modeling_navit_siglip.py

# coding=utf-8
# Copyright 2024 Google AI and The HuggingFace 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.
""" PyTorch Siglip model. """
# Copied from  HuggingFaceM4/siglip-so400m-14-980-flash-attn2-navit and add tgt_sizes


import os
import math
import warnings
from dataclasses import dataclass
from typing import Any, Optional, Tuple, Union

import numpy as np
import torch
import torch.nn.functional as F
import torch.utils.checkpoint
from torch import nn
from torch.nn.init import _calculate_fan_in_and_fan_out

from transformers.activations import ACT2FN
from transformers.modeling_attn_mask_utils import _prepare_4d_attention_mask
from transformers.modeling_outputs import BaseModelOutput, BaseModelOutputWithPooling
from transformers.modeling_utils import PreTrainedModel
from transformers.configuration_utils import PretrainedConfig
from transformers.utils import (
    ModelOutput,
    add_start_docstrings,
    add_start_docstrings_to_model_forward,
    logging,
    replace_return_docstrings,
)
from transformers.utils import logging

logger = logging.get_logger(__name__)

class SiglipVisionConfig(PretrainedConfig):
    r"""
    This is the configuration class to store the configuration of a [`SiglipVisionModel`]. It is used to instantiate a
    Siglip vision encoder according to the specified arguments, defining the model architecture. Instantiating a
    configuration with the defaults will yield a similar configuration to that of the vision encoder of the Siglip
    [google/siglip-base-patch16-224](https://huggingface.co/google/siglip-base-patch16-224) architecture.
    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
    documentation from [`PretrainedConfig`] for more information.
    Args:
        hidden_size (`int`, *optional*, defaults to 768):
            Dimensionality of the encoder layers and the pooler layer.
        intermediate_size (`int`, *optional*, defaults to 3072):
            Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
        num_hidden_layers (`int`, *optional*, defaults to 12):
            Number of hidden layers in the Transformer encoder.
        num_attention_heads (`int`, *optional*, defaults to 12):
            Number of attention heads for each attention layer in the Transformer encoder.
        num_channels (`int`, *optional*, defaults to 3):
            Number of channels in the input images.
        image_size (`int`, *optional*, defaults to 224):
            The size (resolution) of each image.
        patch_size (`int`, *optional*, defaults to 16):
            The size (resolution) of each patch.
        hidden_act (`str` or `function`, *optional*, defaults to `"gelu_pytorch_tanh"`):
            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
            `"relu"`, `"selu"` and `"gelu_new"` ``"quick_gelu"` are supported.
        layer_norm_eps (`float`, *optional*, defaults to 1e-06):
            The epsilon used by the layer normalization layers.
        attention_dropout (`float`, *optional*, defaults to 0.0):
            The dropout ratio for the attention probabilities.
    Example:
    ```python
    >>> from transformers import SiglipVisionConfig, SiglipVisionModel
    >>> # Initializing a SiglipVisionConfig with google/siglip-base-patch16-224 style configuration
    >>> configuration = SiglipVisionConfig()
    >>> # Initializing a SiglipVisionModel (with random weights) from the google/siglip-base-patch16-224 style configuration
    >>> model = SiglipVisionModel(configuration)
    >>> # Accessing the model configuration
    >>> configuration = model.config
    ```"""

    model_type = "siglip_vision_model"

    def __init__(
        self,
        hidden_size=768,
        intermediate_size=3072,
        num_hidden_layers=12,
        num_attention_heads=12,
        num_channels=3,
        image_size=224,
        patch_size=16,
        hidden_act="gelu_pytorch_tanh",
        layer_norm_eps=1e-6,
        attention_dropout=0.0,
        **kwargs,
    ):
        super().__init__(**kwargs)

        self.hidden_size = hidden_size
        self.intermediate_size = intermediate_size
        self.num_hidden_layers = num_hidden_layers
        self.num_attention_heads = num_attention_heads
        self.num_channels = num_channels
        self.patch_size = patch_size
        self.image_size = image_size
        self.attention_dropout = attention_dropout
        self.layer_norm_eps = layer_norm_eps
        self.hidden_act = hidden_act

    @classmethod
    def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
        cls._set_token_in_kwargs(kwargs)

        config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)

        # get the vision config dict if we are loading from SiglipConfig
        if config_dict.get("model_type") == "siglip":
            config_dict = config_dict["vision_config"]

        if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
            logger.warning(
                f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
                f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
            )

        return cls.from_dict(config_dict, **kwargs)
        

_CHECKPOINT_FOR_DOC = "google/siglip-base-patch16-224"

SIGLIP_PRETRAINED_MODEL_ARCHIVE_LIST = [
    "google/siglip-base-patch16-224",
    # See all SigLIP models at https://huggingface.co/models?filter=siglip
]


# Copied from transformers.models.llama.modeling_llama._get_unpad_data
def _get_unpad_data(attention_mask):
    seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
    indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
    max_seqlen_in_batch = seqlens_in_batch.max().item()
    cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.torch.int32), (1, 0))
    return (
        indices,
        cu_seqlens,
        max_seqlen_in_batch,
    )


def _trunc_normal_(tensor, mean, std, a, b):
    # Cut & paste from PyTorch official master until it's in a few official releases - RW
    # Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf
    def norm_cdf(x):
        # Computes standard normal cumulative distribution function
        return (1.0 + math.erf(x / math.sqrt(2.0))) / 2.0

    if (mean < a - 2 * std) or (mean > b + 2 * std):
        warnings.warn(
            "mean is more than 2 std from [a, b] in nn.init.trunc_normal_. "
            "The distribution of values may be incorrect.",
            stacklevel=2,
        )

    # Values are generated by using a truncated uniform distribution and
    # then using the inverse CDF for the normal distribution.
    # Get upper and lower cdf values
    l = norm_cdf((a - mean) / std)
    u = norm_cdf((b - mean) / std)

    # Uniformly fill tensor with values from [l, u], then translate to
    # [2l-1, 2u-1].
    tensor.uniform_(2 * l - 1, 2 * u - 1)

    # Use inverse cdf transform for normal distribution to get truncated
    # standard normal
    if tensor.dtype in [torch.float16, torch.bfloat16]:
        # The `erfinv_` op is not (yet?) defined in float16+cpu, bfloat16+gpu
        og_dtype = tensor.dtype
        tensor = tensor.to(torch.float32)
        tensor.erfinv_()
        tensor = tensor.to(og_dtype)
    else:
        tensor.erfinv_()

    # Transform to proper mean, std
    tensor.mul_(std * math.sqrt(2.0))
    tensor.add_(mean)

    # Clamp to ensure it's in the proper range
    if tensor.dtype == torch.float16:
        # The `clamp_` op is not (yet?) defined in float16+cpu
        tensor = tensor.to(torch.float32)
        tensor.clamp_(min=a, max=b)
        tensor = tensor.to(torch.float16)
    else:
        tensor.clamp_(min=a, max=b)


def trunc_normal_tf_(
    tensor: torch.Tensor, mean: float = 0.0, std: float = 1.0, a: float = -2.0, b: float = 2.0
) -> torch.Tensor:
    """Fills the input Tensor with values drawn from a truncated
    normal distribution. The values are effectively drawn from the
    normal distribution :math:`\\mathcal{N}(\text{mean}, \text{std}^2)`
    with values outside :math:`[a, b]` redrawn until they are within
    the bounds. The method used for generating the random values works
    best when :math:`a \\leq \text{mean} \\leq b`.
    NOTE: this 'tf' variant behaves closer to Tensorflow / JAX impl where the
    bounds [a, b] are applied when sampling the normal distribution with mean=0, std=1.0
    and the result is subsquently scaled and shifted by the mean and std args.
    Args:
        tensor: an n-dimensional `torch.Tensor`
        mean: the mean of the normal distribution
        std: the standard deviation of the normal distribution
        a: the minimum cutoff value
        b: the maximum cutoff value
    """
    with torch.no_grad():
        _trunc_normal_(tensor, 0, 1.0, a, b)
        tensor.mul_(std).add_(mean)


def variance_scaling_(tensor, scale=1.0, mode="fan_in", distribution="normal"):
    fan_in, fan_out = _calculate_fan_in_and_fan_out(tensor)
    if mode == "fan_in":
        denom = fan_in
    elif mode == "fan_out":
        denom = fan_out
    elif mode == "fan_avg":
        denom = (fan_in + fan_out) / 2

    variance = scale / denom

    if distribution == "truncated_normal":
        # constant is stddev of standard normal truncated to (-2, 2)
        trunc_normal_tf_(tensor, std=math.sqrt(variance) / 0.87962566103423978)
    elif distribution == "normal":
        with torch.no_grad():
            tensor.normal_(std=math.sqrt(variance))
    elif distribution == "uniform":
        bound = math.sqrt(3 * variance)
        with torch.no_grad():
            tensor.uniform_(-bound, bound)
    else:
        raise ValueError(f"invalid distribution {distribution}")


def lecun_normal_(tensor):
    variance_scaling_(tensor, mode="fan_in", distribution="truncated_normal")


def default_flax_embed_init(tensor):
    variance_scaling_(tensor, mode="fan_in", distribution="normal")


@dataclass
# Copied from transformers.models.clip.modeling_clip.CLIPVisionModelOutput with CLIP->Siglip
class SiglipVisionModelOutput(ModelOutput):
    """
    Base class for vision model's outputs that also contains image embeddings of the pooling of the last hidden states.
    Args:
        image_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim)` *optional* returned when model is initialized with `with_projection=True`):
            The image embeddings obtained by applying the projection layer to the pooler_output.
        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.
        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
            Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
        attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
            sequence_length)`.
            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
            heads.
    """

    image_embeds: Optional[torch.FloatTensor] = None
    last_hidden_state: torch.FloatTensor = None
    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
    attentions: Optional[Tuple[torch.FloatTensor]] = None


class SiglipVisionEmbeddings(nn.Module):
    def __init__(self, config: SiglipVisionConfig):
        super().__init__()
        self.config = config
        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, tgt_sizes: Optional[torch.IntTensor]=None) -> torch.Tensor:
        batch_size = pixel_values.size(0)

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

        max_im_h, max_im_w = pixel_values.size(2), pixel_values.size(3)
        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)
        position_ids = torch.full(
            size=(
                batch_size,
                max_nb_patches_h * max_nb_patches_w,
            ),
            fill_value=0,
        )

        for batch_idx, p_attn_mask in enumerate(patch_attention_mask):
            if tgt_sizes is not None:
                nb_patches_h = tgt_sizes[batch_idx][0]
                nb_patches_w = tgt_sizes[batch_idx][1]
            else:
                nb_patches_h = p_attn_mask[:, 0].sum()
                nb_patches_w = p_attn_mask[0].sum()

            fractional_coords_h = torch.arange(0, 1 - 1e-6, 1 / nb_patches_h)
            fractional_coords_w = torch.arange(0, 1 - 1e-6, 1 / nb_patches_w)

            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).flatten()
            position_ids[batch_idx][p_attn_mask.view(-1).cpu()] = pos_ids

        position_ids = position_ids.to(self.position_embedding.weight.device)

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


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

    # Copied from transformers.models.clip.modeling_clip.CLIPAttention.__init__
    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)

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        output_attentions: Optional[bool] = False,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        """Input shape: Batch x Time x Channel"""

        batch_size, q_len, _ = hidden_states.size()

        query_states = self.q_proj(hidden_states)
        key_states = self.k_proj(hidden_states)
        value_states = self.v_proj(hidden_states)

        query_states = query_states.view(batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)
        key_states = key_states.view(batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)
        value_states = value_states.view(batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)

        k_v_seq_len = key_states.shape[-2]
        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) * self.scale

        if attn_weights.size() != (batch_size, self.num_heads, q_len, k_v_seq_len):
            raise ValueError(
                f"Attention weights should be of size {(batch_size, self.num_heads, q_len, k_v_seq_len)}, but is"
                f" {attn_weights.size()}"
            )

        if attention_mask is not None:
            if attention_mask.size() != (batch_size, 1, q_len, k_v_seq_len):
                raise ValueError(
                    f"Attention mask should be of size {(batch_size, 1, q_len, k_v_seq_len)}, but is {attention_mask.size()}"
                )
            attn_weights = attn_weights + attention_mask

        # upcast attention to fp32
        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
        attn_weights = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)
        attn_output = torch.matmul(attn_weights, value_states)

        if attn_output.size() != (batch_size, self.num_heads, q_len, self.head_dim):
            raise ValueError(
                f"`attn_output` should be of size {(batch_size, self.num_heads, q_len, self.head_dim)}, but is"
                f" {attn_output.size()}"
            )

        attn_output = attn_output.transpose(1, 2).contiguous()
        attn_output = attn_output.reshape(batch_size, q_len, self.embed_dim)

        attn_output = self.out_proj(attn_output)

        return attn_output, attn_weights


# Copied from transformers.models.clip.modeling_clip.CLIPMLP with CLIP->Siglip
class SiglipMLP(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


# Copied from transformers.models.clip.modeling_clip.CLIPEncoderLayer with CLIP->Siglip
class SiglipEncoderLayer(nn.Module):
    def __init__(self, config: SiglipVisionConfig):
        super().__init__()
        self.embed_dim = config.hidden_size
        self._use_flash_attention_2 = False
        self.self_attn = SiglipAttention(config)
        self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
        self.mlp = SiglipMLP(config)
        self.layer_norm2 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.Tensor,
        output_attentions: Optional[bool] = False,
    ) -> Tuple[torch.FloatTensor]:
        """
        Args:
            hidden_states (`torch.FloatTensor`):
                Input to the layer of shape `(batch, seq_len, embed_dim)`.
            attention_mask (`torch.FloatTensor`):
                Attention mask of shape `(batch, 1, q_len, k_v_seq_len)` where padding elements are indicated by very large negative values.
            output_attentions (`bool`, *optional*, defaults to `False`):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
        """
        residual = hidden_states

        hidden_states = self.layer_norm1(hidden_states)
        hidden_states, attn_weights = self.self_attn(
            hidden_states=hidden_states,
            attention_mask=attention_mask,
            output_attentions=output_attentions,
        )
        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

        outputs = (hidden_states,)

        if output_attentions:
            outputs += (attn_weights,)

        return outputs


class SiglipPreTrainedModel(PreTrainedModel):
    """
    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
    models.
    """

    config_class = SiglipVisionConfig
    base_model_prefix = "siglip"
    supports_gradient_checkpointing = True

    def _init_weights(self, module):
        """Initialize the weights"""

        if isinstance(module, SiglipVisionEmbeddings):
            width = self.config.hidden_size
            nn.init.normal_(module.position_embedding.weight, std=1 / np.sqrt(width))
        elif isinstance(module, nn.Embedding):
            default_flax_embed_init(module.weight)
        elif isinstance(module, SiglipAttention):
            nn.init.normal_(module.q_proj.weight)
            nn.init.normal_(module.k_proj.weight)
            nn.init.normal_(module.v_proj.weight)
            nn.init.normal_(module.out_proj.weight)
            nn.init.zeros_(module.q_proj.bias)
            nn.init.zeros_(module.k_proj.bias)
            nn.init.zeros_(module.v_proj.bias)
            nn.init.zeros_(module.out_proj.bias)
        elif isinstance(module, SiglipMLP):
            nn.init.normal_(module.fc1.weight)
            nn.init.normal_(module.fc2.weight)
            nn.init.normal_(module.fc1.bias, std=1e-6)
            nn.init.normal_(module.fc2.bias, std=1e-6)
        elif isinstance(module, (nn.Linear, nn.Conv2d)):
            lecun_normal_(module.weight)
            if module.bias is not None:
                nn.init.zeros_(module.bias)
        elif isinstance(module, nn.LayerNorm):
            module.bias.data.zero_()
            module.weight.data.fill_(1.0)


SIGLIP_START_DOCSTRING = r"""
    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
    etc.)
    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
    and behavior.
    Parameters:
        config ([`SiglipVisionConfig`]): Model configuration class with all the parameters of the model.
            Initializing with a config file does not load the weights associated with the model, only the
            configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
"""


SIGLIP_VISION_INPUTS_DOCSTRING = r"""
    Args:
        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
            Pixel values. Padding will be ignored by default should you provide it. Pixel values can be obtained using
            [`AutoImageProcessor`]. See [`CLIPImageProcessor.__call__`] for details.
        output_attentions (`bool`, *optional*):
            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
            tensors for more detail.
        output_hidden_states (`bool`, *optional*):
            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
            more detail.
        return_dict (`bool`, *optional*):
            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
"""


# Copied from transformers.models.clip.modeling_clip.CLIPEncoder with CLIP->Siglip
class SiglipEncoder(nn.Module):
    """
    Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a
    [`SiglipEncoderLayer`].
    Args:
        config: SiglipConfig
    """

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

    # Ignore copy
    def forward(
        self,
        inputs_embeds,
        attention_mask: Optional[torch.Tensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, BaseModelOutput]:
        r"""
        Args:
            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
                This is useful if you want more control over how to convert `input_ids` indices into associated vectors
                than the model's internal embedding lookup matrix.
            attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
                Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
                - 1 for tokens that are **not masked**,
                - 0 for tokens that are **masked**.
                [What are attention masks?](../glossary#attention-mask)
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
            output_hidden_states (`bool`, *optional*):
                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
                for more detail.
            return_dict (`bool`, *optional*):
                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
        """
        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

        encoder_states = () if output_hidden_states else None
        all_attentions = () if output_attentions else None

        hidden_states = inputs_embeds
        for encoder_layer in self.layers:
            if output_hidden_states:
                encoder_states = encoder_states + (hidden_states,)
            if self.gradient_checkpointing and self.training:
                layer_outputs = self._gradient_checkpointing_func(
                    encoder_layer.__call__,
                    hidden_states,
                    attention_mask,
                    output_attentions,
                )
            else:
                layer_outputs = encoder_layer(
                    hidden_states,
                    attention_mask,
                    output_attentions=output_attentions,
                )

            hidden_states = layer_outputs[0]

            if output_attentions:
                all_attentions = all_attentions + (layer_outputs[1],)

        if output_hidden_states:
            encoder_states = encoder_states + (hidden_states,)

        if not return_dict:
            return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None)
        return BaseModelOutput(
            last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions
        )

@add_start_docstrings(
    """The vision model from SigLIP without any head or projection on top.""",
    SIGLIP_START_DOCSTRING
)
class SiglipVisionTransformer(SiglipPreTrainedModel):
    config_class = SiglipVisionConfig
    main_input_name = "pixel_values"
    _supports_flash_attn_2 = True

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

        self.embeddings = SiglipVisionEmbeddings(config)
        self.encoder = SiglipEncoder(config)
        self.post_layernorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
        self._use_flash_attention_2 = False

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

    def get_input_embeddings(self) -> nn.Module:
        return self.embeddings.patch_embedding

    @add_start_docstrings_to_model_forward(SIGLIP_VISION_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=SiglipVisionConfig)
    def forward(
        self,
        pixel_values,
        patch_attention_mask: Optional[torch.BoolTensor] = None,
        tgt_sizes: Optional[torch.IntTensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, BaseModelOutputWithPooling]:
        r"""
        Returns:
        """
        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

        batch_size = pixel_values.size(0)
        if patch_attention_mask is None:
            patch_attention_mask = torch.ones(
                size=(
                    batch_size,
                    pixel_values.size(2) // self.config.patch_size,
                    pixel_values.size(3) // self.config.patch_size,
                ),
                dtype=torch.bool,
                device=pixel_values.device,
            )

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

        patch_attention_mask = patch_attention_mask.view(batch_size, -1)
        # The call to `_upad_input` in `_flash_attention_forward` is expensive
        # So when the `patch_attention_mask` is full of 1s (i.e. attending to the whole sequence),
        # avoiding passing the attention_mask, which is equivalent to attending to the full sequence
        if not torch.any(~patch_attention_mask):
            attention_mask=None
        else:
            attention_mask = _prepare_4d_attention_mask(patch_attention_mask, hidden_states.dtype)

        encoder_outputs = self.encoder(
            inputs_embeds=hidden_states,
            attention_mask=attention_mask,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        last_hidden_state = encoder_outputs[0]
        last_hidden_state = self.post_layernorm(last_hidden_state)

        if not return_dict:
            return (last_hidden_state, None) + encoder_outputs[1:]

        return BaseModelOutputWithPooling(
            last_hidden_state=last_hidden_state,
            pooler_output=None,
            hidden_states=encoder_outputs.hidden_states,
            attentions=encoder_outputs.attentions,
        )