modeling_siglip2.py

#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
#           This file was automatically generated from src/transformers/models/siglip2/modular_siglip2.py.
#               Do NOT edit this file manually as any edits will be overwritten by the generation of
#             the file from the modular. If any change should be done, please apply the change to the
#                          modular_siglip2.py file directly. One of our CI enforces this.
#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
# coding=utf-8
# Copyright 2025 The HuggingFace Inc. team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import math
import warnings
from dataclasses import dataclass
from typing import Any, Callable, Optional, Tuple, Union

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, auto_docstring, can_return_tuple, logging
from transformers.cache_utils import Cache, DynamicCache, SlidingWindowCache, StaticCache

from .configuration_siglip2 import Siglip2VisionConfig


logger = logging.get_logger(__name__)

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
    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
    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 subsequently 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")

# Copied from transformers.models.llama.modeling_llama.rotate_half
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(q, k, cos, sin, unsqueeze_dim=1):
    cos = cos.unsqueeze(unsqueeze_dim)
    sin = sin.unsqueeze(unsqueeze_dim)
    q_embed = (q * cos) + (rotate_half(q) * sin)
    k_embed = (k * cos) + (rotate_half(k) * sin)
    return q_embed, k_embed

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

        self.patch_embedding = nn.Linear(
            in_features=config.num_channels * self.patch_size * self.patch_size,
            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,
        )

        # (height, width, embed_dim) -> (1, embed_dim, height, width) for interpolation
        positional_embeddings = positional_embeddings.permute(2, 0, 1).unsqueeze(0)

        # Upcast to float32 on CPU because antialias is not supported for bfloat16/float16 on CPU
        if positional_embeddings.device.type == "cpu":
            positional_embeddings = positional_embeddings.to(torch.float32)

        for i in range(batch_size):
            # (1, dim, height, width) -> (1, dim, target_height, target_width)
            height, width = spatial_shapes[i]
            resized_embeddings = F.interpolate(
                positional_embeddings,
                size=(height, width),
                mode="bilinear",
                align_corners=False,
                antialias=True,
            )

            # (1, dim, target_height, target_width) -> (target_height * target_width, dim)
            resized_embeddings = resized_embeddings.reshape(embed_dim, height * width).transpose(0, 1)

            # Cast to original dtype
            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
        """

        # Apply patch embeddings to already patchified pixel values
        target_dtype = self.patch_embedding.weight.dtype
        patch_embeds = self.patch_embedding(pixel_values.to(dtype=target_dtype))

        # Get positional resized and padded positional embeddings
        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]
        )

        # Add positional embeddings to patch embeddings
        embeddings = patch_embeds + resized_positional_embeddings
        return embeddings

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

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

    def __init__(self, config: Union[Siglip2VisionConfig], layer_idx: int):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx
        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,
        position_embeddings: Optional[torch.Tensor] = None,
        past_key_value: Optional[Cache] = None,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
        """Input shape: Batch x Time x Channel"""

        batch_size, seq_length, embed_dim = hidden_states.shape

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

        if position_embeddings is not None:
            cos, sin = position_embeddings
            query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)

        if past_key_value is not None:
            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx)
    
        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,
            query_states,
            key_states,
            value_states,
            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 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 VisionRotaryEmbedding(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, x, position_ids: int) -> torch.Tensor:
        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
        position_ids_expanded = position_ids[:, None, :].float()

        device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
        with torch.autocast(device_type=device_type, enabled=False):  # Force float32
            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
            emb = torch.cat((freqs, freqs), dim=-1)
            cos, sin = emb.cos(), emb.sin()

        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)

    def _apply(self, fn, recurse=True):
        for key, buf in self._buffers.items():
            if buf is not None:
                # self._buffers[key] = fn(buf)
                value = self._buffers[key]
                value_ = fn(buf)
                self._buffers[key] = value.to(value_.device)

        return self

class Siglip2EncoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: Union[Siglip2VisionConfig], layer_idx):
        super().__init__()
        self.embed_dim = config.hidden_size
        self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
        self.self_attn = Siglip2Attention(config, layer_idx)
        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,
        output_attentions: Optional[bool] = False,
        position_embeddings: Optional[torch.Tensor] = None,
        past_key_value: Optional[Cache] = None,
    ) -> 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,
            position_embeddings=position_embeddings,
            past_key_value=past_key_value,
        )
        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 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: Siglip2VisionConfig):
        super().__init__()
        self.config = config
        self.layers = nn.ModuleList([Siglip2EncoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)])
        self.gradient_checkpointing = False

    # Ignore copy
    @can_return_tuple
    def forward(
        self,
        inputs_embeds,
        attention_mask: Optional[torch.Tensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        position_embeddings: Optional[torch.Tensor] = None,
        past_key_value: Optional[Cache] = 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
        for encoder_layer in self.layers:
            if output_hidden_states:
                encoder_states = encoder_states + (hidden_states,)

            layer_outputs = encoder_layer(
                hidden_states,
                attention_mask,
                output_attentions=output_attentions,
                position_embeddings=position_embeddings,
                past_key_value=past_key_value,
            )

            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,)

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

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

        head_dim = config.hidden_size // config.num_attention_heads
        self.rotary_pos_emb = VisionRotaryEmbedding(head_dim)

        self.embeddings = Siglip2VisionEmbeddings(config)
        self.encoder = Siglip2Encoder(config)
        self.post_layernorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
        self.use_head = True if not hasattr(config, "vision_use_head") else config.vision_use_head
        self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"

    @can_return_tuple
    @auto_docstring
    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,
        past_key_value: Optional[Cache] = None,
    ) -> BaseModelOutputWithPooling:
        r"""
        spatial_shapes (`torch.LongTensor` of shape `(batch_size, 2)`):
            Tensor containing the spatial dimensions (height, width) of the input images.
        """
        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
        )

        hidden_states = self.embeddings(pixel_values, spatial_shapes)

        position_embeddings = self.rotary_pos_emb(hidden_states.shape[1])

        if attention_mask is not None and not self._use_flash_attention_2:
            # [batch_size, seq_len] -> [batch_size, 1, tgt_seq_len, src_seq_len]
            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,
            attention_mask=encoder_attention_mask,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            position_embeddings=position_embeddings,
            past_key_value=past_key_value
        )

        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,
        )


    

@auto_docstring(
    custom_intro="""
    The vision model from Siglip2 without any head or projection on top.
    """
)
class Siglip2VisionModel(PreTrainedModel):
    config_class = Siglip2VisionConfig
    main_input_name = "pixel_values"
    base_model_prefix = "siglip2"
    supports_gradient_checkpointing = True
    _no_split_modules = ["Siglip2EncoderLayer", "Siglip2VisionEmbeddings", "Siglip2EncoderLayer"]
    _supports_flash_attn_2 = True
    _supports_sdpa = True
    
    def _init_weights(self, module):
        """Initialize the weights"""
        if isinstance(module, Siglip2VisionEmbeddings):
            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, 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, (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)

    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
    @auto_docstring
    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,
        past_key_value: Optional[Cache] = None,
    ) -> BaseModelOutputWithPooling:
        r"""
        pixel_attention_mask (`torch.Tensor` of shape `(batch_size, image_size, image_size)`, *optional*):
            Mask to avoid performing attention on padding pixel indices.
        spatial_shapes (`torch.LongTensor` of shape `(batch_size, 2)`):
            Tensor containing the spatial dimensions (height, width) of the input images.

        Examples:

        ```python
        >>> from PIL import Image
        >>> import requests
        >>> from transformers import AutoProcessor, Siglip2VisionModel

        >>> model = Siglip2VisionModel.from_pretrained("google/siglip2-base-patch16-224")
        >>> processor = AutoProcessor.from_pretrained("google/siglip2-base-patch16-224")

        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
        >>> image = Image.open(requests.get(url, stream=True).raw)

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

        >>> outputs = model(**inputs)
        >>> last_hidden_state = outputs.last_hidden_state
        >>> pooled_output = outputs.pooler_output  # pooled features
        ```"""
        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,
            past_key_value=past_key_value,
        )