modeling_siglip2.py

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

import os
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
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_layers import GradientCheckpointingLayer
from transformers.modeling_outputs import BaseModelOutput, BaseModelOutputWithPooling, ImageClassifierOutput
from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
from transformers.utils import (
    ModelOutput,
    add_start_docstrings,
    add_start_docstrings_to_model_forward,
    can_return_tuple,
    logging,
    replace_return_docstrings,
    is_flash_attn_2_available,
    is_flash_attn_greater_or_equal_2_10,
)

from .configuration_siglip2 import Siglip2Config, Siglip2TextConfig, Siglip2VisionConfig


logger = logging.get_logger(__name__)

_CONFIG_FOR_DOC = "Siglip2Config"

is_aiter_available = False

if is_flash_attn_2_available():
    try:
        from aiter import flash_attn_varlen_func
        is_aiter_available = True
    except ImportError:
        from flash_attn import flash_attn_varlen_func
    from flash_attn.layers.rotary import apply_rotary_emb

else:
    flash_attn_varlen_func = None
    apply_rotary_emb = None


if is_flash_attn_2_available():
    from transformers.modeling_flash_attention_utils import _flash_attention_forward
else:
    flash_attn_varlen_func = None

@dataclass
class Siglip2VisionOutput(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: Optional[torch.FloatTensor] = None
    hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
    attentions: Optional[Tuple[torch.FloatTensor, ...]] = None


@dataclass
class Siglip2TextOutput(ModelOutput):
    """
    Base class for text model's outputs that also contains a pooling of the last hidden states.

    Args:
        text_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim)` 
        *optional* returned when model is initialized with `with_projection=True`):
            The text 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.
    """

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


@dataclass
class Siglip2Output(ModelOutput):
    """
    Args:
        loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `return_loss` is `True`):
            Contrastive loss for image-text similarity.
        logits_per_image (`torch.FloatTensor` of shape `(image_batch_size, text_batch_size)`):
            The scaled dot product scores between `image_embeds` and `text_embeds`. This represents the image-text
            similarity scores.
        logits_per_text (`torch.FloatTensor` of shape `(text_batch_size, image_batch_size)`):
            The scaled dot product scores between `text_embeds` and `image_embeds`. This represents the text-image
            similarity scores.
        text_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim`):
            The text embeddings obtained by applying the projection layer to the pooled output of [`Siglip2TextModel`].
        image_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim`):
            The image embeddings obtained by applying the projection layer to 
            the pooled output of [`Siglip2VisionModel`].
        text_model_output (`BaseModelOutputWithPooling`):
            The output of the [`Siglip2TextModel`].
        vision_model_output (`BaseModelOutputWithPooling`):
            The output of the [`Siglip2VisionModel`].
    """

    loss: Optional[torch.FloatTensor] = None
    logits_per_image: Optional[torch.FloatTensor] = None
    logits_per_text: Optional[torch.FloatTensor] = None
    text_embeds: Optional[torch.FloatTensor] = None
    image_embeds: Optional[torch.FloatTensor] = None
    text_model_output: BaseModelOutputWithPooling = None
    vision_model_output: BaseModelOutputWithPooling = None

    def to_tuple(self) -> Tuple[Any]:
        return tuple(
            self[k] if k not in ["text_model_output", "vision_model_output"] else getattr(self, k).to_tuple()
            for k in self.keys()
        )

class Siglip2VisionEmbeddings(nn.Module):
    def __init__(self, config: Siglip2VisionConfig):
        super().__init__()
        self.config = config
        self.embed_dim = config.hidden_size
        self.patch_size = config.patch_size

        if hasattr(config, 'in_features') and config.in_features > 0:
            self.in_features = config.in_features
        else:
            self.in_features = config.num_channels * self.patch_size * self.patch_size

        self.patch_embedding = nn.Linear(
            in_features=self.in_features,
            out_features=self.embed_dim,
        )

        self.num_patches = config.num_patches
        self.position_embedding_size = int(self.num_patches**0.5)
        self.position_embedding = nn.Embedding(self.num_patches, self.embed_dim)

    @staticmethod
    def resize_positional_embeddings(
        positional_embeddings: torch.Tensor,
        spatial_shapes: torch.LongTensor,
        max_length: int,
    ) -> torch.Tensor:
        """
        Resize positional embeddings to image-specific size and pad to a fixed size.

        Args:
            positional_embeddings (`torch.Tensor`):
                Position embeddings of shape (height, width, embed_dim)
            spatial_shapes (`torch.LongTensor`):
                Spatial shapes of shape (batch_size, 2) to resize the positional embeddings to
            max_length (`int`):
                Maximum length of the positional embeddings to pad resized positional embeddings to

        Returns:
            `torch.Tensor`: Embeddings of shape (batch_size, max_length, embed_dim)
        """
        batch_size = spatial_shapes.shape[0]
        embed_dim = positional_embeddings.shape[-1]
        source_dtype = positional_embeddings.dtype

        resulted_positional_embeddings = torch.empty(
            (batch_size, max_length, embed_dim),
            device=positional_embeddings.device,
            dtype=source_dtype,
        )

        positional_embeddings = positional_embeddings.permute(2, 0, 1).unsqueeze(0)
        if positional_embeddings.device.type == "cpu":
            positional_embeddings = positional_embeddings.to(torch.float32)
        for i in range(batch_size):
            height, width = spatial_shapes[i]
            resized_embeddings = F.interpolate(
                positional_embeddings,
                size=(height, width),
                mode="bilinear",
                align_corners=False,
                antialias=True,
            )
            resized_embeddings = resized_embeddings.reshape(embed_dim, height * width).transpose(0, 1)
            resized_embeddings = resized_embeddings.to(source_dtype)

            resulted_positional_embeddings[i, : height * width] = resized_embeddings
            resulted_positional_embeddings[i, height * width :] = resized_embeddings[0]

        return resulted_positional_embeddings


    def forward(self, pixel_values: torch.FloatTensor, spatial_shapes: torch.LongTensor) -> torch.Tensor:
        """
        Args:
            pixel_values (`torch.FloatTensor`):
                Pixel values of shape (batch_size, max_num_patches, num_channels * patch_size * patch_size)
            spatial_shapes (`List[Tuple[int, int]]`):
                Spatial shapes of shape (batch_size, 2) to resize the positional embeddings to
        """

        target_dtype = self.patch_embedding.weight.dtype
        patch_embeds = self.patch_embedding(pixel_values.to(dtype=target_dtype))
        positional_embeddings = self.position_embedding.weight.reshape(
            self.position_embedding_size, self.position_embedding_size, -1
        )

        resized_positional_embeddings = self.resize_positional_embeddings(
            positional_embeddings, spatial_shapes, max_length=pixel_values.shape[1]
        )
        embeddings = patch_embeds + resized_positional_embeddings
        return embeddings


class Siglip2VisionEmbeddingsWoPos(nn.Module):
    def __init__(self, config: Siglip2VisionConfig):
        super().__init__()
        self.config = config
        self.embed_dim = config.hidden_size
        self.patch_size = config.patch_size

        if hasattr(config, 'in_features') and config.in_features > 0:
            self.in_features = config.in_features
        else:
            self.in_features = config.num_channels * self.patch_size * self.patch_size

        self.patch_embedding = nn.Linear(
            in_features=self.in_features,
            out_features=self.embed_dim,
        )

    def forward(self, pixel_values: torch.FloatTensor) -> torch.Tensor:
        target_dtype = self.patch_embedding.weight.dtype
        patch_embeds = self.patch_embedding(pixel_values.to(dtype=target_dtype))
        patch_embeds = patch_embeds.view(-1, self.embed_dim)
        return patch_embeds


def eager_attention_forward(
    module: nn.Module,
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    attention_mask: Optional[torch.Tensor],
    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

def apply_rotary_pos_emb_flashatt(
        q: torch.Tensor, k: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor
    ) -> Tuple[torch.Tensor, torch.Tensor]:
    cos = cos.chunk(2, dim=-1)[0].contiguous()
    sin = sin.chunk(2, dim=-1)[0].contiguous()
    q_embed = apply_rotary_emb(q.float(), cos.float(), sin.float()).type_as(q)
    k_embed = apply_rotary_emb(k.float(), cos.float(), sin.float()).type_as(k)
    return q_embed, k_embed

class Siglip2Attention(nn.Module):
    """Multi-headed attention from 'Attention Is All You Need' paper"""

    def __init__(self, config: Union[Siglip2VisionConfig, Siglip2TextConfig]):
        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.is_causal = False

        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]]:
        """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 = eager_attention_forward
        if self.config._attn_implementation != "eager":
            if self.config._attn_implementation == "sdpa" and output_attentions:
                logger.warning_once(
                    "`torch.nn.functional.scaled_dot_product_attention` does not support"
                    "`output_attentions=True`. Falling back to 'eager attention. This warning"
                    'can be removed using the argument `attn_implementation="eager"` when loading the model.'
                )
            else:
                attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]

        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)

        if not output_attentions:
            attn_weights = None

        return attn_output, attn_weights

class Vision_FlashAttention2(nn.Module):
    def __init__(self, config: Union[Siglip2VisionConfig, Siglip2TextConfig]) -> None:
        super().__init__()
        dim = config.hidden_size
        self.num_heads = config.num_attention_heads
        self.k_proj = nn.Linear(dim, dim)
        self.v_proj = nn.Linear(dim, dim)
        self.q_proj = nn.Linear(dim, dim)
        self.out_proj = nn.Linear(dim, dim)

    def forward(
        self,
        hidden_states: torch.Tensor,
        cu_seqlens: torch.Tensor,
        rotary_pos_emb: Optional[torch.Tensor] = None,
        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
    ) -> torch.Tensor:

        seq_length = hidden_states.shape[0]
        q = self.q_proj(hidden_states).reshape(seq_length, self.num_heads, -1)
        k = self.k_proj(hidden_states).reshape(seq_length, self.num_heads, -1)
        v = self.v_proj(hidden_states).reshape(seq_length, self.num_heads, -1)
        

        if position_embeddings is None:
            logger.warning_once(
                "The attention layers in this model are transitioning from computing the RoPE embeddings internally "
                "through `rotary_pos_emb` (2D tensor of RoPE theta values), to using externally computed "
                "`position_embeddings` (Tuple of tensors, containing cos and sin). In v4.54 `rotary_pos_emb` will be "
                "removed and `position_embeddings` will be mandatory."
            )
            emb = torch.cat((rotary_pos_emb, rotary_pos_emb), dim=-1)
            cos = emb.cos()
            sin = emb.sin()
        else:
            cos, sin = position_embeddings
        
        q, k = apply_rotary_pos_emb_flashatt(q.unsqueeze(0), k.unsqueeze(0), cos, sin)
        q = q.squeeze(0)
        k = k.squeeze(0)

        max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max().item()
        if is_aiter_available:
            attn_output = flash_attn_varlen_func(q, k, v, cu_seqlens, cu_seqlens,
                            max_seqlen, max_seqlen, return_lse=True)[0].reshape(
                            seq_length, -1)
        else:
            attn_output = flash_attn_varlen_func(q, k, v, cu_seqlens, cu_seqlens,
                            max_seqlen, max_seqlen).reshape(
                            seq_length, -1)   
        attn_output = self.out_proj(attn_output)
        return attn_output, None



def rotate_half(x):
    """Rotates half the hidden dims of the input."""
    x1 = x[..., : x.shape[-1] // 2]
    x2 = x[..., x.shape[-1] // 2 :]
    return torch.cat((-x2, x1), dim=-1)


def apply_rotary_pos_emb_vision(
    q: torch.Tensor, k: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor
) -> Tuple[torch.Tensor, torch.Tensor]:
    orig_q_dtype = q.dtype
    orig_k_dtype = k.dtype
    q, k = q.float(), k.float()
    cos, sin = cos.unsqueeze(-2).float(), sin.unsqueeze(-2).float()
    q_embed = (q * cos) + (rotate_half(q) * sin)
    k_embed = (k * cos) + (rotate_half(k) * sin)
    q_embed = q_embed.to(orig_q_dtype)
    k_embed = k_embed.to(orig_k_dtype)
    return q_embed, k_embed


class Vision_EagerAttention(nn.Module):
    def __init__(self, config: Union[Siglip2VisionConfig, Siglip2TextConfig]) -> None:
        super().__init__()
        dim = config.hidden_size
        self.num_heads = config.num_attention_heads
        self.k_proj = nn.Linear(dim, dim)
        self.v_proj = nn.Linear(dim, dim)
        self.q_proj = nn.Linear(dim, dim)
        self.out_proj = nn.Linear(dim, dim)
        self.head_dim = dim // self.num_heads

    def forward(
        self,
        hidden_states: torch.Tensor,
        cu_seqlens: torch.Tensor,
        rotary_pos_emb: Optional[torch.Tensor] = None,
        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
    ) -> torch.Tensor:
        seq_length = hidden_states.shape[0]
        q = self.q_proj(hidden_states).reshape(seq_length, self.num_heads, -1)
        k = self.k_proj(hidden_states).reshape(seq_length, self.num_heads, -1)
        v = self.v_proj(hidden_states).reshape(seq_length, self.num_heads, -1)
        
        if position_embeddings is None:
            logger.warning_once(
                "The attention layers in this model are transitioning from computing the RoPE embeddings internally "
                "through `rotary_pos_emb` (2D tensor of RoPE theta values), to using externally computed "
                "`position_embeddings` (Tuple of tensors, containing cos and sin). In v4.54 `rotary_pos_emb` will be "
                "removed and `position_embeddings` will be mandatory."
            )
            emb = torch.cat((rotary_pos_emb, rotary_pos_emb), dim=-1)
            cos = emb.cos()
            sin = emb.sin()
        else:
            cos, sin = position_embeddings
        q, k = apply_rotary_pos_emb_vision(q, k, cos, sin)

        attention_mask = torch.full(
            [1, seq_length, seq_length], torch.finfo(q.dtype).min, device=q.device, dtype=q.dtype
        )
        for i in range(1, len(cu_seqlens)):
            attention_mask[..., cu_seqlens[i - 1] : cu_seqlens[i], cu_seqlens[i - 1] : cu_seqlens[i]] = 0

        q = q.transpose(0, 1)
        k = k.transpose(0, 1)
        v = v.transpose(0, 1)
        attn_weights = torch.matmul(q, k.transpose(1, 2)) / math.sqrt(self.head_dim)
        attn_weights = attn_weights + attention_mask
        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(q.dtype)
        attn_output = torch.matmul(attn_weights, v)
        attn_output = attn_output.transpose(0, 1)
        attn_output = attn_output.reshape(seq_length, -1)
        attn_output = self.out_proj(attn_output)
        return attn_output, None


VISION_ATTENTION_CLASSES = {
    'eager': Vision_EagerAttention,
    'flash_attention_2': Vision_FlashAttention2,
}


class Siglip2MLP(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 Siglip2EncoderLayer(GradientCheckpointingLayer):
class Siglip2EncoderLayer(nn.Module):
    def __init__(self, config: Union[Siglip2VisionConfig, Siglip2TextConfig]):
        super().__init__()
        self.embed_dim = config.hidden_size
        self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
        self.self_attn = VISION_ATTENTION_CLASSES[config._attn_implementation](config=config)
        self.layer_norm2 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
        self.mlp = Siglip2MLP(config)

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.Tensor,
        cu_seqlens: torch.Tensor,
        rotary_pos_emb: Optional[torch.Tensor] = None,
        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
        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,
            cu_seqlens=cu_seqlens,
            rotary_pos_emb=rotary_pos_emb,
            position_embeddings=position_embeddings,
        )
        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 VisionRope(nn.Module):
    def __init__(self, dim: int, theta: float = 10000.0) -> None:
        super().__init__()
        inv_freq = 1.0 / (theta ** (torch.arange(0, dim, 2, dtype=torch.float) / dim))
        self.register_buffer("inv_freq", inv_freq, persistent=False)

    def forward(self, seqlen: int) -> torch.Tensor:
        seq = torch.arange(seqlen, device=self.inv_freq.device, dtype=self.inv_freq.dtype)
        freqs = torch.outer(seq, self.inv_freq)
        return freqs

class Siglip2Encoder(nn.Module):
    """
    Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a
    [`Siglip2EncoderLayer`].

    Args:
        config: Siglip2Config
    """

    def __init__(self, config: Siglip2Config):
        super().__init__()
        self.config = config
        self.layers = nn.ModuleList([Siglip2EncoderLayer(config) for _ in range(config.num_hidden_layers)])
        self.gradient_checkpointing = False
        self.spatial_merge_size = 2
        self.spatial_merge_unit = self.spatial_merge_size * self.spatial_merge_size
        self.patch_size = config.patch_size
        self.window_size = self.patch_size * 2 * 8
        
        assert(config.hidden_size%(config.num_attention_heads*2) == 0)
        self.rotary_pos_emb = VisionRope(config.hidden_size//config.num_attention_heads//2)

    def rot_pos_emb(self, spatial_shapes):
        pos_ids = []

        for h, w in spatial_shapes:
            t = 1
            hpos_ids = torch.arange(h).unsqueeze(1).expand(-1, w)
            
            hpos_ids = hpos_ids.reshape(
                h // self.spatial_merge_size,
                self.spatial_merge_size,
                w // self.spatial_merge_size,
                self.spatial_merge_size,
            )
            hpos_ids = hpos_ids.permute(0, 2, 1, 3)
            hpos_ids = hpos_ids.flatten()

            wpos_ids = torch.arange(w).unsqueeze(0).expand(h, -1)
            wpos_ids = wpos_ids.reshape(
                h // self.spatial_merge_size,
                self.spatial_merge_size,
                w // self.spatial_merge_size,
                self.spatial_merge_size,
            )
            wpos_ids = wpos_ids.permute(0, 2, 1, 3)
            wpos_ids = wpos_ids.flatten()

            pos_ids.append(torch.stack([hpos_ids, wpos_ids], dim=-1).repeat(t, 1))
        pos_ids = torch.cat(pos_ids, dim=0)
        max_grid_size = spatial_shapes.max()
        rotary_pos_emb_full = self.rotary_pos_emb(max_grid_size)
        rotary_pos_emb = rotary_pos_emb_full[pos_ids].flatten(1)
        return rotary_pos_emb

    def get_window_index(self, spatial_shapes):
        window_index: list = []
        cu_window_seqlens: list = [0]
        window_index_id = 0
        vit_merger_window_size = self.window_size // self.spatial_merge_size // self.patch_size

        for grid_h, grid_w in spatial_shapes:
            grid_t = 1
            llm_grid_h, llm_grid_w = (
                grid_h // self.spatial_merge_size,
                grid_w // self.spatial_merge_size,
            )
            index = torch.arange(grid_t * llm_grid_h * llm_grid_w).reshape(grid_t, llm_grid_h, llm_grid_w)
            pad_h = (vit_merger_window_size - llm_grid_h % vit_merger_window_size) % vit_merger_window_size
            pad_w = (vit_merger_window_size - llm_grid_w % vit_merger_window_size) % vit_merger_window_size
            num_windows_h = (llm_grid_h + pad_h) // vit_merger_window_size
            num_windows_w = (llm_grid_w + pad_w) // vit_merger_window_size
            index_padded = F.pad(index, (0, pad_w, 0, pad_h), "constant", -100)
            index_padded = index_padded.reshape(
                grid_t,
                num_windows_h,
                vit_merger_window_size,
                num_windows_w,
                vit_merger_window_size,
            )

            index_padded = index_padded.permute(0, 1, 3, 2, 4).reshape(
                grid_t,
                num_windows_h * num_windows_w,
                vit_merger_window_size,
                vit_merger_window_size,
            )
            seqlens = (index_padded != -100).sum([2, 3]).reshape(-1)
            index_padded = index_padded.reshape(-1)
            index_new = index_padded[index_padded != -100]
            window_index.append(index_new + window_index_id)
            cu_seqlens_tmp = seqlens.cumsum(0) * self.spatial_merge_unit + cu_window_seqlens[-1]
            cu_window_seqlens.extend(cu_seqlens_tmp.tolist())
            window_index_id += (grid_t * llm_grid_h * llm_grid_w).item()

        window_index = torch.cat(window_index, dim=0)

        return window_index, cu_window_seqlens

    @can_return_tuple
    def forward(
        self,
        inputs_embeds,
        spatial_shapes: torch.LongTensor,
        attention_mask: Optional[torch.Tensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
    ) -> 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
        )

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

        hidden_states = inputs_embeds
        rotary_pos_emb = self.rot_pos_emb(spatial_shapes)
        window_index, cu_window_seqlens = self.get_window_index(spatial_shapes)
        cu_window_seqlens = torch.tensor(
            cu_window_seqlens,
            device=hidden_states.device,
            dtype=spatial_shapes.dtype if torch.jit.is_tracing() else torch.int32,
        )
        cu_window_seqlens = torch.unique_consecutive(cu_window_seqlens)
        
        seq_len, _ = hidden_states.size()
        hidden_states = hidden_states.reshape(seq_len // self.spatial_merge_unit, self.spatial_merge_unit, -1)
        hidden_states = hidden_states[window_index, :, :]
        hidden_states = hidden_states.reshape(seq_len, -1)
        rotary_pos_emb = rotary_pos_emb.reshape(seq_len // self.spatial_merge_unit, self.spatial_merge_unit, -1)
        rotary_pos_emb = rotary_pos_emb[window_index, :, :]
        rotary_pos_emb = rotary_pos_emb.reshape(seq_len, -1)
        emb = torch.cat((rotary_pos_emb, rotary_pos_emb), dim=-1)
        position_embeddings = (emb.cos(), emb.sin())

        cu_seqlens = torch.repeat_interleave(spatial_shapes[:, 0] * spatial_shapes[:, 1], 1).cumsum(
            dim=0,
            dtype=spatial_shapes.dtype if torch.jit.is_tracing() else torch.int32,
        )
        cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0)

        for layer_num, encoder_layer in enumerate(self.layers):
            if output_hidden_states:
                encoder_states = encoder_states + (hidden_states,)

            if (1+layer_num) % 8 == 0 or layer_num == len(self.layers) - 1:
                cu_seqlens_now = cu_seqlens
            else:
                cu_seqlens_now = cu_window_seqlens

            layer_outputs = encoder_layer(
                hidden_states,
                attention_mask,
                cu_seqlens=cu_seqlens_now, 
                position_embeddings=position_embeddings
            )

            hidden_states = layer_outputs[0]
            if output_attentions:
                all_attentions = all_attentions + (layer_outputs[1],)


        hidden_states = hidden_states.reshape(seq_len // self.spatial_merge_unit, self.spatial_merge_unit, -1)
        reverse_indices = torch.argsort(window_index)
        hidden_states = hidden_states[reverse_indices, :, :]
        hidden_states = hidden_states.reshape(seq_len, -1)
        
        if output_hidden_states:
            encoder_states = encoder_states + (hidden_states,)

        return BaseModelOutput(
            last_hidden_state=hidden_states,
            hidden_states=encoder_states,
            attentions=all_attentions,
        )


SIGLIP2_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.
        interpolate_pos_encoding (`bool`, *optional*, defaults to `False`):
            Whether to interpolate the pre-trained position encodings.
        return_dict (`bool`, *optional*):
            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
"""


class Siglip2VisionTransformer(nn.Module):
    def __init__(self, config: Siglip2VisionConfig):
        super().__init__()
        self.config = config
        embed_dim = config.hidden_size

        self.embeddings = Siglip2VisionEmbeddingsWoPos(config)
        self.encoder = Siglip2Encoder(config)
        self.post_layernorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
        self.use_head = False
        if self.use_head:
            self.head = Siglip2MultiheadAttentionPoolingHead(config)
        self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"

    @can_return_tuple
    @add_start_docstrings_to_model_forward(SIGLIP2_VISION_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=Siglip2VisionConfig)
    def forward(
        self,
        pixel_values: torch.FloatTensor,
        attention_mask: torch.Tensor,
        spatial_shapes: torch.LongTensor,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
    ) -> 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
        )

        bs, length, dim = pixel_values.shape
        hidden_states = self.embeddings(pixel_values)

        if attention_mask is not None and not self._use_flash_attention_2:
            encoder_attention_mask = _prepare_4d_attention_mask(attention_mask, hidden_states.dtype)
        else:
            encoder_attention_mask = attention_mask

        encoder_outputs: BaseModelOutput = self.encoder(
            inputs_embeds=hidden_states,
            spatial_shapes=spatial_shapes,
            attention_mask=encoder_attention_mask,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
        )

        last_hidden_state = encoder_outputs.last_hidden_state

        last_hidden_state = self.post_layernorm(last_hidden_state)

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


class Siglip2TextEmbeddings(nn.Module):
    def __init__(self, config: Siglip2TextConfig):
        super().__init__()
        embed_dim = config.hidden_size

        self.token_embedding = nn.Embedding(config.vocab_size, embed_dim)
        self.position_embedding = nn.Embedding(config.max_position_embeddings, embed_dim)

        self.register_buffer(
            "position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)), persistent=False
        )

    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
    ) -> torch.Tensor:
        seq_length = input_ids.shape[-1] if input_ids is not None else inputs_embeds.shape[-2]
        max_position_embedding = self.position_embedding.weight.shape[0]
        if seq_length > max_position_embedding:
            raise ValueError(
                f"Sequence length must be less than max_position_embeddings (got `sequence length`: "
                f"{seq_length} and max_position_embeddings: {max_position_embedding}"
            )

        if position_ids is None:
            position_ids = self.position_ids[:, :seq_length]
 
        if inputs_embeds is None:
            inputs_embeds = self.token_embedding(input_ids)

        position_embeddings = self.position_embedding(position_ids)
        embeddings = inputs_embeds + position_embeddings

        return embeddings


def _trunc_normal_(tensor, mean, std, a, b):
    def norm_cdf(x):
        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,
        )

    l = norm_cdf((a - mean) / std)
    u = norm_cdf((b - mean) / std)

    tensor.uniform_(2 * l - 1, 2 * u - 1)

    tensor.erfinv_()

    tensor.mul_(std * math.sqrt(2.0))
    tensor.add_(mean)

    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:
    """
    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":
        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")


SIGLIP2_TEXT_INPUTS_DOCSTRING = r"""
    Args:
        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
            it.

            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
            [`PreTrainedTokenizer.__call__`] for details.

            [What are input IDs?](../glossary#input-ids)
        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)
        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
            config.max_position_embeddings - 1]`.

            [What are position IDs?](../glossary#position-ids)
        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.
"""


class Siglip2TextTransformer(nn.Module):
    def __init__(self, config: Siglip2TextConfig):
        super().__init__()
        self.config = config
        embed_dim = config.hidden_size
        self.embeddings = Siglip2TextEmbeddings(config)
        self.encoder = Siglip2Encoder(config)
        self.final_layer_norm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)

        self.head = nn.Linear(embed_dim, config.projection_size)
        self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"

    @can_return_tuple
    @add_start_docstrings_to_model_forward(SIGLIP2_TEXT_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=Siglip2TextConfig)
    def forward(
        self,
        input_ids: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.Tensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
    ) -> 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
        )

        if input_ids is None:
            raise ValueError("You have to specify input_ids")

        input_shape = input_ids.size()
        input_ids = input_ids.view(-1, input_shape[-1])

        hidden_states = self.embeddings(input_ids=input_ids, position_ids=position_ids)

        if attention_mask is not None and not self._use_flash_attention_2:
            attention_mask = _prepare_4d_attention_mask(attention_mask, hidden_states.dtype)

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

        last_hidden_state = encoder_outputs.last_hidden_state

        last_hidden_state = self.final_layer_norm(last_hidden_state)

        pooled_output = last_hidden_state[:, -1, :]
        pooled_output = self.head(pooled_output)

        return BaseModelOutputWithPooling(
            last_hidden_state=last_hidden_state,
            pooler_output=pooled_output,
            hidden_states=encoder_outputs.hidden_states,
            attentions=encoder_outputs.attentions,
        )


SIGLIP2_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 ([`Siglip2Config`]): 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.
"""

SIGLIP2_INPUTS_DOCSTRING = r"""
    Args:
        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
            it.

            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
            [`PreTrainedTokenizer.__call__`] for details.

            [What are input IDs?](../glossary#input-ids)
        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)
        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
            config.max_position_embeddings - 1]`.

            [What are position IDs?](../glossary#position-ids)
        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.
        return_loss (`bool`, *optional*):
            Whether or not to return the contrastive loss.
        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.
        interpolate_pos_encoding (`bool`, *optional*, defaults to `False`):
            Whether to interpolate the pre-trained position encodings.
        return_dict (`bool`, *optional*):
            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
"""


class Siglip2PreTrainedModel(PreTrainedModel):

    config_class = Siglip2Config
    base_model_prefix = "siglip2"
    supports_gradient_checkpointing = True

    _no_split_modules = [
        "Siglip2TextEmbeddings",
        "Siglip2EncoderLayer",
        "Siglip2VisionEmbeddings",
        "Siglip2EncoderLayer",
        "Siglip2MultiheadAttentionPoolingHead",
    ]
    _supports_flash_attn_2 = True
    _supports_sdpa = True

    def _init_weights(self, module):
        """Initialize the weights"""
        if isinstance(module, Siglip2VisionEmbeddings):
            width = (
                self.config.vision_config.hidden_size
                if isinstance(self.config, Siglip2Config)
                else 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, Siglip2Attention):
            nn.init.xavier_uniform_(module.q_proj.weight)
            nn.init.xavier_uniform_(module.k_proj.weight)
            nn.init.xavier_uniform_(module.v_proj.weight)
            nn.init.xavier_uniform_(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, Siglip2MLP):
            nn.init.xavier_uniform_(module.fc1.weight)
            nn.init.xavier_uniform_(module.fc2.weight)
            nn.init.normal_(module.fc1.bias, std=1e-6)
            nn.init.normal_(module.fc2.bias, std=1e-6)
        elif isinstance(module, Siglip2MultiheadAttentionPoolingHead):
            nn.init.xavier_uniform_(module.probe.data)
            nn.init.xavier_uniform_(module.attention.in_proj_weight.data)
            nn.init.zeros_(module.attention.in_proj_bias.data)
        elif isinstance(module, Siglip2Model):
            logit_scale_init = torch.log(torch.tensor(1.0))
            module.logit_scale.data.fill_(logit_scale_init)
            module.logit_bias.data.zero_()
        elif isinstance(module, Siglip2ForImageClassification):
            nn.init.normal_(
                module.classifier.weight,
                std=self.config.vision_config.hidden_size**-0.5 * self.config.initializer_factor,
            )
        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)


@add_start_docstrings(
    """The text model from Siglip2 without any head or projection on top.""",
    SIGLIP2_START_DOCSTRING,
)
class Siglip2TextModel(Siglip2PreTrainedModel):
    config_class = Siglip2TextConfig

    def __init__(self, config: Siglip2TextConfig):
        super().__init__(config)
        self.text_model = Siglip2TextTransformer(config)
        # Initialize weights and apply final processing
        self.post_init()

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

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

    @can_return_tuple
    @add_start_docstrings_to_model_forward(SIGLIP2_TEXT_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=Siglip2TextConfig)
    def forward(
        self,
        input_ids: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.Tensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
    ) -> BaseModelOutputWithPooling:
        r"""
        Returns:

        """

        return self.text_model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
        )


class Siglip2MultiheadAttentionPoolingHead(nn.Module):
    """Multihead Attention Pooling."""

    def __init__(self, config: Siglip2VisionConfig):
        super().__init__()

        self.probe = nn.Parameter(torch.randn(1, 1, config.hidden_size))
        self.attention = torch.nn.MultiheadAttention(config.hidden_size, config.num_attention_heads, batch_first=True)
        self.layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.mlp = Siglip2MLP(config)
        self.num_heads = config.num_attention_heads

    def forward(self, hidden_state: torch.Tensor, attention_mask: Optional[torch.Tensor] = None) -> torch.Tensor:
        batch_size = hidden_state.shape[0]
        probe = self.probe.repeat(batch_size, 1, 1)

        if attention_mask is not None:
            target_len, source_len = probe.shape[1], hidden_state.shape[1]
            attention_mask = _prepare_4d_attention_mask(attention_mask, hidden_state.dtype, target_len)
            attention_mask = attention_mask.repeat(1, self.num_heads, target_len, 1)
            attention_mask = attention_mask.reshape(-1, target_len, source_len)

        hidden_state = self.attention(probe, hidden_state, hidden_state, attn_mask=attention_mask)[0]

        residual = hidden_state
        hidden_state = self.layernorm(hidden_state)
        hidden_state = residual + self.mlp(hidden_state)

        return hidden_state[:, 0]


@add_start_docstrings(
    """The vision model from Siglip2 without any head or projection on top.""",
    SIGLIP2_START_DOCSTRING,
)
class Siglip2VisionModel(Siglip2PreTrainedModel):
    config_class = Siglip2VisionConfig
    main_input_name = "pixel_values"

    def __init__(self, config: Siglip2VisionConfig):
        super().__init__(config)

        self.vision_model = Siglip2VisionTransformer(config)

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

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

    @can_return_tuple
    @add_start_docstrings_to_model_forward(SIGLIP2_VISION_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=Siglip2VisionConfig)
    def forward(
        self,
        pixel_values: torch.FloatTensor,
        pixel_attention_mask: torch.Tensor,
        spatial_shapes: torch.LongTensor,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
    ) -> BaseModelOutputWithPooling:
        r"""
        Returns:

        ```"""
        return self.vision_model(
            pixel_values=pixel_values,
            attention_mask=pixel_attention_mask,
            spatial_shapes=spatial_shapes,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
        )


@add_start_docstrings(SIGLIP2_START_DOCSTRING)
class Siglip2Model(Siglip2PreTrainedModel):
    config_class = Siglip2Config

    def __init__(self, config: Siglip2Config):
        super().__init__(config)

        if not isinstance(config.text_config, Siglip2TextConfig):
            raise TypeError(
                "config.text_config is expected to be of type Siglip2TextConfig but is of type"
                f" {type(config.text_config)}."
            )

        if not isinstance(config.vision_config, Siglip2VisionConfig):
            raise TypeError(
                "config.vision_config is expected to be of type Siglip2VisionConfig but is of type"
                f" {type(config.vision_config)}."
            )

        text_config = config.text_config
        vision_config = config.vision_config

        text_model = Siglip2TextModel._from_config(text_config)
        vision_model = Siglip2VisionModel._from_config(vision_config)

        self.text_model = text_model.text_model
        self.vision_model = vision_model.vision_model

        self.logit_scale = nn.Parameter(torch.randn(1))
        self.logit_bias = nn.Parameter(torch.randn(1))

        self.post_init()

    @add_start_docstrings_to_model_forward(SIGLIP2_TEXT_INPUTS_DOCSTRING)
    def get_text_features(
        self,
        input_ids: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.Tensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
    ) -> torch.FloatTensor:
        r"""
        Returns:
            text_features (`torch.FloatTensor` of shape `(batch_size, output_dim`): The text embeddings obtained by
            applying the projection layer to the pooled output of [`Siglip2TextModel`].

        """
        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
        )

        text_outputs: BaseModelOutputWithPooling = self.text_model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
        )

        pooled_output = text_outputs.pooler_output

        return pooled_output

    @add_start_docstrings_to_model_forward(SIGLIP2_VISION_INPUTS_DOCSTRING)
    def get_image_features(
        self,
        pixel_values: Optional[torch.FloatTensor] = None,
        pixel_attention_mask: Optional[torch.Tensor] = None,
        spatial_shapes: Optional[torch.LongTensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
    ) -> torch.FloatTensor:
        r"""
        Returns:
            image_features (`torch.FloatTensor` of shape `(batch_size, output_dim`): The image embeddings obtained by
            applying the projection layer to the pooled output of [`Siglip2VisionModel`].

        """
        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
        )

        vision_outputs: BaseModelOutputWithPooling = self.vision_model(
            pixel_values=pixel_values,
            attention_mask=pixel_attention_mask,
            spatial_shapes=spatial_shapes,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
        )

        pooled_output = vision_outputs.pooler_output

        return pooled_output

    @can_return_tuple
    @add_start_docstrings_to_model_forward(SIGLIP2_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=Siglip2Output, config_class=Siglip2Config)
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        pixel_values: Optional[torch.FloatTensor] = None,
        pixel_attention_mask: Optional[torch.Tensor] = None,
        spatial_shapes: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        return_loss: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
    ) -> Siglip2Output:
        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
        )

        vision_outputs: BaseModelOutputWithPooling = self.vision_model(
            pixel_values=pixel_values,
            attention_mask=pixel_attention_mask,
            spatial_shapes=spatial_shapes,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
        )

        text_outputs: BaseModelOutputWithPooling = self.text_model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
        )

        image_embeds = vision_outputs.pooler_output
        text_embeds = text_outputs.pooler_output

        image_embeds = image_embeds / image_embeds.norm(p=2, dim=-1, keepdim=True)
        text_embeds = text_embeds / text_embeds.norm(p=2, dim=-1, keepdim=True)

        logits_per_text = torch.matmul(text_embeds, image_embeds.t().to(text_embeds.device))

        logit_scale, logit_bias = self.logit_scale.to(text_embeds.device), self.logit_bias.to(text_embeds.device)
        logits_per_text = logits_per_text * logit_scale.exp() + logit_bias

        logits_per_image = logits_per_text.t()

        loss = None
        if return_loss:
            eye = torch.eye(logits_per_text.size(0), device=logits_per_text.device)
            m1_diag1 = -torch.ones_like(logits_per_text) + 2 * eye
            loglik = torch.nn.functional.logsigmoid(m1_diag1 * logits_per_text)
            nll = -torch.sum(loglik, dim=-1)
            loss = nll.mean()

        return Siglip2Output(
            loss=loss,
            logits_per_image=logits_per_image,
            logits_per_text=logits_per_text,
            text_embeds=text_embeds,
            image_embeds=image_embeds,
            text_model_output=text_outputs,
            vision_model_output=vision_outputs,
        )


@add_start_docstrings(
    """
    Siglip2 vision encoder with an image classification head on top (a 
    linear layer on top of the pooled final hidden states of
    the patch tokens) e.g. for ImageNet.
    """,
    SIGLIP2_START_DOCSTRING,
)
class Siglip2ForImageClassification(Siglip2PreTrainedModel):
    main_input_name = "pixel_values"

    def __init__(self, config: Siglip2Config) -> None:
        super().__init__(config)

        self.num_labels = config.num_labels

        vision_model = Siglip2VisionModel._from_config(config.vision_config)
        self.vision_model = vision_model.vision_model

        self.classifier = (
            nn.Linear(config.vision_config.hidden_size, config.num_labels) if config.num_labels > 0 else nn.Identity()
        )

        self.post_init()

    @can_return_tuple
    @add_start_docstrings_to_model_forward(SIGLIP2_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=ImageClassifierOutput, config_class=_CONFIG_FOR_DOC)
    def forward(
        self,
        pixel_values: Optional[torch.Tensor] = None,
        pixel_attention_mask: Optional[torch.Tensor] = None,
        spatial_shapes: Optional[torch.LongTensor] = None,
        labels: Optional[torch.Tensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
    ) -> ImageClassifierOutput:
        r"""
        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
            Labels for computing the image classification/regression loss. Indices should be in `[0, ...,
            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).

        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
        )

        outputs: BaseModelOutputWithPooling = self.vision_model(
            pixel_values,
            attention_mask=pixel_attention_mask,
            spatial_shapes=spatial_shapes,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
        )

        sequence_output = outputs.last_hidden_state

        if pixel_attention_mask is not None:
            pool_mask = pixel_attention_mask[..., None].to(sequence_output.device)
            sequence_output = torch.sum(sequence_output * pool_mask, dim=1) / torch.sum(pool_mask, dim=1)
        else:
            sequence_output = torch.mean(sequence_output, dim=1)

        logits = self.classifier(sequence_output)

        loss = None
        if labels is not None:
            labels = labels.to(logits.device)
            if self.config.problem_type is None:
                if self.num_labels == 1:
                    self.config.problem_type = "regression"
                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
                    self.config.problem_type = "single_label_classification"
                else:
                    self.config.problem_type = "multi_label_classification"

            if self.config.problem_type == "regression":
                loss_fct = MSELoss()
                if self.num_labels == 1:
                    loss = loss_fct(logits.squeeze(), labels.squeeze())
                else:
                    loss = loss_fct(logits, labels)
            elif self.config.problem_type == "single_label_classification":
                loss_fct = CrossEntropyLoss()
                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
            elif self.config.problem_type == "multi_label_classification":
                loss_fct = BCEWithLogitsLoss()
                loss = loss_fct(logits, labels)

        return ImageClassifierOutput(
            loss=loss,
            logits=logits,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )


__all__ = [
    "Siglip2Model",
    "Siglip2PreTrainedModel",
    "Siglip2TextModel",
    "Siglip2VisionModel",
    "Siglip2ForImageClassification",
]