidden_states": SmolVLMEncoderLayer,
        "attentions": SmolVLMVisionAttention,
    }

    def __init__(self, config: SmolVLMVisionConfig):
        super().__init__(config)
        embed_dim = config.hidden_size

        self.embeddings = SmolVLMVisionEmbeddings(config)
        self.encoder = SmolVLMEncoder(config)
        self.patch_size = config.patch_size
        self.post_layernorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)

        self.post_init()

    def get_input_embeddings(self):
        return self.embeddings

    def set_input_embeddings(self, value):
        self.embeddings = value

    @check_model_inputs(tie_last_hidden_states=False)
    def forward(
        self,
        pixel_values,
        patch_attention_mask: torch.BoolTensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | BaseModelOutput:
        batch_size = pixel_values.size(0)
        if patch_attention_mask is None:
            patch_size = self.patch_size
            patch_attention_mask = torch.ones(
                (
                    batch_size,
                    pixel_values.size(2) // patch_size,
                    pixel_values.size(3) // patch_size,
                )
            )
            patch_attention_mask = patch_attention_mask.to(dtype=torch.bool, device=pixel_values.device)

        hidden_states = self.embeddings(pixel_values=pixel_values, patch_attention_mask=patch_attention_mask)

        patch_attention_mask = patch_attention_mask.view(batch_size, -1)
        # Create the correct attention mask based on the attention implementation
        patch_attention_mask = create_bidirectional_mask(
            config=self.config,
            input_embeds=hidden_states,
            attention_mask=patch_attention_mask,
        )

        encoder_outputs: BaseModelOutput = self.encoder(
            inputs_embeds=hidden_states,
            attention_mask=patch_attention_mask,
        )

        last_hidden_state = encoder_outputs.last_hidden_state
        last_hidden_state = self.post_layernorm(last_hidden_state)

        return BaseModelOutput(
            last_hidden_state=last_hidden_state,
        )


@dataclass
@auto_docstring(
    custom_intro="""
    Base class for SmolVLM model's outputs that may also contain a past key/values (to speed up sequential decoding).
    """
)
class SmolVLMBaseModelOutputWithPast(ModelOutput):
    r"""
    last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
        Sequence of hidden-states at the output of the last layer of the model.
        If `past_key_values` is used only the last hidden-state of the sequences of shape `(batch_size, 1,
        hidden_size)` is output.
    past_key_values (`Cache`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
        It is a [`~cache_utils.Cache`] instance. For more details, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache).

        Contains pre-computed hidden-states (key and values in the self-attention blocks and optionally if
        `config.is_encoder_decoder=True` in the cross-attention blocks) that can be used (see `past_key_values`
        input) to speed up sequential decoding.
    image_hidden_states (`tuple(torch.FloatTensor)`, *optional*):
        Tuple of `torch.FloatTensor` (one for the output of the image embeddings, `(batch_size, num_images,
        sequence_length, hidden_size)`.
        image_hidden_states of the model produced by the vision encoder
    """

    last_hidden_state: torch.FloatTensor | None = None
    past_key_values: Cache | None = None
    hidden_states: tuple[torch.FloatTensor] | None = None
    attentions: tuple[torch.FloatTensor] | None = None
    image_hidden_states: tuple[torch.FloatTensor] | None = None


class SmolVLMSimpleMLP(nn.Module):
    def __init__(self, config):
        super().__init__()
        input_size = config.vision_config.hidden_size * (config.scale_factor**2)
        output_size = config.text_config.hidden_size
        self.proj = nn.Linear(input_size, output_size, bias=False)

    def forward(self, x):
        return self.proj(x)


class SmolVLMConnector(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.scale_factor = config.scale_factor
        self.modality_projection = SmolVLMSimpleMLP(config)

    def pixel_shuffle(self, x, scale_factor=2):
        bsz, seq, embed_dim = x.size()
        height = width = int(seq**0.5)
        x = x.view(bsz, height, width, embed_dim)
        x = x.view(bsz, height, int(width / scale_factor), embed_dim * scale_factor)
        x = x.permute(0, 2, 1, 3)
        x = x.reshape(bsz, int(width / scale_factor), int(height / scale_factor), embed_dim * (scale_factor**2))
        x = x.permute(0, 2, 1, 3)
        x = x.reshape(bsz, int(seq / (scale_factor**2)), embed_dim * (scale_factor**2))
        return x

    def forward(self, image_hidden_states):
        image_hidden_states = self.pixel_shuffle(image_hidden_states, self.scale_factor)
        image_hidden_states = self.modality_projection(image_hidden_states)
        return image_hidden_states


@auto_docstring(
    custom_intro="""
    SmolVLM model consisting of a SIGLIP vision encoder and Llama3 language decoder
    """
)
class SmolVLMModel(SmolVLMPreTrainedModel):
    """
    A subclass of Idefics3Model. We do *not* remove or block the call to inputs_merger
    in forward. Instead, we override inputs_merger here with custom logic.
    """

    def __init__(self, config: SmolVLMConfig):
        super().__init__(config)
        self.padding_idx = self.config.text_config.pad_token_id
        self.vocab_size = self.config.text_config.vocab_size

        self.vision_model = SmolVLMVisionTransformer._from_config(config.vision_config)
        self.connector = SmolVLMConnector(config)
        self.text_model = AutoModel.from_config(config.text_config)

        self.image_seq_len = int(
            ((config.vision_config.image_size // config.vision_config.patch_size) ** 2) / (config.scale_factor**2)
        )
        self.image_token_id = self.config.image_token_id

        self.post_init()

    def get_input_embeddings(self):
        return self.text_model.get_input_embeddings()

    def set_input_embeddings(self, value):
        self.text_model.set_input_embeddings(value)

    def inputs_merger(
        self, input_ids: torch.LongTensor, inputs_embeds: torch.Tensor, image_hidden_states: torch.Tensor
    ):
        """
        This method aims at merging the token embeddings with the image hidden states into one single sequence of vectors that are fed to the transformer LM.
        The merging happens as follows:
        - The text token sequence is: `tok_1 tok_2 tok_3 <fake_token_around_image> <image> <image> ... <image> <fake_token_around_image> tok_4`.
        - We get the image hidden states for the image through the vision encoder and that hidden state, after a pixel shuffle operation, is then projected into the text embedding space.
        We thus have a sequence of image hidden states of size (1, image_seq_len, hidden_dim), where 1 is for batch_size of 1 image and hidden_dim is the hidden_dim of the LM transformer.
        - The merging happens so that we obtain the following sequence: `vector_tok_1 vector_tok_2 vector_tok_3 vector_fake_tok_around_image {sequence of image_seq_len image hidden states} vector_fake_toke_around_image vector_tok_4`. That sequence is fed to the LM.
        - To fit the format of that sequence, `input_ids`, `input_embeds`, `attention_mask` are all 3 adapted to insert the image hidden states.
        """
        _, patch_size, _ = image_hidden_states.shape

        if input_ids is None:
            image_mask = inputs_embeds == self.get_input_embeddings()(
                torch.tensor(self.config.image_token_id, dtype=torch.long, device=inputs_embeds.device)
            )
            image_mask = image_mask[..., 0]  # slice off the hidden dim
        else:
            image_mask = input_ids == self.config.image_token_id

        num_image_tokens = image_mask.sum(dim=1)
        torch_compilable_check(
            torch.all(num_image_tokens % patch_size == 0),
            "At least one sample has <image> tokens not divisible by patch_size.",
        )
        blocks_per_sample = num_image_tokens // patch_size

        offsets = torch.nn.functional.pad(blocks_per_sample.cumsum(dim=0), (1, 0), value=0)
        block_offset = offsets[:-1]
        row_cum = image_mask.cumsum(dim=-1)
        chunk_idx = (row_cum - 1) // patch_size
        local_idx = (row_cum - 1) % patch_size
        block_idx = block_offset.unsqueeze(1) + chunk_idx

        image_embeds = torch.zeros_like(inputs_embeds)
        image_embeds[image_mask] = image_hidden_states[block_idx[image_mask], local_idx[image_mask], :]

        merged_embeds = torch.where(image_mask.unsqueeze(-1), image_embeds, inputs_embeds)
        return merged_embeds

    @can_return_tuple
    @auto_docstring(
        custom_intro="Encodes images into continuous embeddings that can be forwarded to the language model."
    )
    def get_image_features(
        self,
        pixel_values: torch.FloatTensor,
        pixel_attention_mask: torch.LongTensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | BaseModelOutputWithPooling:
        r"""
        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)`):
            The tensors corresponding to the input images.
        pixel_attention_mask (`torch.LongTensor`, *optional*):
            The attention mask indicating padded regions in the image.
        """
        batch_size, num_images, num_channels, height, width = pixel_values.shape
        pixel_values = pixel_values.to(dtype=self.dtype)  # fp16 compatibility
        pixel_values = pixel_values.view(batch_size * num_images, *pixel_values.shape[2:])

        # Remove padding images - padding images are full 0.
        nb_values_per_image = pixel_values.shape[1:].numel()
        real_images_inds = (pixel_values == 0.0).sum(dim=(-1, -2, -3)) != nb_values_per_image

        # If no images, leave one empty image.
        real_images_inds[0] |= ~torch.any(real_images_inds)

        pixel_values = pixel_values[real_images_inds].contiguous()
        # Handle the vision attention mask
        if pixel_attention_mask is None:
            pixel_attention_mask = torch.ones(
                size=[pixel_values.shape[i] for i in (0, 2, 3)],
                dtype=torch.bool,
                device=pixel_values.device,
            )
        else:
            # Remove padding images from the mask
            pixel_attention_mask = pixel_attention_mask.view(batch_size * num_images, *pixel_attention_mask.shape[2:])
            pixel_attention_mask = pixel_attention_mask[real_images_inds].contiguous()
        patch_size = self.config.vision_config.patch_size
        patches_subgrid = pixel_attention_mask.unfold(dimension=1, size=patch_size, step=patch_size)
        patches_subgrid = patches_subgrid.unfold(dimension=2, size=patch_size, step=patch_size)
        patch_attention_mask = (patches_subgrid.sum(dim=(-1, -2)) > 0).bool()

        # Get sequence from the vision encoder
        image_outputs = self.vision_model(
            pixel_values=pixel_values, patch_attention_mask=patch_attention_mask, return_dict=True, **kwargs
        )
        image_hidden_states = image_outputs.last_hidden_state

        # Modality projection & resampling
        image_features = self.connector(image_hidden_states)
        image_outputs.pooler_output = image_features

        return image_outputs

    @can_return_tuple
    @auto_docstring(
        custom_intro="""
        Inputs fed to the model can have an arbitrary number of images. To account for this, pixel_values fed to
        the model have image padding -> (batch_size, max_num_images, 3, max_heights, max_widths) where
        max_num_images is the maximum number of images among the batch_size samples in the batch.
        Padding images are not needed beyond padding the pixel_values at the entrance of the model.
        For efficiency, we only pass through the vision_model's forward the real images by
        discarding the padding images i.e. pixel_values of size (image_batch_size, 3, height, width) where
        image_batch_size would be 7 when num_images_per_sample=[1, 3, 1, 2] and max_num_images would be 3.
        """
    )
    def forward(
        self,
        input_ids: torch.LongTensor | 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,
        pixel_values: torch.FloatTensor | None = None,
        pixel_attention_mask: torch.BoolTensor | None = None,
        image_hidden_states: torch.FloatTensor | None = None,
        use_cache: bool | None = None,
        output_attentions: bool | None = None,
        output_hidden_states: bool | None = None,
        return_dict: bool | None = None,
        cache_position: torch.LongTensor | None = None,
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> tuple | SmolVLMBaseModelOutputWithPast:
        r"""
        pixel_attention_mask (`torch.Tensor` of shape `(batch_size, image_size, image_size)`, *optional*):
            Mask to avoid performing attention on padding pixel indices.
        image_hidden_states (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)`):
            The hidden states of the image encoder after modality projection.
        """
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        use_cache = use_cache if use_cache is not None else self.config.use_cache
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        if self.training and self.text_model.gradient_checkpointing and use_cache:
            logger.warning_once(
                "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
            )
            use_cache = False

        # retrieve input_ids and inputs_embeds
        if input_ids is not None:
            batch_size, seq_length = input_ids.shape
        elif inputs_embeds is not None:
            batch_size, seq_length, _ = inputs_embeds.shape
        else:
            raise ValueError("You have to specify either input_ids or inputs_embeds")

        if use_cache and past_key_values is None:
            past_key_values = DynamicCache(config=self.config)

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

        # START VISUAL INPUTS INTEGRATION
        if pixel_values is not None and image_hidden_states is not None:
            raise ValueError("You cannot specify both pixel_values and image_hidden_states at the same time")

        if pixel_values is not None:
            image_hidden_states = self.get_image_features(
                pixel_values, pixel_attention_mask, return_dict=True
            ).pooler_output
            image_hidden_states = image_hidden_states.to(inputs_embeds.device)
        elif image_hidden_states is not None:
            image_hidden_states = image_hidden_states.to(dtype=self.dtype, device=inputs_embeds.device)

        if image_hidden_states is not None:
            # When we generate, we don't want to replace the potential image_token_id that we generated by images
            # that simply don't exist
            inputs_embeds = self.inputs_merger(
                input_ids=input_ids,
                inputs_embeds=inputs_embeds,
                image_hidden_states=image_hidden_states,
            )

        outputs = self.text_model(
            inputs_embeds=inputs_embeds,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=True,
            cache_position=cache_position,
            **kwargs,
        )

        return SmolVLMBaseModelOutputWithPast(
            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_hidden_states,
        )


@dataclass
@auto_docstring(
    custom_intro="""
    Base class for Idefics causal language model (or autoregressive) outputs.
    """
)
class SmolVLMCausalLMOutputWithPast(ModelOutput):
    r"""
    loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
        Language modeling loss (for next-token prediction).
    logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
        Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
    past_key_values (`Cache`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
        It is a [`~cache_utils.Cache`] instance. For more details, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache).

        Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
        `past_key_values` input) to speed up sequential decoding.
    image_hidden_states (`tuple(torch.FloatTensor)`, *optional*):
        Tuple of `torch.FloatTensor` (one for the output of the image embeddings, `(batch_size, num_images,
        sequence_length, hidden_size)`.
        image_hidden_states of the model produced by the vision encoder
    """

    loss: torch.FloatTensor | None = None
    logits: torch.FloatTensor | None = None
    past_key_values: Cache | None = None
    hidden_states: tuple[torch.FloatTensor] | None = None
    attentions: tuple[torch.FloatTensor] | None = None
    image_hidden_states: tuple[torch.FloatTensor] | None = None


@auto_docstring(
    custom_intro="""
    The SmolVLM Model with a language modeling head. It is made up a SigLIP vision encoder, with a language modeling head on top.
    """
)
class SmolVLMForConditionalGeneration(SmolVLMPreTrainedModel, GenerationMixin):
    _tied_weights_keys = {"lm_head.weight": "model.text_model.embed_tokens.weight"}

    def __init__(self, config):
        super().__init__(config)
        self.model = SmolVLMModel(config)
        self.image_token_id = self.config.image_token_id
        self.lm_head = nn.Linear(config.text_config.hidden_size, config.text_config.vocab_size, bias=False)
        self.vocab_size = config.text_config.vocab_size
        self.model.text_model.generation_config = GenerationConfig.from_model_config(config)

        # Initialize weights and apply final processing
        self.post_init()

    def get_input_embeddings(self):
        return self.model.text_model.get_input_embeddings()

    def set_input_embeddings(self, value):
        self.model.text_model.set_input_embeddings(value)

    @auto_docstring
    def get_image_features(
        self,
        pixel_values: torch.FloatTensor,
        pixel_attention_mask: torch.LongTensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | BaseModelOutputWithPooling:
        r"""
        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)`):
            The tensors corresponding to the input images.
        pixel_attention_mask (`torch.LongTensor`, *optional*):
            The attention mask indicating padded regions in the image.
        """
        return self.model.get_image_features(
            pixel_values=pixel_values, pixel_attention_mask=pixel_attention_mask, **kwargs
        )

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | 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,
        pixel_values: torch.FloatTensor | None = None,
        pixel_attention_mask: torch.BoolTensor | None = None,
        image_hidden_states: torch.FloatTensor | None = None,
        labels: torch.LongTensor | None = None,
        use_cache: bool | None = None,
        output_attentions: bool | None = None,
        output_hidden_states: bool | None = None,
        cache_position: torch.LongTensor | None = None,
        return_dict: bool | None = None,
        logits_to_keep: int | torch.Tensor = 0,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | SmolVLMCausalLMOutputWithPast:
        r"""
        pixel_attention_mask (`torch.Tensor` of shape `(batch_size, image_size, image_size)`, *optional*):
            Mask to avoid performing attention on padding pixel indices.
        image_hidden_states (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)`):
            The hidden states of the image encoder after modality projection.
        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
            config.vocab_size]` or `model.image_token_id`. Tokens with indices set to `model.image_token_id` are
            ignored (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.

        Example:

        ```python
        >>> import httpx
        >>> from io import BytesIO
        >>> import torch
        >>> from PIL import Image
        >>> from io import BytesIO

        >>> from transformers import AutoProcessor, AutoModelForImageTextToText
        >>> from transformers.image_utils import load_image

        >>> # Note that passing the image urls (instead of the actual pil images) to the processor is also possible
        >>> image1 = load_image("https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg")
        >>> image2 = load_image("https://cdn.britannica.com/59/94459-050-DBA42467/Skyline-Chicago.jpg")
        >>> image3 = load_image("https://cdn.britannica.com/68/170868-050-8DDE8263/Golden-Gate-Bridge-San-Francisco.jpg")

        >>> processor = AutoProcessor.from_pretrained("HuggingFaceTB/SmolVLM2-2.2B-Instruct")
        >>> model = AutoModelForImageTextToText.from_pretrained("HuggingFaceTB/SmolVLM2-2.2B-Instruct", dtype=torch.bfloat16, device_map="auto")

        >>> # Create inputs
        >>> messages = [
        ...     {
        ...         "role": "user",
        ...         "content": [
        ...             {"type": "video", "path": path/to/video},
        ...             {"type": "text", "text": "What is happening in this video?"},
        ...         ]
        ...     }
        ... ]

        >>> inputs = processor.apply_chat_template([messages], add_generation_prompt=True)

        >>> # Generate
        >>> generated_ids = model.generate(**inputs, max_new_tokens=256)
        >>> generated_texts = processor.batch_decode(generated_ids, skip_special_tokens=True)

        >>> print(generated_texts)
        ```"""
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
        outputs = self.model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            pixel_values=pixel_values,
            pixel_attention_mask=pixel_attention_mask,
            image_hidden_states=image_hidden_states,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            cache_position=cache_position,
            return_dict=True,
            **kwargs,
        )

        hidden_states = outputs[0]
        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
        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 SmolVLMCausalLMOutputWithPast(
            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,
        )

    def prepare_inputs_for_generation(
        self,
        input_ids,
        past_key_values=None,
        attention_mask=None,
        inputs_embeds=None,
        cache_position=None,
        pixel_values=None,
        pixel_attention_mask=None,
        image_hidden_states=None,
        logits_to_keep=None,
        is_first_iteration=False,
        **kwargs,
    ):
        # Overwritten -- there are mutually exclusive inputs (if the logic to make `image_hidden_states` take
        # precedence is moved to the model, we can remove this fn)

        model_inputs = super().prepare_inputs_for_generation(
            input_ids,
            past_key_values=past_key_values,
            attention_mask=attention_mask,
            inputs_embeds=inputs_embeds,
            cache_position=cache_position,
            pixel_values=pixel_values,
            pixel_attention_mask=pixel_attention_mask,
            image_hidden_states=image_hidden_states,
            logits_to_keep=logits_to_keep,
            is_first_iteration=is_first_iteration,
            **kwargs,
        )

        if image_hidden_states is not None or not is_first_iteration:
            model_inputs["pixel_values"] = None
            model_inputs["pixel_attention_mask"] = None

        return model_inputs


__all__ = ["SmolVLMForConditionalGeneration", "SmolVLMPreTrainedModel", "SmolVLMModel", "SmolVLMVisionTransformer"]