transformers 5.0.0rc3__py3-none-any.whl → 5.1.0__py3-none-any.whl

transformers/models/pp_doclayout_v3/modeling_pp_doclayout_v3.py

@@ -0,0 +1,2083 @@
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+#           This file was automatically generated from src/transformers/models/pp_doclayout_v3/modular_pp_doclayout_v3.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_pp_doclayout_v3.py file directly. One of our CI enforces this.
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+# Copyright 2026 The PaddlePaddle Team and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+import warnings
+from collections.abc import Callable
+from dataclasses import dataclass
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from torch import Tensor, nn
+
+from ... import initialization as init
+from ...activations import ACT2CLS, ACT2FN
+from ...backbone_utils import load_backbone
+from ...image_transforms import center_to_corners_format, corners_to_center_format
+from ...integrations import use_kernel_forward_from_hub
+from ...modeling_outputs import BaseModelOutput
+from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
+from ...processing_utils import Unpack
+from ...pytorch_utils import compile_compatible_method_lru_cache
+from ...utils import (
+    ModelOutput,
+    TransformersKwargs,
+    auto_docstring,
+    torch_compilable_check,
+    torch_int,
+)
+from ...utils.generic import can_return_tuple, check_model_inputs
+from .configuration_pp_doclayout_v3 import PPDocLayoutV3Config
+
+
+class PPDocLayoutV3GlobalPointer(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.head_size = config.global_pointer_head_size
+        self.dense = nn.Linear(config.d_model, self.head_size * 2)
+        self.dropout = nn.Dropout(config.gp_dropout_value)
+
+    def forward(self, inputs):
+        batch_size, sequence_length, _ = inputs.shape
+        query_key_projection = self.dense(inputs).reshape(batch_size, sequence_length, 2, self.head_size)
+        query_key_projection = self.dropout(query_key_projection)
+        queries, keys = torch.unbind(query_key_projection, dim=2)
+
+        logits = (queries @ keys.transpose(-2, -1)) / (self.head_size**0.5)
+        mask = torch.tril(torch.ones(sequence_length, sequence_length, device=logits.device)).bool()
+        logits = logits.masked_fill(mask.unsqueeze(0), -1e4)
+
+        return logits
+
+
+@use_kernel_forward_from_hub("MultiScaleDeformableAttention")
+class MultiScaleDeformableAttention(nn.Module):
+    def forward(
+        self,
+        value: Tensor,
+        value_spatial_shapes: Tensor,
+        value_spatial_shapes_list: list[tuple],
+        level_start_index: Tensor,
+        sampling_locations: Tensor,
+        attention_weights: Tensor,
+        im2col_step: int,
+    ):
+        batch_size, _, num_heads, hidden_dim = value.shape
+        _, num_queries, num_heads, num_levels, num_points, _ = sampling_locations.shape
+        value_list = value.split([height * width for height, width in value_spatial_shapes_list], dim=1)
+        sampling_grids = 2 * sampling_locations - 1
+        sampling_value_list = []
+        for level_id, (height, width) in enumerate(value_spatial_shapes_list):
+            # batch_size, height*width, num_heads, hidden_dim
+            # -> batch_size, height*width, num_heads*hidden_dim
+            # -> batch_size, num_heads*hidden_dim, height*width
+            # -> batch_size*num_heads, hidden_dim, height, width
+            value_l_ = (
+                value_list[level_id]
+                .flatten(2)
+                .transpose(1, 2)
+                .reshape(batch_size * num_heads, hidden_dim, height, width)
+            )
+            # batch_size, num_queries, num_heads, num_points, 2
+            # -> batch_size, num_heads, num_queries, num_points, 2
+            # -> batch_size*num_heads, num_queries, num_points, 2
+            sampling_grid_l_ = sampling_grids[:, :, :, level_id].transpose(1, 2).flatten(0, 1)
+            # batch_size*num_heads, hidden_dim, num_queries, num_points
+            sampling_value_l_ = nn.functional.grid_sample(
+                value_l_,
+                sampling_grid_l_,
+                mode="bilinear",
+                padding_mode="zeros",
+                align_corners=False,
+            )
+            sampling_value_list.append(sampling_value_l_)
+        # (batch_size, num_queries, num_heads, num_levels, num_points)
+        # -> (batch_size, num_heads, num_queries, num_levels, num_points)
+        # -> (batch_size, num_heads, 1, num_queries, num_levels*num_points)
+        attention_weights = attention_weights.transpose(1, 2).reshape(
+            batch_size * num_heads, 1, num_queries, num_levels * num_points
+        )
+        output = (
+            (torch.stack(sampling_value_list, dim=-2).flatten(-2) * attention_weights)
+            .sum(-1)
+            .view(batch_size, num_heads * hidden_dim, num_queries)
+        )
+        return output.transpose(1, 2).contiguous()
+
+
+class PPDocLayoutV3MultiscaleDeformableAttention(nn.Module):
+    """
+    Multiscale deformable attention as proposed in Deformable DETR.
+    """
+
+    def __init__(self, config: PPDocLayoutV3Config, num_heads: int, n_points: int):
+        super().__init__()
+
+        self.attn = MultiScaleDeformableAttention()
+
+        if config.d_model % num_heads != 0:
+            raise ValueError(
+                f"embed_dim (d_model) must be divisible by num_heads, but got {config.d_model} and {num_heads}"
+            )
+        dim_per_head = config.d_model // num_heads
+        # check if dim_per_head is power of 2
+        if not ((dim_per_head & (dim_per_head - 1) == 0) and dim_per_head != 0):
+            warnings.warn(
+                "You'd better set embed_dim (d_model) in PPDocLayoutV3MultiscaleDeformableAttention to make the"
+                " dimension of each attention head a power of 2 which is more efficient in the authors' CUDA"
+                " implementation."
+            )
+
+        self.im2col_step = 64
+
+        self.d_model = config.d_model
+        self.n_levels = config.num_feature_levels
+        self.n_heads = num_heads
+        self.n_points = n_points
+
+        self.sampling_offsets = nn.Linear(config.d_model, num_heads * self.n_levels * n_points * 2)
+        self.attention_weights = nn.Linear(config.d_model, num_heads * self.n_levels * n_points)
+        self.value_proj = nn.Linear(config.d_model, config.d_model)
+        self.output_proj = nn.Linear(config.d_model, config.d_model)
+
+        self.disable_custom_kernels = config.disable_custom_kernels
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor | None = None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        position_embeddings: torch.Tensor | None = None,
+        reference_points=None,
+        spatial_shapes=None,
+        spatial_shapes_list=None,
+        level_start_index=None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        # add position embeddings to the hidden states before projecting to queries and keys
+        if position_embeddings is not None:
+            hidden_states = hidden_states + position_embeddings
+
+        batch_size, num_queries, _ = hidden_states.shape
+        batch_size, sequence_length, _ = encoder_hidden_states.shape
+        total_elements = sum(height * width for height, width in spatial_shapes_list)
+        torch_compilable_check(
+            total_elements == sequence_length,
+            "Make sure to align the spatial shapes with the sequence length of the encoder hidden states",
+        )
+
+        value = self.value_proj(encoder_hidden_states)
+        if attention_mask is not None:
+            # we invert the attention_mask
+            value = value.masked_fill(~attention_mask[..., None], float(0))
+        value = value.view(batch_size, sequence_length, self.n_heads, self.d_model // self.n_heads)
+        sampling_offsets = self.sampling_offsets(hidden_states).view(
+            batch_size, num_queries, self.n_heads, self.n_levels, self.n_points, 2
+        )
+        attention_weights = self.attention_weights(hidden_states).view(
+            batch_size, num_queries, self.n_heads, self.n_levels * self.n_points
+        )
+        attention_weights = F.softmax(attention_weights, -1).view(
+            batch_size, num_queries, self.n_heads, self.n_levels, self.n_points
+        )
+        # batch_size, num_queries, n_heads, n_levels, n_points, 2
+        num_coordinates = reference_points.shape[-1]
+        if num_coordinates == 2:
+            offset_normalizer = torch.stack([spatial_shapes[..., 1], spatial_shapes[..., 0]], -1)
+            sampling_locations = (
+                reference_points[:, :, None, :, None, :]
+                + sampling_offsets / offset_normalizer[None, None, None, :, None, :]
+            )
+        elif num_coordinates == 4:
+            sampling_locations = (
+                reference_points[:, :, None, :, None, :2]
+                + sampling_offsets / self.n_points * reference_points[:, :, None, :, None, 2:] * 0.5
+            )
+        else:
+            raise ValueError(f"Last dim of reference_points must be 2 or 4, but got {reference_points.shape[-1]}")
+
+        output = self.attn(
+            value,
+            spatial_shapes,
+            spatial_shapes_list,
+            level_start_index,
+            sampling_locations,
+            attention_weights,
+            self.im2col_step,
+        )
+
+        output = self.output_proj(output)
+
+        return output, attention_weights
+
+
+@auto_docstring
+class PPDocLayoutV3PreTrainedModel(PreTrainedModel):
+    config: PPDocLayoutV3Config
+    base_model_prefix = "pp_doclayout_v3"
+    main_input_name = "pixel_values"
+    input_modalities = ("image",)
+    _no_split_modules = [r"PPDocLayoutV3HybridEncoder", r"PPDocLayoutV3DecoderLayer"]
+    _supports_sdpa = True
+    _supports_flash_attn = True
+    _supports_attention_backend = True
+    _supports_flex_attn = True
+
+    @torch.no_grad()
+    def _init_weights(self, module):
+        """Initialize the weights"""
+        if isinstance(module, PPDocLayoutV3MultiscaleDeformableAttention):
+            init.constant_(module.sampling_offsets.weight, 0.0)
+            default_dtype = torch.get_default_dtype()
+            thetas = torch.arange(module.n_heads, dtype=torch.int64).to(default_dtype) * (
+                2.0 * math.pi / module.n_heads
+            )
+            grid_init = torch.stack([thetas.cos(), thetas.sin()], -1)
+            grid_init = (
+                (grid_init / grid_init.abs().max(-1, keepdim=True)[0])
+                .view(module.n_heads, 1, 1, 2)
+                .repeat(1, module.n_levels, module.n_points, 1)
+            )
+            for i in range(module.n_points):
+                grid_init[:, :, i, :] *= i + 1
+
+            init.copy_(module.sampling_offsets.bias, grid_init.view(-1))
+            init.constant_(module.attention_weights.weight, 0.0)
+            init.constant_(module.attention_weights.bias, 0.0)
+            init.xavier_uniform_(module.value_proj.weight)
+            init.constant_(module.value_proj.bias, 0.0)
+            init.xavier_uniform_(module.output_proj.weight)
+            init.constant_(module.output_proj.bias, 0.0)
+
+        elif isinstance(module, PPDocLayoutV3Model):
+            prior_prob = self.config.initializer_bias_prior_prob or 1 / (self.config.num_labels + 1)
+            bias = float(-math.log((1 - prior_prob) / prior_prob))
+            init.xavier_uniform_(module.enc_score_head.weight)
+            init.constant_(module.enc_score_head.bias, bias)
+            init.xavier_uniform_(module.decoder.class_embed.weight)
+            init.constant_(module.decoder.class_embed.bias, bias)
+
+        elif isinstance(module, (nn.Linear, nn.Conv2d, nn.BatchNorm2d)):
+            init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                init.zeros_(module.bias)
+            if getattr(module, "running_mean", None) is not None:
+                init.zeros_(module.running_mean)
+                init.ones_(module.running_var)
+                init.zeros_(module.num_batches_tracked)
+
+        elif isinstance(module, nn.LayerNorm):
+            init.ones_(module.weight)
+            init.zeros_(module.bias)
+
+        if isinstance(module, nn.Embedding):
+            init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
+            if module.padding_idx is not None:
+                init.zeros_(module.weight.data[module.padding_idx])
+
+
+@dataclass
+class PPDocLayoutV3DecoderOutput(ModelOutput):
+    r"""
+    intermediate_hidden_states (`torch.FloatTensor` of shape `(batch_size, config.decoder_layers, num_queries, hidden_size)`):
+        Stacked intermediate hidden states (output of each layer of the decoder).
+    intermediate_logits (`torch.FloatTensor` of shape `(batch_size, config.decoder_layers, sequence_length, config.num_labels)`):
+        Stacked intermediate logits (logits of each layer of the decoder).
+    intermediate_reference_points (`torch.FloatTensor` of shape `(batch_size, config.decoder_layers, sequence_length, hidden_size)`):
+        Stacked intermediate reference points (reference points of each layer of the decoder).
+    intermediate_predicted_corners (`torch.FloatTensor` of shape `(batch_size, config.decoder_layers, num_queries, 4)`):
+        Stacked intermediate predicted corners (predicted corners of each layer of the decoder).
+    initial_reference_points (`torch.FloatTensor` of shape `(batch_size, config.decoder_layers, num_queries, 4)`):
+        Stacked initial reference points (initial reference points of each layer of the decoder).
+    cross_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` and `config.add_cross_attention=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 of the decoder's cross-attention layer, after the attention softmax,
+        used to compute the weighted average in the cross-attention heads.
+    decoder_out_order_logits (`torch.FloatTensor` of shape `(batch_size, config.decoder_layers, config.num_queries, config.num_queries)`):
+        Stacked order logits (order logits of each layer of the decoder).
+    decoder_out_masks (`torch.FloatTensor` of shape `(batch_size, config.decoder_layers, config.num_queries, 200, 200)`):
+        Stacked masks (masks of each layer of the decoder).
+    """
+
+    last_hidden_state: torch.FloatTensor | None = None
+    intermediate_hidden_states: torch.FloatTensor | None = None
+    intermediate_logits: torch.FloatTensor | None = None
+    intermediate_reference_points: torch.FloatTensor | None = None
+    intermediate_predicted_corners: torch.FloatTensor | None = None
+    initial_reference_points: torch.FloatTensor | None = None
+    hidden_states: tuple[torch.FloatTensor] | None = None
+    attentions: tuple[torch.FloatTensor] | None = None
+    cross_attentions: tuple[torch.FloatTensor] | None = None
+
+    decoder_out_order_logits: torch.FloatTensor | None = None
+    decoder_out_masks: torch.FloatTensor | None = None
+
+
+@dataclass
+@auto_docstring(
+    custom_intro="""
+    Base class for outputs of the PP-DocLayoutV3 model.
+    """
+)
+class PPDocLayoutV3ModelOutput(ModelOutput):
+    r"""
+    last_hidden_state (`torch.FloatTensor` of shape `(batch_size, num_queries, hidden_size)`):
+        Sequence of hidden-states at the output of the last layer of the decoder of the model.
+    intermediate_hidden_states (`torch.FloatTensor` of shape `(batch_size, config.decoder_layers, num_queries, hidden_size)`):
+        Stacked intermediate hidden states (output of each layer of the decoder).
+    intermediate_logits (`torch.FloatTensor` of shape `(batch_size, config.decoder_layers, sequence_length, config.num_labels)`):
+        Stacked intermediate logits (logits of each layer of the decoder).
+    intermediate_reference_points (`torch.FloatTensor` of shape `(batch_size, config.decoder_layers, num_queries, 4)`):
+        Stacked intermediate reference points (reference points of each layer of the decoder).
+    intermediate_predicted_corners (`torch.FloatTensor` of shape `(batch_size, config.decoder_layers, num_queries, 4)`):
+        Stacked intermediate predicted corners (predicted corners of each layer of the decoder).
+    initial_reference_points (`torch.FloatTensor` of shape `(batch_size, num_queries, 4)`):
+        Initial reference points used for the first decoder layer.
+    init_reference_points (`torch.FloatTensor` of shape `(batch_size, num_queries, 4)`):
+        Initial reference points sent through the Transformer decoder.
+    enc_topk_logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.num_labels)`):
+        Predicted bounding boxes scores where the top `config.two_stage_num_proposals` scoring bounding boxes are
+        picked as region proposals in the encoder stage. Output of bounding box binary classification (i.e.
+        foreground and background).
+    enc_topk_bboxes (`torch.FloatTensor` of shape `(batch_size, sequence_length, 4)`):
+        Logits of predicted bounding boxes coordinates in the encoder stage.
+    enc_outputs_class (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.num_labels)`, *optional*, returned when `config.with_box_refine=True` and `config.two_stage=True`):
+        Predicted bounding boxes scores where the top `config.two_stage_num_proposals` scoring bounding boxes are
+        picked as region proposals in the first stage. Output of bounding box binary classification (i.e.
+        foreground and background).
+    enc_outputs_coord_logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, 4)`, *optional*, returned when `config.with_box_refine=True` and `config.two_stage=True`):
+        Logits of predicted bounding boxes coordinates in the first stage.
+    denoising_meta_values (`dict`):
+        Extra dictionary for the denoising related values.
+    out_order_logits (`torch.FloatTensor` of shape `(batch_size, config.decoder_layers, config.num_queries, config.num_queries)`):
+        Stacked order logits (order logits of each layer of the decoder).
+    out_masks (`torch.FloatTensor` of shape `(batch_size, config.decoder_layers, config.num_queries, 200, 200)`):
+        Stacked masks (masks of each layer of the decoder).
+    """
+
+    last_hidden_state: torch.FloatTensor | None = None
+    intermediate_hidden_states: torch.FloatTensor | None = None
+    intermediate_logits: torch.FloatTensor | None = None
+    intermediate_reference_points: torch.FloatTensor | None = None
+    intermediate_predicted_corners: torch.FloatTensor | None = None
+    initial_reference_points: torch.FloatTensor | None = None
+    decoder_hidden_states: tuple[torch.FloatTensor] | None = None
+    decoder_attentions: tuple[torch.FloatTensor] | None = None
+    cross_attentions: tuple[torch.FloatTensor] | None = None
+    encoder_last_hidden_state: torch.FloatTensor | None = None
+    encoder_hidden_states: tuple[torch.FloatTensor] | None = None
+    encoder_attentions: tuple[torch.FloatTensor] | None = None
+    init_reference_points: torch.FloatTensor | None = None
+    enc_topk_logits: torch.FloatTensor | None = None
+    enc_topk_bboxes: torch.FloatTensor | None = None
+    enc_outputs_class: torch.FloatTensor | None = None
+    enc_outputs_coord_logits: torch.FloatTensor | None = None
+    denoising_meta_values: dict | None = None
+
+    out_order_logits: torch.FloatTensor | None = None
+    out_masks: torch.FloatTensor | None = None
+
+
+class PPDocLayoutV3MLPPredictionHead(nn.Module):
+    """
+    Very simple multi-layer perceptron (MLP, also called FFN), used to predict the normalized center coordinates,
+    height and width of a bounding box w.r.t. an image.
+
+    """
+
+    def __init__(self, input_dim, hidden_dim, output_dim, num_layers):
+        super().__init__()
+        self.num_layers = num_layers
+        h = [hidden_dim] * (num_layers - 1)
+        self.layers = nn.ModuleList(nn.Linear(n, k) for n, k in zip([input_dim] + h, h + [output_dim]))
+
+    def forward(self, x):
+        for i, layer in enumerate(self.layers):
+            x = nn.functional.relu(layer(x)) if i < self.num_layers - 1 else layer(x)
+        return x
+
+
+class PPDocLayoutV3ConvLayer(nn.Module):
+    def __init__(
+        self, in_channels: int, out_channels: int, kernel_size: int = 3, stride: int = 1, activation: str = "relu"
+    ):
+        super().__init__()
+        self.convolution = nn.Conv2d(
+            in_channels, out_channels, kernel_size=kernel_size, stride=stride, padding=kernel_size // 2, bias=False
+        )
+        self.normalization = nn.BatchNorm2d(out_channels)
+        self.activation = ACT2FN[activation] if activation is not None else nn.Identity()
+
+    def forward(self, input: Tensor) -> Tensor:
+        hidden_state = self.convolution(input)
+        hidden_state = self.normalization(hidden_state)
+        hidden_state = self.activation(hidden_state)
+        return hidden_state
+
+
+class PPDocLayoutV3ScaleHead(nn.Module):
+    def __init__(self, in_channels, feature_channels, fpn_stride, base_stride, align_corners=False):
+        super().__init__()
+        head_length = max(1, int(np.log2(fpn_stride) - np.log2(base_stride)))
+        self.layers = nn.ModuleList()
+        for k in range(head_length):
+            in_c = in_channels if k == 0 else feature_channels
+            self.layers.append(PPDocLayoutV3ConvLayer(in_c, feature_channels, 3, 1, "silu"))
+            if fpn_stride != base_stride:
+                self.layers.append(nn.Upsample(scale_factor=2, mode="bilinear", align_corners=align_corners))
+
+    def forward(self, x):
+        for layer in self.layers:
+            x = layer(x)
+        return x
+
+
+class PPDocLayoutV3MaskFeatFPN(nn.Module):
+    def __init__(
+        self,
+        in_channels=[256, 256, 256],
+        fpn_strides=[32, 16, 8],
+        feature_channels=256,
+        dropout_ratio=0.0,
+        out_channels=256,
+        align_corners=False,
+    ):
+        super().__init__()
+
+        reorder_index = np.argsort(fpn_strides, axis=0).tolist()
+        in_channels = [in_channels[i] for i in reorder_index]
+        fpn_strides = [fpn_strides[i] for i in reorder_index]
+
+        self.reorder_index = reorder_index
+        self.fpn_strides = fpn_strides
+        self.dropout_ratio = dropout_ratio
+        self.align_corners = align_corners
+        if self.dropout_ratio > 0:
+            self.dropout = nn.Dropout2d(dropout_ratio)
+
+        self.scale_heads = nn.ModuleList()
+        for i in range(len(fpn_strides)):
+            self.scale_heads.append(
+                PPDocLayoutV3ScaleHead(
+                    in_channels=in_channels[i],
+                    feature_channels=feature_channels,
+                    fpn_stride=fpn_strides[i],
+                    base_stride=fpn_strides[0],
+                    align_corners=align_corners,
+                )
+            )
+        self.output_conv = PPDocLayoutV3ConvLayer(feature_channels, out_channels, 3, 1, "silu")
+
+    def forward(self, inputs):
+        x = [inputs[i] for i in self.reorder_index]
+
+        output = self.scale_heads[0](x[0])
+        for i in range(1, len(self.fpn_strides)):
+            output = output + F.interpolate(
+                self.scale_heads[i](x[i]), size=output.shape[2:], mode="bilinear", align_corners=self.align_corners
+            )
+
+        if self.dropout_ratio > 0:
+            output = self.dropout(output)
+        output = self.output_conv(output)
+        return output
+
+
+class PPDocLayoutV3EncoderMaskOutput(nn.Module):
+    def __init__(self, in_channels, num_prototypes):
+        super().__init__()
+        self.base_conv = PPDocLayoutV3ConvLayer(in_channels, in_channels, 3, 1, "silu")
+        self.conv = nn.Conv2d(in_channels, num_prototypes, kernel_size=1)
+
+    def forward(self, x):
+        x = self.base_conv(x)
+        x = self.conv(x)
+        return x
+
+
+class PPDocLayoutV3MLP(nn.Module):
+    def __init__(
+        self, config: PPDocLayoutV3Config, hidden_size: int, intermediate_size: int, activation_function: str
+    ):
+        super().__init__()
+        self.fc1 = nn.Linear(hidden_size, intermediate_size)
+        self.fc2 = nn.Linear(intermediate_size, hidden_size)
+        self.activation_fn = ACT2FN[activation_function]
+        self.activation_dropout = config.activation_dropout
+        self.dropout = config.dropout
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.activation_fn(self.fc1(hidden_states))
+        hidden_states = nn.functional.dropout(hidden_states, p=self.activation_dropout, training=self.training)
+        hidden_states = self.fc2(hidden_states)
+        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+        return hidden_states
+
+
+def eager_attention_forward(
+    module: nn.Module,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attention_mask: torch.Tensor | None,
+    scaling: float | None = None,
+    dropout: float = 0.0,
+    **kwargs: Unpack[TransformersKwargs],
+):
+    if scaling is None:
+        scaling = query.size(-1) ** -0.5
+
+    # Take the dot product between "query" and "key" to get the raw attention scores.
+    attn_weights = torch.matmul(query, key.transpose(2, 3)) * scaling
+
+    if attention_mask is not None:
+        attention_mask = attention_mask[:, :, :, : key.shape[-2]]
+        attn_weights = attn_weights + attention_mask
+
+    attn_weights = nn.functional.softmax(attn_weights, dim=-1)
+    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 PPDocLayoutV3SelfAttention(nn.Module):
+    """
+    Multi-headed self-attention from 'Attention Is All You Need' paper.
+
+    In PP_DOCLAYOUT_V3, position embeddings are added to both queries and keys (but not values) in self-attention.
+    """
+
+    def __init__(
+        self,
+        config: PPDocLayoutV3Config,
+        hidden_size: int,
+        num_attention_heads: int,
+        dropout: float = 0.0,
+        bias: bool = True,
+    ):
+        super().__init__()
+        self.config = config
+        self.head_dim = hidden_size // num_attention_heads
+        self.scaling = self.head_dim**-0.5
+        self.attention_dropout = dropout
+        self.is_causal = False
+
+        self.k_proj = nn.Linear(hidden_size, hidden_size, bias=bias)
+        self.v_proj = nn.Linear(hidden_size, hidden_size, bias=bias)
+        self.q_proj = nn.Linear(hidden_size, hidden_size, bias=bias)
+        self.o_proj = nn.Linear(hidden_size, hidden_size, bias=bias)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor | None = None,
+        position_embeddings: torch.Tensor | None = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        """
+        Position embeddings are added to both queries and keys (but not values).
+        """
+        input_shape = hidden_states.shape[:-1]
+        hidden_shape = (*input_shape, -1, self.head_dim)
+
+        query_key_input = hidden_states + position_embeddings if position_embeddings is not None else hidden_states
+
+        query_states = self.q_proj(query_key_input).view(hidden_shape).transpose(1, 2)
+        key_states = self.k_proj(query_key_input).view(hidden_shape).transpose(1, 2)
+        value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+
+        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
+            self.config._attn_implementation, eager_attention_forward
+        )
+
+        attn_output, attn_weights = attention_interface(
+            self,
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            dropout=0.0 if not self.training else self.attention_dropout,
+            scaling=self.scaling,
+            **kwargs,
+        )
+
+        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
+        attn_output = self.o_proj(attn_output)
+        return attn_output, attn_weights
+
+
+class PPDocLayoutV3ConvNormLayer(nn.Module):
+    def __init__(self, config, in_channels, out_channels, kernel_size, stride, padding=None, activation=None):
+        super().__init__()
+        self.conv = nn.Conv2d(
+            in_channels,
+            out_channels,
+            kernel_size,
+            stride,
+            padding=(kernel_size - 1) // 2 if padding is None else padding,
+            bias=False,
+        )
+        self.norm = nn.BatchNorm2d(out_channels, config.batch_norm_eps)
+        self.activation = nn.Identity() if activation is None else ACT2CLS[activation]()
+
+    def forward(self, hidden_state):
+        hidden_state = self.conv(hidden_state)
+        hidden_state = self.norm(hidden_state)
+        hidden_state = self.activation(hidden_state)
+        return hidden_state
+
+
+class PPDocLayoutV3EncoderLayer(nn.Module):
+    def __init__(self, config: PPDocLayoutV3Config):
+        super().__init__()
+        self.normalize_before = config.normalize_before
+        self.hidden_size = config.encoder_hidden_dim
+
+        # self-attention
+        self.self_attn = PPDocLayoutV3SelfAttention(
+            config=config,
+            hidden_size=self.hidden_size,
+            num_attention_heads=config.num_attention_heads,
+            dropout=config.dropout,
+        )
+        self.self_attn_layer_norm = nn.LayerNorm(self.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = config.dropout
+        self.mlp = PPDocLayoutV3MLP(
+            config, self.hidden_size, config.encoder_ffn_dim, config.encoder_activation_function
+        )
+        self.final_layer_norm = nn.LayerNorm(self.hidden_size, eps=config.layer_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor,
+        spatial_position_embeddings: torch.Tensor | None = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> torch.Tensor:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, hidden_size)`
+            attention_mask (`torch.FloatTensor`): attention mask of size
+                `(batch, 1, target_len, source_len)` where padding elements are indicated by very large negative
+                values.
+            spatial_position_embeddings (`torch.FloatTensor`, *optional*):
+                Spatial position embeddings (2D positional encodings of image locations), to be added to both
+                the queries and keys in self-attention (but not to values).
+        """
+        residual = hidden_states
+        if self.normalize_before:
+            hidden_states = self.self_attn_layer_norm(hidden_states)
+
+        hidden_states, _ = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_embeddings=spatial_position_embeddings,
+            **kwargs,
+        )
+
+        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+        hidden_states = residual + hidden_states
+        if not self.normalize_before:
+            hidden_states = self.self_attn_layer_norm(hidden_states)
+
+        if self.normalize_before:
+            hidden_states = self.final_layer_norm(hidden_states)
+        residual = hidden_states
+
+        hidden_states = self.mlp(hidden_states)
+
+        hidden_states = residual + hidden_states
+        if not self.normalize_before:
+            hidden_states = self.final_layer_norm(hidden_states)
+
+        if self.training:
+            if torch.isinf(hidden_states).any() or torch.isnan(hidden_states).any():
+                clamp_value = torch.finfo(hidden_states.dtype).max - 1000
+                hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)
+
+        return hidden_states
+
+
+class PPDocLayoutV3RepVggBlock(nn.Module):
+    """
+    RepVGG architecture block introduced by the work "RepVGG: Making VGG-style ConvNets Great Again".
+    """
+
+    def __init__(self, config: PPDocLayoutV3Config):
+        super().__init__()
+
+        activation = config.activation_function
+        hidden_channels = int(config.encoder_hidden_dim * config.hidden_expansion)
+        self.conv1 = PPDocLayoutV3ConvNormLayer(config, hidden_channels, hidden_channels, 3, 1, padding=1)
+        self.conv2 = PPDocLayoutV3ConvNormLayer(config, hidden_channels, hidden_channels, 1, 1, padding=0)
+        self.activation = nn.Identity() if activation is None else ACT2CLS[activation]()
+
+    def forward(self, x):
+        y = self.conv1(x) + self.conv2(x)
+        return self.activation(y)
+
+
+class PPDocLayoutV3CSPRepLayer(nn.Module):
+    """
+    Cross Stage Partial (CSP) network layer with RepVGG blocks.
+    """
+
+    def __init__(self, config: PPDocLayoutV3Config):
+        super().__init__()
+
+        in_channels = config.encoder_hidden_dim * 2
+        out_channels = config.encoder_hidden_dim
+        num_blocks = 3
+        activation = config.activation_function
+
+        hidden_channels = int(out_channels * config.hidden_expansion)
+        self.conv1 = PPDocLayoutV3ConvNormLayer(config, in_channels, hidden_channels, 1, 1, activation=activation)
+        self.conv2 = PPDocLayoutV3ConvNormLayer(config, in_channels, hidden_channels, 1, 1, activation=activation)
+        self.bottlenecks = nn.Sequential(*[PPDocLayoutV3RepVggBlock(config) for _ in range(num_blocks)])
+        if hidden_channels != out_channels:
+            self.conv3 = PPDocLayoutV3ConvNormLayer(config, hidden_channels, out_channels, 1, 1, activation=activation)
+        else:
+            self.conv3 = nn.Identity()
+
+    def forward(self, hidden_state):
+        hidden_state_1 = self.conv1(hidden_state)
+        hidden_state_1 = self.bottlenecks(hidden_state_1)
+        hidden_state_2 = self.conv2(hidden_state)
+        return self.conv3(hidden_state_1 + hidden_state_2)
+
+
+class PPDocLayoutV3SinePositionEmbedding(nn.Module):
+    """
+    2D sinusoidal position embedding used in RT-DETR hybrid encoder.
+    """
+
+    def __init__(self, embed_dim: int = 256, temperature: int = 10000):
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.temperature = temperature
+
+    @compile_compatible_method_lru_cache(maxsize=32)
+    def forward(
+        self,
+        width: int,
+        height: int,
+        device: torch.device | str,
+        dtype: torch.dtype,
+    ) -> torch.Tensor:
+        """
+        Generate 2D sinusoidal position embeddings.
+
+        Returns:
+            Position embeddings of shape (1, height*width, embed_dim)
+        """
+        grid_w = torch.arange(torch_int(width), device=device).to(dtype)
+        grid_h = torch.arange(torch_int(height), device=device).to(dtype)
+        grid_w, grid_h = torch.meshgrid(grid_w, grid_h, indexing="xy")
+        if self.embed_dim % 4 != 0:
+            raise ValueError("Embed dimension must be divisible by 4 for 2D sin-cos position embedding")
+        pos_dim = self.embed_dim // 4
+        omega = torch.arange(pos_dim, device=device).to(dtype) / pos_dim
+        omega = 1.0 / (self.temperature**omega)
+
+        out_w = grid_w.flatten()[..., None] @ omega[None]
+        out_h = grid_h.flatten()[..., None] @ omega[None]
+
+        return torch.concat([out_h.sin(), out_h.cos(), out_w.sin(), out_w.cos()], dim=1)[None, :, :]
+
+
+class PPDocLayoutV3AIFILayer(nn.Module):
+    """
+    AIFI (Attention-based Intra-scale Feature Interaction) layer used in RT-DETR hybrid encoder.
+    """
+
+    def __init__(self, config: PPDocLayoutV3Config):
+        super().__init__()
+        self.config = config
+        self.encoder_hidden_dim = config.encoder_hidden_dim
+        self.eval_size = config.eval_size
+
+        self.position_embedding = PPDocLayoutV3SinePositionEmbedding(
+            embed_dim=self.encoder_hidden_dim,
+            temperature=config.positional_encoding_temperature,
+        )
+        self.layers = nn.ModuleList([PPDocLayoutV3EncoderLayer(config) for _ in range(config.encoder_layers)])
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> torch.Tensor:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor` of shape `(batch_size, channels, height, width)`):
+                Feature map to process.
+        """
+        batch_size = hidden_states.shape[0]
+        height, width = hidden_states.shape[2:]
+
+        hidden_states = hidden_states.flatten(2).permute(0, 2, 1)
+
+        if self.training or self.eval_size is None:
+            pos_embed = self.position_embedding(
+                width=width,
+                height=height,
+                device=hidden_states.device,
+                dtype=hidden_states.dtype,
+            )
+        else:
+            pos_embed = None
+
+        for layer in self.layers:
+            hidden_states = layer(
+                hidden_states,
+                attention_mask=None,
+                spatial_position_embeddings=pos_embed,
+                **kwargs,
+            )
+
+        hidden_states = (
+            hidden_states.permute(0, 2, 1).reshape(batch_size, self.encoder_hidden_dim, height, width).contiguous()
+        )
+
+        return hidden_states
+
+
+class PPDocLayoutV3HybridEncoder(PPDocLayoutV3PreTrainedModel):
+    """
+    Main difference to `RTDetrHybridEncoder`:
+        1. Mask Feature Head: Added `PPDocLayoutV3MaskFeatFPN` module (`self.mask_feature_head`) for document - specific mask feature generation.
+        2. Extra Conv Layers: Introduced `self.encoder_mask_lateral` and `self.encoder_mask_output` for mask feature processing and output.
+    """
+
+    _can_record_outputs = {
+        "hidden_states": PPDocLayoutV3AIFILayer,
+        "attentions": PPDocLayoutV3SelfAttention,
+    }
+
+    def __init__(self, config: PPDocLayoutV3Config):
+        super().__init__(config)
+        self.config = config
+        self.in_channels = config.encoder_in_channels
+        self.feat_strides = config.feat_strides
+        self.encoder_hidden_dim = config.encoder_hidden_dim
+        self.encode_proj_layers = config.encode_proj_layers
+        self.positional_encoding_temperature = config.positional_encoding_temperature
+        self.eval_size = config.eval_size
+        self.out_channels = [self.encoder_hidden_dim for _ in self.in_channels]
+        self.out_strides = self.feat_strides
+        self.num_fpn_stages = len(self.in_channels) - 1
+        self.num_pan_stages = len(self.in_channels) - 1
+
+        # AIFI (Attention-based Intra-scale Feature Interaction) layers
+        self.aifi = nn.ModuleList([PPDocLayoutV3AIFILayer(config) for _ in range(len(self.encode_proj_layers))])
+
+        # top-down FPN
+        self.lateral_convs = nn.ModuleList()
+        self.fpn_blocks = nn.ModuleList()
+        for _ in range(self.num_fpn_stages):
+            lateral_conv = PPDocLayoutV3ConvNormLayer(
+                config,
+                in_channels=self.encoder_hidden_dim,
+                out_channels=self.encoder_hidden_dim,
+                kernel_size=1,
+                stride=1,
+                activation=config.activation_function,
+            )
+            fpn_block = PPDocLayoutV3CSPRepLayer(config)
+            self.lateral_convs.append(lateral_conv)
+            self.fpn_blocks.append(fpn_block)
+
+        # bottom-up PAN
+        self.downsample_convs = nn.ModuleList()
+        self.pan_blocks = nn.ModuleList()
+        for _ in range(self.num_pan_stages):
+            downsample_conv = PPDocLayoutV3ConvNormLayer(
+                config,
+                in_channels=self.encoder_hidden_dim,
+                out_channels=self.encoder_hidden_dim,
+                kernel_size=3,
+                stride=2,
+                activation=config.activation_function,
+            )
+            pan_block = PPDocLayoutV3CSPRepLayer(config)
+            self.downsample_convs.append(downsample_conv)
+            self.pan_blocks.append(pan_block)
+
+        feat_strides = config.feat_strides
+        mask_feature_channels = config.mask_feature_channels
+        self.mask_feature_head = PPDocLayoutV3MaskFeatFPN(
+            [self.encoder_hidden_dim] * len(feat_strides),
+            feat_strides,
+            feature_channels=mask_feature_channels[0],
+            out_channels=mask_feature_channels[1],
+        )
+        self.encoder_mask_lateral = PPDocLayoutV3ConvLayer(config.x4_feat_dim, mask_feature_channels[1], 3, 1, "silu")
+        self.encoder_mask_output = PPDocLayoutV3EncoderMaskOutput(
+            in_channels=mask_feature_channels[1], num_prototypes=config.num_prototypes
+        )
+
+        self.post_init()
+
+    @check_model_inputs(tie_last_hidden_states=False)
+    def forward(
+        self,
+        inputs_embeds=None,
+        x4_feat=None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> BaseModelOutput:
+        r"""
+        Args:
+            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
+                Flattened feature map (output of the backbone + projection layer) that is passed to the encoder.
+        """
+        feature_maps = inputs_embeds
+
+        # AIFI: Apply transformer encoder to specified feature levels
+        if self.config.encoder_layers > 0:
+            for i, enc_ind in enumerate(self.encode_proj_layers):
+                feature_maps[enc_ind] = self.aifi[i](feature_maps[enc_ind], **kwargs)
+
+        # top-down FPN
+        fpn_feature_maps = [feature_maps[-1]]
+        for idx, (lateral_conv, fpn_block) in enumerate(zip(self.lateral_convs, self.fpn_blocks)):
+            backbone_feature_map = feature_maps[self.num_fpn_stages - idx - 1]
+            top_fpn_feature_map = fpn_feature_maps[-1]
+            # apply lateral block
+            top_fpn_feature_map = lateral_conv(top_fpn_feature_map)
+            fpn_feature_maps[-1] = top_fpn_feature_map
+            # apply fpn block
+            top_fpn_feature_map = F.interpolate(top_fpn_feature_map, scale_factor=2.0, mode="nearest")
+            fused_feature_map = torch.concat([top_fpn_feature_map, backbone_feature_map], dim=1)
+            new_fpn_feature_map = fpn_block(fused_feature_map)
+            fpn_feature_maps.append(new_fpn_feature_map)
+
+        fpn_feature_maps.reverse()
+
+        # bottom-up PAN
+        pan_feature_maps = [fpn_feature_maps[0]]
+        for idx, (downsample_conv, pan_block) in enumerate(zip(self.downsample_convs, self.pan_blocks)):
+            top_pan_feature_map = pan_feature_maps[-1]
+            fpn_feature_map = fpn_feature_maps[idx + 1]
+            downsampled_feature_map = downsample_conv(top_pan_feature_map)
+            fused_feature_map = torch.concat([downsampled_feature_map, fpn_feature_map], dim=1)
+            new_pan_feature_map = pan_block(fused_feature_map)
+            pan_feature_maps.append(new_pan_feature_map)
+
+        mask_feat = self.mask_feature_head(pan_feature_maps)
+        mask_feat = F.interpolate(mask_feat, scale_factor=2, mode="bilinear", align_corners=False)
+        mask_feat += self.encoder_mask_lateral(x4_feat[0])
+        mask_feat = self.encoder_mask_output(mask_feat)
+
+        return PPDocLayoutV3HybridEncoderOutput(
+            last_hidden_state=pan_feature_maps,
+            mask_feat=mask_feat,
+        )
+
+
+class PPDocLayoutV3DecoderLayer(nn.Module):
+    def __init__(self, config: PPDocLayoutV3Config):
+        super().__init__()
+        self.hidden_size = config.d_model
+
+        # self-attention
+        self.self_attn = PPDocLayoutV3SelfAttention(
+            config=config,
+            hidden_size=self.hidden_size,
+            num_attention_heads=config.decoder_attention_heads,
+            dropout=config.attention_dropout,
+        )
+        self.dropout = config.dropout
+
+        self.self_attn_layer_norm = nn.LayerNorm(self.hidden_size, eps=config.layer_norm_eps)
+        # cross-attention
+        self.encoder_attn = PPDocLayoutV3MultiscaleDeformableAttention(
+            config,
+            num_heads=config.decoder_attention_heads,
+            n_points=config.decoder_n_points,
+        )
+        self.encoder_attn_layer_norm = nn.LayerNorm(self.hidden_size, eps=config.layer_norm_eps)
+        # feedforward neural networks
+        self.mlp = PPDocLayoutV3MLP(
+            config, self.hidden_size, config.decoder_ffn_dim, config.decoder_activation_function
+        )
+        self.final_layer_norm = nn.LayerNorm(self.hidden_size, eps=config.layer_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        object_queries_position_embeddings: torch.Tensor | None = None,
+        reference_points=None,
+        spatial_shapes=None,
+        spatial_shapes_list=None,
+        level_start_index=None,
+        encoder_hidden_states: torch.Tensor | None = None,
+        encoder_attention_mask: torch.Tensor | None = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> torch.Tensor:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`):
+                Input to the layer of shape `(batch, seq_len, hidden_size)`.
+            object_queries_position_embeddings (`torch.FloatTensor`, *optional*):
+                Position embeddings for the object query slots. These are added to both queries and keys
+                in the self-attention layer (not values).
+            reference_points (`torch.FloatTensor`, *optional*):
+                Reference points.
+            spatial_shapes (`torch.LongTensor`, *optional*):
+                Spatial shapes.
+            level_start_index (`torch.LongTensor`, *optional*):
+                Level start index.
+            encoder_hidden_states (`torch.FloatTensor`):
+                cross attention input to the layer of shape `(batch, seq_len, hidden_size)`
+            encoder_attention_mask (`torch.FloatTensor`): encoder attention mask of size
+                `(batch, 1, target_len, source_len)` where padding elements are indicated by very large negative
+                values.
+        """
+        residual = hidden_states
+
+        # Self Attention
+        hidden_states, _ = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=encoder_attention_mask,
+            position_embeddings=object_queries_position_embeddings,
+            **kwargs,
+        )
+
+        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+        hidden_states = residual + hidden_states
+        hidden_states = self.self_attn_layer_norm(hidden_states)
+
+        residual = hidden_states
+
+        # Cross-Attention
+        hidden_states, _ = self.encoder_attn(
+            hidden_states=hidden_states,
+            encoder_hidden_states=encoder_hidden_states,
+            position_embeddings=object_queries_position_embeddings,
+            reference_points=reference_points,
+            spatial_shapes=spatial_shapes,
+            spatial_shapes_list=spatial_shapes_list,
+            level_start_index=level_start_index,
+            **kwargs,
+        )
+
+        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+        hidden_states = residual + hidden_states
+
+        hidden_states = self.encoder_attn_layer_norm(hidden_states)
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+        hidden_states = self.final_layer_norm(hidden_states)
+
+        return hidden_states
+
+
+def inverse_sigmoid(x, eps=1e-5):
+    x = x.clamp(min=0, max=1)
+    x1 = x.clamp(min=eps)
+    x2 = (1 - x).clamp(min=eps)
+    return torch.log(x1 / x2)
+
+
+class PPDocLayoutV3Decoder(PPDocLayoutV3PreTrainedModel):
+    """
+    Main difference to `RTDetrDecoder`:
+        A new mask generation process is introduced at each decoder layer.
+    """
+
+    _can_record_outputs = {
+        "hidden_states": PPDocLayoutV3DecoderLayer,
+        "attentions": PPDocLayoutV3SelfAttention,
+        "cross_attentions": PPDocLayoutV3MultiscaleDeformableAttention,
+    }
+
+    def __init__(self, config: PPDocLayoutV3Config):
+        super().__init__(config)
+
+        self.dropout = config.dropout
+        self.layers = nn.ModuleList([PPDocLayoutV3DecoderLayer(config) for _ in range(config.decoder_layers)])
+        self.query_pos_head = PPDocLayoutV3MLPPredictionHead(4, 2 * config.d_model, config.d_model, num_layers=2)
+
+        # hack implementation for iterative bounding box refinement and two-stage Deformable DETR
+        self.bbox_embed = None
+        self.class_embed = None
+
+        self.num_queries = config.num_queries
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @check_model_inputs()
+    def forward(
+        self,
+        inputs_embeds=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        reference_points=None,
+        spatial_shapes=None,
+        spatial_shapes_list=None,
+        level_start_index=None,
+        order_head=None,
+        global_pointer=None,
+        mask_query_head=None,
+        norm=None,
+        mask_feat=None,
+        **kwargs: Unpack[TransformersKwargs],
+    ):
+        r"""
+        Args:
+            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, num_queries, hidden_size)`):
+                The query embeddings that are passed into the decoder.
+            encoder_hidden_states (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+                Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention
+                of the decoder.
+            encoder_attention_mask (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Mask to avoid performing cross-attention on padding pixel_values of the encoder. Mask values selected
+                in `[0, 1]`:
+                - 1 for pixels that are real (i.e. **not masked**),
+                - 0 for pixels that are padding (i.e. **masked**).
+            reference_points (`torch.FloatTensor` of shape `(batch_size, num_queries, 4)` is `as_two_stage` else `(batch_size, num_queries, 2)` or , *optional*):
+                Reference point in range `[0, 1]`, top-left (0,0), bottom-right (1, 1), including padding area.
+            spatial_shapes (`torch.FloatTensor` of shape `(num_feature_levels, 2)`):
+                Spatial shapes of the feature maps.
+            level_start_index (`torch.LongTensor` of shape `(num_feature_levels)`, *optional*):
+                Indexes for the start of each feature level. In range `[0, sequence_length]`.
+        """
+        if inputs_embeds is not None:
+            hidden_states = inputs_embeds
+
+        # decoder layers
+        intermediate = ()
+        intermediate_reference_points = ()
+        intermediate_logits = ()
+        decoder_out_order_logits = ()
+        decoder_out_masks = ()
+
+        reference_points = F.sigmoid(reference_points)
+
+        # https://github.com/lyuwenyu/RT-DETR/blob/94f5e16708329d2f2716426868ec89aa774af016/rtdetr_pytorch/src/zoo/rtdetr/rtdetr_decoder.py#L252
+        for idx, decoder_layer in enumerate(self.layers):
+            reference_points_input = reference_points.unsqueeze(2)
+            object_queries_position_embeddings = self.query_pos_head(reference_points)
+
+            hidden_states = decoder_layer(
+                hidden_states,
+                object_queries_position_embeddings=object_queries_position_embeddings,
+                encoder_hidden_states=encoder_hidden_states,
+                reference_points=reference_points_input,
+                spatial_shapes=spatial_shapes,
+                spatial_shapes_list=spatial_shapes_list,
+                level_start_index=level_start_index,
+                encoder_attention_mask=encoder_attention_mask,
+                **kwargs,
+            )
+
+            # hack implementation for iterative bounding box refinement
+            if self.bbox_embed is not None:
+                predicted_corners = self.bbox_embed(hidden_states)
+                new_reference_points = F.sigmoid(predicted_corners + inverse_sigmoid(reference_points))
+                reference_points = new_reference_points.detach()
+
+            intermediate += (hidden_states,)
+            intermediate_reference_points += (
+                (new_reference_points,) if self.bbox_embed is not None else (reference_points,)
+            )
+
+            # get_pred_class_order_and_mask
+            out_query = norm(hidden_states)
+            mask_query_embed = mask_query_head(out_query)
+            batch_size, mask_dim, _ = mask_query_embed.shape
+            _, _, mask_h, mask_w = mask_feat.shape
+            out_mask = torch.bmm(mask_query_embed, mask_feat.flatten(start_dim=2)).reshape(
+                batch_size, mask_dim, mask_h, mask_w
+            )
+            decoder_out_masks += (out_mask,)
+
+            if self.class_embed is not None:
+                logits = self.class_embed(out_query)
+                intermediate_logits += (logits,)
+
+            if order_head is not None and global_pointer is not None:
+                valid_query = out_query[:, -self.num_queries :] if self.num_queries is not None else out_query
+                order_logits = global_pointer(order_head[idx](valid_query))
+                decoder_out_order_logits += (order_logits,)
+
+        # Keep batch_size as first dimension
+        intermediate = torch.stack(intermediate, dim=1)
+        intermediate_reference_points = torch.stack(intermediate_reference_points, dim=1)
+        if self.class_embed is not None:
+            intermediate_logits = torch.stack(intermediate_logits, dim=1)
+        if order_head is not None and global_pointer is not None:
+            decoder_out_order_logits = torch.stack(decoder_out_order_logits, dim=1)
+        decoder_out_masks = torch.stack(decoder_out_masks, dim=1)
+
+        return PPDocLayoutV3DecoderOutput(
+            last_hidden_state=hidden_states,
+            intermediate_hidden_states=intermediate,
+            intermediate_logits=intermediate_logits,
+            intermediate_reference_points=intermediate_reference_points,
+            decoder_out_order_logits=decoder_out_order_logits,
+            decoder_out_masks=decoder_out_masks,
+        )
+
+
+class PPDocLayoutV3FrozenBatchNorm2d(nn.Module):
+    """
+    BatchNorm2d where the batch statistics and the affine parameters are fixed.
+
+    Copy-paste from torchvision.misc.ops with added eps before rqsrt, without which any other models than
+    torchvision.models.resnet[18,34,50,101] produce nans.
+    """
+
+    def __init__(self, n):
+        super().__init__()
+        self.register_buffer("weight", torch.ones(n))
+        self.register_buffer("bias", torch.zeros(n))
+        self.register_buffer("running_mean", torch.zeros(n))
+        self.register_buffer("running_var", torch.ones(n))
+
+    def _load_from_state_dict(
+        self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs
+    ):
+        num_batches_tracked_key = prefix + "num_batches_tracked"
+        if num_batches_tracked_key in state_dict:
+            del state_dict[num_batches_tracked_key]
+
+        super()._load_from_state_dict(
+            state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs
+        )
+
+    def forward(self, x):
+        # move reshapes to the beginning
+        # to make it user-friendly
+        weight = self.weight.reshape(1, -1, 1, 1)
+        bias = self.bias.reshape(1, -1, 1, 1)
+        running_var = self.running_var.reshape(1, -1, 1, 1)
+        running_mean = self.running_mean.reshape(1, -1, 1, 1)
+        epsilon = 1e-5
+        scale = weight * (running_var + epsilon).rsqrt()
+        bias = bias - running_mean * scale
+        return x * scale + bias
+
+
+def replace_batch_norm(model):
+    r"""
+    Recursively replace all `torch.nn.BatchNorm2d` with `PPDocLayoutV3FrozenBatchNorm2d`.
+
+    Args:
+        model (torch.nn.Module):
+            input model
+    """
+    for name, module in model.named_children():
+        if isinstance(module, nn.BatchNorm2d):
+            new_module = PPDocLayoutV3FrozenBatchNorm2d(module.num_features)
+
+            if module.weight.device != torch.device("meta"):
+                new_module.weight.copy_(module.weight)
+                new_module.bias.copy_(module.bias)
+                new_module.running_mean.copy_(module.running_mean)
+                new_module.running_var.copy_(module.running_var)
+
+            model._modules[name] = new_module
+
+        if len(list(module.children())) > 0:
+            replace_batch_norm(module)
+
+
+class PPDocLayoutV3ConvEncoder(nn.Module):
+    """
+    Convolutional backbone using the modeling_pp_doclayout_v3_resnet.py.
+
+    nn.BatchNorm2d layers are replaced by PPDocLayoutV3FrozenBatchNorm2d as defined above.
+    https://github.com/lyuwenyu/RT-DETR/blob/main/PPDocLayoutV3_pytorch/src/nn/backbone/presnet.py#L142
+    """
+
+    def __init__(self, config):
+        super().__init__()
+
+        backbone = load_backbone(config)
+
+        if config.freeze_backbone_batch_norms:
+            # replace batch norm by frozen batch norm
+            with torch.no_grad():
+                replace_batch_norm(backbone)
+        self.model = backbone
+        self.intermediate_channel_sizes = self.model.channels
+
+    def forward(self, pixel_values: torch.Tensor, pixel_mask: torch.Tensor):
+        # send pixel_values through the model to get list of feature maps
+        features = self.model(pixel_values).feature_maps
+
+        out = []
+        for feature_map in features:
+            # downsample pixel_mask to match shape of corresponding feature_map
+            mask = nn.functional.interpolate(pixel_mask[None].float(), size=feature_map.shape[-2:]).to(torch.bool)[0]
+            out.append((feature_map, mask))
+        return out
+
+
+def get_contrastive_denoising_training_group(
+    targets,
+    num_classes,
+    num_queries,
+    class_embed,
+    num_denoising_queries=100,
+    label_noise_ratio=0.5,
+    box_noise_scale=1.0,
+):
+    """
+    Creates a contrastive denoising training group using ground-truth samples. It adds noise to labels and boxes.
+
+    Args:
+        targets (`list[dict]`):
+            The target objects, each containing 'class_labels' and 'boxes' for objects in an image.
+        num_classes (`int`):
+            Total number of classes in the dataset.
+        num_queries (`int`):
+            Number of query slots in the transformer.
+        class_embed (`callable`):
+            A function or a model layer to embed class labels.
+        num_denoising_queries (`int`, *optional*, defaults to 100):
+            Number of denoising queries.
+        label_noise_ratio (`float`, *optional*, defaults to 0.5):
+            Ratio of noise applied to labels.
+        box_noise_scale (`float`, *optional*, defaults to 1.0):
+            Scale of noise applied to bounding boxes.
+    Returns:
+        `tuple` comprising various elements:
+        - **input_query_class** (`torch.FloatTensor`) --
+          Class queries with applied label noise.
+        - **input_query_bbox** (`torch.FloatTensor`) --
+          Bounding box queries with applied box noise.
+        - **attn_mask** (`torch.FloatTensor`) --
+           Attention mask for separating denoising and reconstruction queries.
+        - **denoising_meta_values** (`dict`) --
+          Metadata including denoising positive indices, number of groups, and split sizes.
+    """
+
+    if num_denoising_queries <= 0:
+        return None, None, None, None
+
+    num_ground_truths = [len(t["class_labels"]) for t in targets]
+    device = targets[0]["class_labels"].device
+
+    max_gt_num = max(num_ground_truths)
+    if max_gt_num == 0:
+        return None, None, None, None
+
+    num_groups_denoising_queries = num_denoising_queries // max_gt_num
+    num_groups_denoising_queries = 1 if num_groups_denoising_queries == 0 else num_groups_denoising_queries
+    # pad gt to max_num of a batch
+    batch_size = len(num_ground_truths)
+
+    input_query_class = torch.full([batch_size, max_gt_num], num_classes, dtype=torch.int32, device=device)
+    input_query_bbox = torch.zeros([batch_size, max_gt_num, 4], device=device)
+    pad_gt_mask = torch.zeros([batch_size, max_gt_num], dtype=torch.bool, device=device)
+
+    for i in range(batch_size):
+        num_gt = num_ground_truths[i]
+        if num_gt > 0:
+            input_query_class[i, :num_gt] = targets[i]["class_labels"]
+            input_query_bbox[i, :num_gt] = targets[i]["boxes"]
+            pad_gt_mask[i, :num_gt] = 1
+    # each group has positive and negative queries.
+    input_query_class = input_query_class.tile([1, 2 * num_groups_denoising_queries])
+    input_query_bbox = input_query_bbox.tile([1, 2 * num_groups_denoising_queries, 1])
+    pad_gt_mask = pad_gt_mask.tile([1, 2 * num_groups_denoising_queries])
+    # positive and negative mask
+    negative_gt_mask = torch.zeros([batch_size, max_gt_num * 2, 1], device=device)
+    negative_gt_mask[:, max_gt_num:] = 1
+    negative_gt_mask = negative_gt_mask.tile([1, num_groups_denoising_queries, 1])
+    positive_gt_mask = 1 - negative_gt_mask
+    # contrastive denoising training positive index
+    positive_gt_mask = positive_gt_mask.squeeze(-1) * pad_gt_mask
+    denoise_positive_idx = torch.nonzero(positive_gt_mask)[:, 1]
+    denoise_positive_idx = torch.split(
+        denoise_positive_idx, [n * num_groups_denoising_queries for n in num_ground_truths]
+    )
+    # total denoising queries
+    num_denoising_queries = torch_int(max_gt_num * 2 * num_groups_denoising_queries)
+
+    if label_noise_ratio > 0:
+        mask = torch.rand_like(input_query_class, dtype=torch.float) < (label_noise_ratio * 0.5)
+        # randomly put a new one here
+        new_label = torch.randint_like(mask, 0, num_classes, dtype=input_query_class.dtype)
+        input_query_class = torch.where(mask & pad_gt_mask, new_label, input_query_class)
+
+    if box_noise_scale > 0:
+        known_bbox = center_to_corners_format(input_query_bbox)
+        diff = torch.tile(input_query_bbox[..., 2:] * 0.5, [1, 1, 2]) * box_noise_scale
+        rand_sign = torch.randint_like(input_query_bbox, 0, 2) * 2.0 - 1.0
+        rand_part = torch.rand_like(input_query_bbox)
+        rand_part = (rand_part + 1.0) * negative_gt_mask + rand_part * (1 - negative_gt_mask)
+        rand_part *= rand_sign
+        known_bbox += rand_part * diff
+        known_bbox.clip_(min=0.0, max=1.0)
+        input_query_bbox = corners_to_center_format(known_bbox)
+        input_query_bbox = inverse_sigmoid(input_query_bbox)
+
+    input_query_class = class_embed(input_query_class)
+
+    target_size = num_denoising_queries + num_queries
+    attn_mask = torch.full([target_size, target_size], 0, dtype=torch.float, device=device)
+    # match query cannot see the reconstruction
+    attn_mask[num_denoising_queries:, :num_denoising_queries] = -torch.inf
+
+    # reconstructions cannot see each other
+    for i in range(num_groups_denoising_queries):
+        idx_block_start = max_gt_num * 2 * i
+        idx_block_end = max_gt_num * 2 * (i + 1)
+        attn_mask[idx_block_start:idx_block_end, :idx_block_start] = -torch.inf
+        attn_mask[idx_block_start:idx_block_end, idx_block_end:num_denoising_queries] = -torch.inf
+
+    denoising_meta_values = {
+        "dn_positive_idx": denoise_positive_idx,
+        "dn_num_group": num_groups_denoising_queries,
+        "dn_num_split": [num_denoising_queries, num_queries],
+    }
+
+    return input_query_class, input_query_bbox, attn_mask, denoising_meta_values
+
+
+def mask_to_box_coordinate(mask, dtype):
+    mask = mask.bool()
+
+    height, width = mask.shape[-2:]
+
+    y_coords, x_coords = torch.meshgrid(
+        torch.arange(height, device=mask.device), torch.arange(width, device=mask.device), indexing="ij"
+    )
+    x_coords = x_coords.to(dtype)
+    y_coords = y_coords.to(dtype)
+
+    x_coords_masked = x_coords * mask
+    x_max = x_coords_masked.flatten(start_dim=-2).max(dim=-1).values + 1
+    x_min = (
+        torch.where(mask, x_coords_masked, torch.tensor(torch.finfo(dtype).max))
+        .flatten(start_dim=-2)
+        .min(dim=-1)
+        .values
+    )
+
+    y_coords_masked = y_coords * mask
+    y_max = y_coords_masked.flatten(start_dim=-2).max(dim=-1).values + 1
+    y_min = (
+        torch.where(mask, y_coords_masked, torch.tensor(torch.finfo(dtype).max))
+        .flatten(start_dim=-2)
+        .min(dim=-1)
+        .values
+    )
+
+    unnormalized_bbox = torch.stack([x_min, y_min, x_max, y_max], dim=-1)
+
+    is_mask_non_empty = torch.any(mask, dim=(-2, -1)).unsqueeze(-1)
+    unnormalized_bbox = unnormalized_bbox * is_mask_non_empty
+
+    norm_tensor = torch.tensor([width, height, width, height], device=mask.device, dtype=dtype)
+    normalized_bbox_xyxy = unnormalized_bbox / norm_tensor
+
+    x_min_norm, y_min_norm, x_max_norm, y_max_norm = normalized_bbox_xyxy.unbind(dim=-1)
+
+    center_x = (x_min_norm + x_max_norm) / 2
+    center_y = (y_min_norm + y_max_norm) / 2
+    box_width = x_max_norm - x_min_norm
+    box_height = y_max_norm - y_min_norm
+
+    return torch.stack([center_x, center_y, box_width, box_height], dim=-1)
+
+
+@auto_docstring(
+    custom_intro="""
+    PP-DocLayoutV3 Model (consisting of a backbone and encoder-decoder) outputting raw hidden states without any head on top.
+    """
+)
+class PPDocLayoutV3Model(PPDocLayoutV3PreTrainedModel):
+    _tied_weights_keys = {
+        "decoder.class_embed": "enc_score_head",
+        "decoder.bbox_embed": "enc_bbox_head",
+    }
+
+    def __init__(self, config: PPDocLayoutV3Config):
+        super().__init__(config)
+
+        # Create backbone
+        self.backbone = PPDocLayoutV3ConvEncoder(config)
+        intermediate_channel_sizes = self.backbone.intermediate_channel_sizes
+
+        # Create encoder input projection layers
+        # https://github.com/lyuwenyu/RT-DETR/blob/94f5e16708329d2f2716426868ec89aa774af016/PPDocLayoutV3_pytorch/src/zoo/PPDocLayoutV3/hybrid_encoder.py#L212
+        num_backbone_outs = len(intermediate_channel_sizes)
+
+        encoder_input_proj_list = []
+        for i in range(num_backbone_outs):
+            in_channels = intermediate_channel_sizes[i]
+            encoder_input_proj_list.append(
+                nn.Sequential(
+                    nn.Conv2d(in_channels, config.encoder_hidden_dim, kernel_size=1, bias=False),
+                    nn.BatchNorm2d(config.encoder_hidden_dim),
+                )
+            )
+        self.encoder_input_proj = nn.ModuleList(encoder_input_proj_list[1:])
+
+        # Create encoder
+        self.encoder = PPDocLayoutV3HybridEncoder(config)
+
+        # denoising part
+        if config.num_denoising > 0:
+            self.denoising_class_embed = nn.Embedding(
+                config.num_labels + 1, config.d_model, padding_idx=config.num_labels
+            )
+
+        # decoder embedding
+        if config.learn_initial_query:
+            self.weight_embedding = nn.Embedding(config.num_queries, config.d_model)
+
+        # encoder head
+        self.enc_output = nn.Sequential(
+            nn.Linear(config.d_model, config.d_model),
+            nn.LayerNorm(config.d_model, eps=config.layer_norm_eps),
+        )
+        self.enc_score_head = nn.Linear(config.d_model, config.num_labels)
+        self.enc_bbox_head = PPDocLayoutV3MLPPredictionHead(config.d_model, config.d_model, 4, num_layers=3)
+
+        # init encoder output anchors and valid_mask
+        if config.anchor_image_size:
+            self.anchors, self.valid_mask = self.generate_anchors(dtype=self.dtype)
+
+        # Create decoder input projection layers
+        # https://github.com/lyuwenyu/RT-DETR/blob/94f5e16708329d2f2716426868ec89aa774af016/PPDocLayoutV3_pytorch/src/zoo/PPDocLayoutV3/PPDocLayoutV3_decoder.py#L412
+        num_backbone_outs = len(config.decoder_in_channels)
+        decoder_input_proj_list = []
+        for i in range(num_backbone_outs):
+            in_channels = config.decoder_in_channels[i]
+            decoder_input_proj_list.append(
+                nn.Sequential(
+                    nn.Conv2d(in_channels, config.d_model, kernel_size=1, bias=False),
+                    nn.BatchNorm2d(config.d_model, config.batch_norm_eps),
+                )
+            )
+        for _ in range(config.num_feature_levels - num_backbone_outs):
+            decoder_input_proj_list.append(
+                nn.Sequential(
+                    nn.Conv2d(in_channels, config.d_model, kernel_size=3, stride=2, padding=1, bias=False),
+                    nn.BatchNorm2d(config.d_model, config.batch_norm_eps),
+                )
+            )
+            in_channels = config.d_model
+        self.decoder_input_proj = nn.ModuleList(decoder_input_proj_list)
+        self.decoder = PPDocLayoutV3Decoder(config)
+
+        self.decoder_order_head = nn.ModuleList(
+            [nn.Linear(config.d_model, config.d_model) for _ in range(config.decoder_layers)]
+        )
+        self.decoder_global_pointer = PPDocLayoutV3GlobalPointer(config)
+        self.decoder_norm = nn.LayerNorm(config.d_model, eps=config.layer_norm_eps)
+        self.decoder.class_embed = nn.Linear(config.d_model, config.num_labels)
+        self.decoder.bbox_embed = PPDocLayoutV3MLPPredictionHead(config.d_model, config.d_model, 4, num_layers=3)
+
+        self.mask_enhanced = config.mask_enhanced
+        self.mask_query_head = PPDocLayoutV3MLPPredictionHead(
+            config.d_model, config.d_model, config.num_prototypes, num_layers=3
+        )
+
+        self.post_init()
+
+    def freeze_backbone(self):
+        for param in self.backbone.parameters():
+            param.requires_grad_(False)
+
+    def unfreeze_backbone(self):
+        for param in self.backbone.parameters():
+            param.requires_grad_(True)
+
+    @compile_compatible_method_lru_cache(maxsize=32)
+    def generate_anchors(self, spatial_shapes=None, grid_size=0.05, device="cpu", dtype=torch.float32):
+        if spatial_shapes is None:
+            spatial_shapes = [
+                [int(self.config.anchor_image_size[0] / s), int(self.config.anchor_image_size[1] / s)]
+                for s in self.config.feat_strides
+            ]
+        anchors = []
+        for level, (height, width) in enumerate(spatial_shapes):
+            grid_y, grid_x = torch.meshgrid(
+                torch.arange(end=height, device=device).to(dtype),
+                torch.arange(end=width, device=device).to(dtype),
+                indexing="ij",
+            )
+            grid_xy = torch.stack([grid_x, grid_y], -1)
+            grid_xy = grid_xy.unsqueeze(0) + 0.5
+            grid_xy[..., 0] /= width
+            grid_xy[..., 1] /= height
+            wh = torch.ones_like(grid_xy) * grid_size * (2.0**level)
+            anchors.append(torch.concat([grid_xy, wh], -1).reshape(-1, height * width, 4))
+        # define the valid range for anchor coordinates
+        eps = 1e-2
+        anchors = torch.concat(anchors, 1)
+        valid_mask = ((anchors > eps) * (anchors < 1 - eps)).all(-1, keepdim=True)
+        anchors = torch.log(anchors / (1 - anchors))
+        anchors = torch.where(valid_mask, anchors, torch.tensor(torch.finfo(dtype).max, dtype=dtype, device=device))
+
+        return anchors, valid_mask
+
+    @auto_docstring
+    @can_return_tuple
+    def forward(
+        self,
+        pixel_values: torch.FloatTensor,
+        pixel_mask: torch.LongTensor | None = None,
+        encoder_outputs: torch.FloatTensor | None = None,
+        labels: list[dict] | None = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple[torch.FloatTensor] | PPDocLayoutV3ModelOutput:
+        r"""
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing the flattened feature map (output of the backbone + projection layer), you
+            can choose to directly pass a flattened representation of an image.
+        decoder_inputs_embeds (`torch.FloatTensor` of shape `(batch_size, num_queries, hidden_size)`, *optional*):
+            Optionally, instead of initializing the queries with a tensor of zeros, you can choose to directly pass an
+            embedded representation.
+        labels (`list[Dict]` of len `(batch_size,)`, *optional*):
+            Labels for computing the bipartite matching loss. List of dicts, each dictionary containing at least the
+            following 2 keys: 'class_labels' and 'boxes' (the class labels and bounding boxes of an image in the batch
+            respectively). The class labels themselves should be a `torch.LongTensor` of len `(number of bounding boxes
+            in the image,)` and the boxes a `torch.FloatTensor` of shape `(number of bounding boxes in the image, 4)`.
+
+        Examples:
+
+        ```python
+        >>> from transformers import AutoImageProcessor, PPDocLayoutV2Model
+        >>> from PIL import Image
+        >>> import requests
+
+        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+
+        >>> image_processor = AutoImageProcessor.from_pretrained("PekingU/PPDocLayoutV2_r50vd")
+        >>> model = PPDocLayoutV2Model.from_pretrained("PekingU/PPDocLayoutV2_r50vd")
+
+        >>> inputs = image_processor(images=image, return_tensors="pt")
+
+        >>> outputs = model(**inputs)
+
+        >>> last_hidden_states = outputs.last_hidden_state
+        >>> list(last_hidden_states.shape)
+        [1, 300, 256]
+        ```"""
+        batch_size, num_channels, height, width = pixel_values.shape
+        device = pixel_values.device
+
+        if pixel_mask is None:
+            pixel_mask = torch.ones(((batch_size, height, width)), device=device)
+
+        features = self.backbone(pixel_values, pixel_mask)
+        x4_feat = features.pop(0)
+        proj_feats = [self.encoder_input_proj[level](source) for level, (source, mask) in enumerate(features)]
+
+        if encoder_outputs is None:
+            encoder_outputs = self.encoder(
+                proj_feats,
+                x4_feat,
+                **kwargs,
+            )
+        # If the user passed a tuple for encoder_outputs, we wrap it in a PPDocLayoutV3HybridEncoderOutput when return_dict=True
+        elif not isinstance(encoder_outputs, PPDocLayoutV3HybridEncoderOutput):
+            encoder_outputs = PPDocLayoutV3HybridEncoderOutput(
+                last_hidden_state=encoder_outputs[0],
+                hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None,
+                attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None,
+                mask_feat=encoder_outputs[-1],
+            )
+
+        # Equivalent to def _get_encoder_input
+        # https://github.com/lyuwenyu/RT-DETR/blob/94f5e16708329d2f2716426868ec89aa774af016/rtdetr_pytorch/src/zoo/rtdetr/rtdetr_decoder.py#L412
+        sources = []
+        for level, source in enumerate(encoder_outputs.last_hidden_state):
+            sources.append(self.decoder_input_proj[level](source))
+
+        # Lowest resolution feature maps are obtained via 3x3 stride 2 convolutions on the final stage
+        if self.config.num_feature_levels > len(sources):
+            _len_sources = len(sources)
+            sources.append(self.decoder_input_proj[_len_sources](encoder_outputs.last_hidden_state[-1]))
+            for i in range(_len_sources + 1, self.config.num_feature_levels):
+                sources.append(self.decoder_input_proj[i](encoder_outputs.last_hidden_state[-1]))
+
+        # Prepare encoder inputs (by flattening)
+        source_flatten = []
+        spatial_shapes_list = []
+        spatial_shapes = torch.empty((len(sources), 2), device=device, dtype=torch.long)
+        for level, source in enumerate(sources):
+            height, width = source.shape[-2:]
+            spatial_shapes[level, 0] = height
+            spatial_shapes[level, 1] = width
+            spatial_shapes_list.append((height, width))
+            source = source.flatten(2).transpose(1, 2)
+            source_flatten.append(source)
+        source_flatten = torch.cat(source_flatten, 1)
+        level_start_index = torch.cat((spatial_shapes.new_zeros((1,)), spatial_shapes.prod(1).cumsum(0)[:-1]))
+
+        # prepare denoising training
+        if self.training and self.config.num_denoising > 0 and labels is not None:
+            (
+                denoising_class,
+                denoising_bbox_unact,
+                attention_mask,
+                denoising_meta_values,
+            ) = get_contrastive_denoising_training_group(
+                targets=labels,
+                num_classes=self.config.num_labels,
+                num_queries=self.config.num_queries,
+                class_embed=self.denoising_class_embed,
+                num_denoising_queries=self.config.num_denoising,
+                label_noise_ratio=self.config.label_noise_ratio,
+                box_noise_scale=self.config.box_noise_scale,
+            )
+        else:
+            denoising_class, denoising_bbox_unact, attention_mask, denoising_meta_values = None, None, None, None
+
+        batch_size = len(source_flatten)
+        device = source_flatten.device
+        dtype = source_flatten.dtype
+
+        # prepare input for decoder
+        if self.training or self.config.anchor_image_size is None:
+            # Pass spatial_shapes as tuple to make it hashable and make sure
+            # lru_cache is working for generate_anchors()
+            spatial_shapes_tuple = tuple(spatial_shapes_list)
+            anchors, valid_mask = self.generate_anchors(spatial_shapes_tuple, device=device, dtype=dtype)
+        else:
+            anchors, valid_mask = self.anchors, self.valid_mask
+            anchors, valid_mask = anchors.to(device, dtype), valid_mask.to(device, dtype)
+
+        # use the valid_mask to selectively retain values in the feature map where the mask is `True`
+        memory = valid_mask.to(source_flatten.dtype) * source_flatten
+
+        output_memory = self.enc_output(memory)
+
+        enc_outputs_class = self.enc_score_head(output_memory)
+        enc_outputs_coord_logits = self.enc_bbox_head(output_memory) + anchors
+
+        _, topk_ind = torch.topk(enc_outputs_class.max(-1).values, self.config.num_queries, dim=1)
+
+        reference_points_unact = enc_outputs_coord_logits.gather(
+            dim=1, index=topk_ind.unsqueeze(-1).repeat(1, 1, enc_outputs_coord_logits.shape[-1])
+        )
+
+        # _get_pred_class_and_mask
+        batch_ind = torch.arange(memory.shape[0], device=output_memory.device).unsqueeze(1)
+        target = output_memory[batch_ind, topk_ind]
+        out_query = self.decoder_norm(target)
+        mask_query_embed = self.mask_query_head(out_query)
+        batch_size, mask_dim, _ = mask_query_embed.shape
+
+        enc_topk_bboxes = F.sigmoid(reference_points_unact)
+
+        enc_topk_logits = enc_outputs_class.gather(
+            dim=1, index=topk_ind.unsqueeze(-1).repeat(1, 1, enc_outputs_class.shape[-1])
+        )
+
+        # extract region features
+        if self.config.learn_initial_query:
+            target = self.weight_embedding.tile([batch_size, 1, 1])
+        else:
+            target = output_memory.gather(dim=1, index=topk_ind.unsqueeze(-1).repeat(1, 1, output_memory.shape[-1]))
+            target = target.detach()
+
+        if denoising_class is not None:
+            target = torch.concat([denoising_class, target], 1)
+
+        if self.mask_enhanced:
+            _, _, mask_h, mask_w = encoder_outputs.mask_feat.shape
+            enc_out_masks = torch.bmm(mask_query_embed, encoder_outputs.mask_feat.flatten(start_dim=2)).reshape(
+                batch_size, mask_dim, mask_h, mask_w
+            )
+            reference_points = mask_to_box_coordinate(enc_out_masks > 0, dtype=reference_points_unact.dtype)
+            reference_points_unact = inverse_sigmoid(reference_points)
+
+        if denoising_bbox_unact is not None:
+            reference_points_unact = torch.concat([denoising_bbox_unact, reference_points_unact], 1)
+
+        init_reference_points = reference_points_unact.detach()
+
+        # decoder
+        decoder_outputs = self.decoder(
+            inputs_embeds=target,
+            encoder_hidden_states=source_flatten,
+            encoder_attention_mask=attention_mask,
+            reference_points=init_reference_points,
+            spatial_shapes=spatial_shapes,
+            spatial_shapes_list=spatial_shapes_list,
+            level_start_index=level_start_index,
+            order_head=self.decoder_order_head,
+            global_pointer=self.decoder_global_pointer,
+            mask_query_head=self.mask_query_head,
+            norm=self.decoder_norm,
+            mask_feat=encoder_outputs.mask_feat,
+            **kwargs,
+        )
+
+        return PPDocLayoutV3ModelOutput(
+            last_hidden_state=decoder_outputs.last_hidden_state,
+            intermediate_hidden_states=decoder_outputs.intermediate_hidden_states,
+            intermediate_logits=decoder_outputs.intermediate_logits,
+            intermediate_reference_points=decoder_outputs.intermediate_reference_points,
+            intermediate_predicted_corners=decoder_outputs.intermediate_predicted_corners,
+            initial_reference_points=decoder_outputs.initial_reference_points,
+            decoder_hidden_states=decoder_outputs.hidden_states,
+            decoder_attentions=decoder_outputs.attentions,
+            cross_attentions=decoder_outputs.cross_attentions,
+            out_order_logits=decoder_outputs.decoder_out_order_logits,
+            out_masks=decoder_outputs.decoder_out_masks,
+            encoder_last_hidden_state=encoder_outputs.last_hidden_state,
+            encoder_hidden_states=encoder_outputs.hidden_states,
+            encoder_attentions=encoder_outputs.attentions,
+            init_reference_points=init_reference_points,
+            enc_topk_logits=enc_topk_logits,
+            enc_topk_bboxes=enc_topk_bboxes,
+            enc_outputs_class=enc_outputs_class,
+            enc_outputs_coord_logits=enc_outputs_coord_logits,
+            denoising_meta_values=denoising_meta_values,
+        )
+
+
+@dataclass
+@auto_docstring
+class PPDocLayoutV3HybridEncoderOutput(BaseModelOutput):
+    r"""
+    mask_feat (`torch.FloatTensor` of shape `(batch_size, config.num_queries, 200, 200)`):
+        Mask features for each query in the batch.
+    """
+
+    mask_feat: torch.FloatTensor = None
+
+
+@dataclass
+@auto_docstring
+class PPDocLayoutV3ForObjectDetectionOutput(ModelOutput):
+    r"""
+    logits (`torch.FloatTensor` of shape `(batch_size, num_queries, num_classes + 1)`):
+        Classification logits (including no-object) for all queries.
+    pred_boxes (`torch.FloatTensor` of shape `(batch_size, num_queries, 4)`):
+        Normalized boxes coordinates for all queries, represented as (center_x, center_y, width, height). These
+        values are normalized in [0, 1], relative to the size of each individual image in the batch (disregarding
+        possible padding). You can use [`~PPDocLayoutV3ImageProcessorFast.post_process_object_detection`] to retrieve the
+        unnormalized (absolute) bounding boxes.
+    order_logits (`tuple` of `torch.FloatTensor` of shape `(batch_size, num_queries, num_queries)`):
+        Order logits of the final layer of the decoder.
+    out_masks (`torch.FloatTensor` of shape `(batch_size, config.decoder_layers, num_queries, height, width)`):
+        Masks of the final layer of the decoder.
+    last_hidden_state (`torch.FloatTensor` of shape `(batch_size, num_queries, hidden_size)`):
+        Sequence of hidden-states at the output of the last layer of the decoder of the model.
+    intermediate_hidden_states (`torch.FloatTensor` of shape `(batch_size, config.decoder_layers, num_queries, hidden_size)`):
+        Stacked intermediate hidden states (output of each layer of the decoder).
+    intermediate_logits (`torch.FloatTensor` of shape `(batch_size, config.decoder_layers, num_queries, config.num_labels)`):
+        Stacked intermediate logits (logits of each layer of the decoder).
+    intermediate_reference_points (`torch.FloatTensor` of shape `(batch_size, config.decoder_layers, num_queries, 4)`):
+        Stacked intermediate reference points (reference points of each layer of the decoder).
+    intermediate_predicted_corners (`torch.FloatTensor` of shape `(batch_size, config.decoder_layers, num_queries, 4)`):
+        Stacked intermediate predicted corners (predicted corners of each layer of the decoder).
+    initial_reference_points (`torch.FloatTensor` of shape `(batch_size, config.decoder_layers, num_queries, 4)`):
+        Stacked initial reference points (initial reference points of each layer of the decoder).
+    init_reference_points (`torch.FloatTensor` of shape  `(batch_size, num_queries, 4)`):
+        Initial reference points sent through the Transformer decoder.
+    enc_topk_logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.num_labels)`, *optional*, returned when `config.with_box_refine=True` and `config.two_stage=True`):
+        Logits of predicted bounding boxes coordinates in the encoder.
+    enc_topk_bboxes (`torch.FloatTensor` of shape `(batch_size, sequence_length, 4)`, *optional*, returned when `config.with_box_refine=True` and `config.two_stage=True`):
+        Logits of predicted bounding boxes coordinates in the encoder.
+    enc_outputs_class (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.num_labels)`, *optional*, returned when `config.with_box_refine=True` and `config.two_stage=True`):
+        Predicted bounding boxes scores where the top `config.two_stage_num_proposals` scoring bounding boxes are
+        picked as region proposals in the first stage. Output of bounding box binary classification (i.e.
+        foreground and background).
+    enc_outputs_coord_logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, 4)`, *optional*, returned when `config.with_box_refine=True` and `config.two_stage=True`):
+        Logits of predicted bounding boxes coordinates in the first stage.
+    denoising_meta_values (`dict`):
+        Extra dictionary for the denoising related values
+    """
+
+    logits: torch.FloatTensor | None = None
+    pred_boxes: torch.FloatTensor | None = None
+    order_logits: torch.FloatTensor | None = None
+    out_masks: torch.FloatTensor | None = None
+    last_hidden_state: torch.FloatTensor | None = None
+    intermediate_hidden_states: torch.FloatTensor | None = None
+    intermediate_logits: torch.FloatTensor | None = None
+    intermediate_reference_points: torch.FloatTensor | None = None
+    intermediate_predicted_corners: torch.FloatTensor | None = None
+    initial_reference_points: torch.FloatTensor | None = None
+    decoder_hidden_states: tuple[torch.FloatTensor] | None = None
+    decoder_attentions: tuple[torch.FloatTensor] | None = None
+    cross_attentions: tuple[torch.FloatTensor] | None = None
+    encoder_last_hidden_state: torch.FloatTensor | None = None
+    encoder_hidden_states: tuple[torch.FloatTensor] | None = None
+    encoder_attentions: tuple[torch.FloatTensor] | None = None
+    init_reference_points: tuple[torch.FloatTensor] | None = None
+    enc_topk_logits: torch.FloatTensor | None = None
+    enc_topk_bboxes: torch.FloatTensor | None = None
+    enc_outputs_class: torch.FloatTensor | None = None
+    enc_outputs_coord_logits: torch.FloatTensor | None = None
+    denoising_meta_values: dict | None = None
+
+
+@auto_docstring(
+    custom_intro="""
+    PP-DocLayoutV3 Model (consisting of a backbone and encoder-decoder) outputs bounding boxes and logits sorted according to reading order,
+    which are further decoded into scores and classes.
+    """
+)
+class PPDocLayoutV3ForObjectDetection(PPDocLayoutV3PreTrainedModel):
+    # When using clones, all layers > 0 will be clones, but layer 0 *is* required
+    # We can't initialize the model on meta device as some weights are modified during the initialization
+    _no_split_modules = None
+    _keys_to_ignore_on_load_missing = ["num_batches_tracked", "rel_pos_y_bias", "rel_pos_x_bias"]
+
+    def __init__(self, config: PPDocLayoutV3Config):
+        super().__init__(config)
+        self.model = PPDocLayoutV3Model(config)
+
+        self.model.denoising_class_embed = nn.Embedding(config.num_labels, config.d_model)
+        self.num_queries = config.num_queries
+        # if two-stage, the last class_embed and bbox_embed is for region proposal generation
+        self.post_init()
+
+    def _set_aux_loss(self, outputs_class, outputs_coord):
+        return [{"logits": a, "pred_boxes": b} for a, b in zip(outputs_class, outputs_coord)]
+
+    @auto_docstring
+    @can_return_tuple
+    def forward(
+        self,
+        pixel_values: torch.FloatTensor,
+        pixel_mask: torch.LongTensor | None = None,
+        encoder_outputs: torch.FloatTensor | None = None,
+        labels: list[dict] | None = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple[torch.FloatTensor] | PPDocLayoutV3ForObjectDetectionOutput:
+        r"""
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing the flattened feature map (output of the backbone + projection layer), you
+            can choose to directly pass a flattened representation of an image.
+        decoder_inputs_embeds (`torch.FloatTensor` of shape `(batch_size, num_queries, hidden_size)`, *optional*):
+            Optionally, instead of initializing the queries with a tensor of zeros, you can choose to directly pass an
+            embedded representation.
+        labels (`list[Dict]` of len `(batch_size,)`, *optional*):
+            Labels for computing the bipartite matching loss. List of dicts, each dictionary containing at least the
+            following 2 keys: 'class_labels' and 'boxes' (the class labels and bounding boxes of an image in the batch
+            respectively). The class labels themselves should be a `torch.LongTensor` of len `(number of bounding boxes
+            in the image,)` and the boxes a `torch.FloatTensor` of shape `(number of bounding boxes in the image, 4)`.
+
+        Examples:
+
+        ```python
+        >>> from transformers import AutoModelForObjectDetection, AutoImageProcessor
+        >>> from PIL import Image
+        >>> import requests
+        >>> import torch
+
+        >>> url = "https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/layout_demo.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+
+        >>> model_path = "PaddlePaddle/PP-DocLayoutV3_safetensors"
+        >>> image_processor = AutoImageProcessor.from_pretrained(model_path)
+        >>> model = AutoModelForObjectDetection.from_pretrained(model_path)
+
+        >>> # prepare image for the model
+        >>> inputs = image_processor(images=[image], return_tensors="pt")
+
+        >>> # forward pass
+        >>> outputs = model(**inputs)
+
+        >>> # convert outputs (bounding boxes and class logits) to Pascal VOC format (xmin, ymin, xmax, ymax)
+        >>> results = image_processor.post_process_object_detection(outputs, target_sizes=torch.tensor([image.size[::-1]]))
+
+        >>> # print outputs
+        >>> for result in results:
+        ...     for idx, (score, label_id, box) in enumerate(zip(result["scores"], result["labels"], result["boxes"])):
+        ...         score, label = score.item(), label_id.item()
+        ...         box = [round(i, 2) for i in box.tolist()]
+        ...         print(f"Order {idx + 1}: {model.config.id2label[label]}: {score:.2f} {box}")
+        Order 1: text: 0.99 [334.95, 184.78, 897.25, 654.83]
+        Order 2: paragraph_title: 0.97 [337.28, 683.92, 869.16, 798.35]
+        Order 3: text: 0.99 [335.75, 842.82, 892.13, 1454.32]
+        Order 4: text: 0.99 [920.18, 185.28, 1476.38, 464.49]
+        Order 5: text: 0.98 [920.47, 483.68, 1480.63, 765.72]
+        Order 6: text: 0.98 [920.62, 846.8, 1482.09, 1220.67]
+        Order 7: text: 0.97 [920.92, 1239.41, 1469.55, 1378.02]
+        Order 8: footnote: 0.86 [335.03, 1614.68, 1483.33, 1731.73]
+        Order 9: footnote: 0.83 [334.64, 1756.74, 1471.78, 1845.69]
+        Order 10: text: 0.81 [336.8, 1910.52, 661.64, 1939.92]
+        Order 11: footnote: 0.96 [336.24, 2114.42, 1450.14, 2172.12]
+        Order 12: number: 0.88 [106.0, 2257.5, 135.84, 2282.18]
+        Order 13: footer: 0.93 [338.4, 2255.52, 986.15, 2284.37]
+        ```"""
+        outputs = self.model(
+            pixel_values,
+            pixel_mask=pixel_mask,
+            encoder_outputs=encoder_outputs,
+            labels=labels,
+            **kwargs,
+        )
+
+        intermediate_logits = outputs.intermediate_logits
+        intermediate_reference_points = outputs.intermediate_reference_points
+        order_logits = outputs.out_order_logits
+        out_masks = outputs.out_masks
+
+        pred_boxes = intermediate_reference_points[:, -1]
+        logits = intermediate_logits[:, -1]
+        order_logits = order_logits[:, -1]
+        out_masks = out_masks[:, -1]
+
+        if labels is not None:
+            raise ValueError("PPDocLayoutV3ForObjectDetection does not support training")
+
+        return PPDocLayoutV3ForObjectDetectionOutput(
+            logits=logits,
+            pred_boxes=pred_boxes,
+            order_logits=order_logits,
+            out_masks=out_masks,
+            last_hidden_state=outputs.last_hidden_state,
+            intermediate_hidden_states=outputs.intermediate_hidden_states,
+            intermediate_logits=outputs.intermediate_logits,
+            intermediate_reference_points=outputs.intermediate_reference_points,
+            intermediate_predicted_corners=outputs.intermediate_predicted_corners,
+            initial_reference_points=outputs.initial_reference_points,
+            decoder_hidden_states=outputs.decoder_hidden_states,
+            decoder_attentions=outputs.decoder_attentions,
+            cross_attentions=outputs.cross_attentions,
+            encoder_last_hidden_state=outputs.encoder_last_hidden_state,
+            encoder_hidden_states=outputs.encoder_hidden_states,
+            encoder_attentions=outputs.encoder_attentions,
+            init_reference_points=outputs.init_reference_points,
+            enc_topk_logits=outputs.enc_topk_logits,
+            enc_topk_bboxes=outputs.enc_topk_bboxes,
+            enc_outputs_class=outputs.enc_outputs_class,
+            enc_outputs_coord_logits=outputs.enc_outputs_coord_logits,
+            denoising_meta_values=outputs.denoising_meta_values,
+        )
+
+
+__all__ = ["PPDocLayoutV3ForObjectDetection", "PPDocLayoutV3Model", "PPDocLayoutV3PreTrainedModel"]