transformers 5.0.0rc0__py3-none-any.whl → 5.0.0rc2__py3-none-any.whl

transformers/models/paddleocr_vl/modeling_paddleocr_vl.py

@@ -0,0 +1,1682 @@
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+#           This file was automatically generated from src/transformers/models/paddleocr_vl/modular_paddleocr_vl.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_paddleocr_vl.py file directly. One of our CI enforces this.
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+# Copyright 2025 The PaddlePaddle Team and The HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from collections.abc import Callable
+from dataclasses import dataclass
+from typing import Any, Optional, Union
+
+import torch
+from torch import nn
+
+from ... import initialization as init
+from ...activations import ACT2FN, GELUActivation
+from ...cache_utils import Cache, DynamicCache
+from ...generation import GenerationMixin
+from ...integrations import use_kernel_forward_from_hub
+from ...masking_utils import create_bidirectional_mask, create_causal_mask
+from ...modeling_flash_attention_utils import FlashAttentionKwargs
+from ...modeling_layers import GradientCheckpointingLayer
+from ...modeling_outputs import BaseModelOutput, BaseModelOutputWithPast, BaseModelOutputWithPooling, ModelOutput
+from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS
+from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
+from ...processing_utils import Unpack
+from ...utils import TransformersKwargs, auto_docstring, can_return_tuple, logging, torch_int
+from ...utils.generic import check_model_inputs, maybe_autocast
+from .configuration_paddleocr_vl import PaddleOCRTextConfig, PaddleOCRVisionConfig, PaddleOCRVLConfig
+
+
+logger = logging.get_logger(__name__)
+
+
+class PaddleOCRProjector(nn.Module):
+    def __init__(self, config: PaddleOCRVLConfig):
+        super().__init__()
+        self.merge_kernel_size = (config.vision_config.spatial_merge_size, config.vision_config.spatial_merge_size)
+
+        hidden_size = config.vision_config.hidden_size * self.merge_kernel_size[0] * self.merge_kernel_size[1]
+
+        self.pre_norm = torch.nn.LayerNorm(config.vision_config.hidden_size, eps=1e-05)
+        self.linear_1 = nn.Linear(hidden_size, hidden_size, bias=True)
+        self.act = GELUActivation()
+        self.linear_2 = nn.Linear(hidden_size, config.text_config.hidden_size, bias=True)
+
+    def forward(self, image_features: torch.Tensor, image_grid_thw: torch.Tensor) -> torch.Tensor:
+        image_features_chunks = image_features.split(image_grid_thw.prod(dim=1).tolist(), dim=0)
+        m1, m2 = self.merge_kernel_size
+
+        processed_features = []
+        for image_feature, image_grid in zip(image_features_chunks, image_grid_thw):
+            image_feature = self.pre_norm(image_feature)
+            t, h, w = image_grid
+            d = image_feature.shape[-1]
+            h_block = h // m1
+            w_block = w // m2
+
+            image_feature = image_feature.reshape(t, h_block, m1, w_block, m2, d)
+            image_feature = image_feature.transpose(2, 3)
+            image_feature = image_feature.reshape(t * h_block * w_block, m1 * m2 * d)
+
+            hidden_states = self.linear_1(image_feature)
+            hidden_states = self.act(hidden_states)
+            hidden_states = self.linear_2(hidden_states)
+            processed_features.append(hidden_states)
+
+        return torch.cat(processed_features, dim=0)
+
+
+class PaddleOCRVisionRotaryEmbedding(nn.Module):
+    inv_freq: torch.Tensor  # fix linting for `register_buffer`
+
+    def __init__(self, dim: int, theta: float = 10000.0) -> None:
+        super().__init__()
+        self.dim = dim
+        self.theta = theta
+        inv_freq = 1.0 / (theta ** (torch.arange(0, dim, 2, dtype=torch.float) / dim))
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+
+    def forward(self, seqlen: int) -> torch.Tensor:
+        seq = torch.arange(seqlen, device=self.inv_freq.device, dtype=self.inv_freq.dtype)
+        freqs = torch.outer(seq, self.inv_freq)
+        return freqs
+
+
+class PaddleOCRRotaryEmbedding(nn.Module):
+    inv_freq: torch.Tensor  # fix linting for `register_buffer`
+
+    def __init__(self, config: PaddleOCRVLConfig, device=None):
+        super().__init__()
+        self.max_seq_len_cached = config.max_position_embeddings
+        self.original_max_seq_len = config.max_position_embeddings
+
+        self.config = config
+
+        self.rope_type = self.config.rope_parameters["rope_type"]
+        rope_init_fn: Callable = self.compute_default_rope_parameters
+        if self.rope_type != "default":
+            rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+        inv_freq, self.attention_scaling = rope_init_fn(self.config, device)
+
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.register_buffer("original_inv_freq", inv_freq.clone(), persistent=False)
+
+    @staticmethod
+    def compute_default_rope_parameters(
+        config: Optional[PaddleOCRVLConfig] = None,
+        device: Optional["torch.device"] = None,
+        seq_len: Optional[int] = None,
+    ) -> tuple["torch.Tensor", float]:
+        """
+        Computes the inverse frequencies according to the original RoPE implementation
+        Args:
+            config ([`~transformers.PreTrainedConfig`]):
+                The model configuration.
+            device (`torch.device`):
+                The device to use for initialization of the inverse frequencies.
+            seq_len (`int`, *optional*):
+                The current sequence length. Unused for this type of RoPE.
+        Returns:
+            Tuple of (`torch.Tensor`, `float`), containing the inverse frequencies for the RoPE embeddings and the
+            post-processing scaling factor applied to the computed cos/sin (unused in this type of RoPE).
+        """
+        base = config.rope_parameters["rope_theta"]
+        dim = getattr(config, "head_dim", None) or config.hidden_size // config.num_attention_heads
+
+        attention_factor = 1.0  # Unused in this type of RoPE
+
+        # Compute the inverse frequencies
+        inv_freq = 1.0 / (
+            base ** (torch.arange(0, dim, 2, dtype=torch.int64).to(device=device, dtype=torch.float) / dim)
+        )
+        return inv_freq, attention_factor
+
+    # Ignore copy
+    def forward(self, x, position_ids):
+        # In contrast to other models, PaddleOCR has different position ids for the grids
+        # So we expand the inv_freq to shape (3, ...)
+        inv_freq_expanded = self.inv_freq[None, None, :, None].float().expand(3, position_ids.shape[1], -1, 1)
+        position_ids_expanded = position_ids[:, :, None, :].float()  # shape (3, bs, 1, positions)
+
+        device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
+        with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
+            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(2, 3)
+            emb = torch.cat((freqs, freqs), dim=-1)
+            cos = emb.cos() * self.attention_scaling
+            sin = emb.sin() * self.attention_scaling
+
+        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+
+class PaddleOCRMLP(nn.Module):
+    def __init__(self, config: PaddleOCRTextConfig):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = config.intermediate_size
+
+        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=config.use_bias)
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=config.use_bias)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=config.use_bias)
+        self.act_fn = ACT2FN[config.hidden_act]
+
+    def forward(self, x):
+        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+        return down_proj
+
+
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """
+    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
+    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
+    """
+    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
+    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+
+def eager_attention_forward(
+    module: nn.Module,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attention_mask: Optional[torch.Tensor],
+    scaling: float,
+    dropout: float = 0.0,
+    **kwargs,
+):
+    key_states = repeat_kv(key, module.num_key_value_groups)
+    value_states = repeat_kv(value, module.num_key_value_groups)
+
+    attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
+    if attention_mask is not None:
+        causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
+        attn_weights = attn_weights + causal_mask
+
+    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
+    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
+    attn_output = torch.matmul(attn_weights, value_states)
+    attn_output = attn_output.transpose(1, 2).contiguous()
+
+    return attn_output, attn_weights
+
+
+def rotate_half(x):
+    """Rotates half the hidden dims of the input."""
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+def apply_multimodal_rotary_pos_emb(q, k, cos, sin, mrope_section, unsqueeze_dim=1):
+    """Applies Rotary Position Embedding with Multimodal Sections to the query and key tensors (https://qwenlm.github.io/blog/qwen2-vl/).
+
+    Explanation:
+        Multimodal 3D rotary position embedding is an extension to 1D rotary position embedding. The input embedding
+        sequence contains vision (images / videos) embedding and text embedding or just contains text embedding. For
+        vision embedding part, we apply rotary position embedding on temporal, height and width dimension separately.
+        Here we split the channel dimension to 3 chunks for the temporal, height and width rotary position embedding.
+        For text embedding part, we just apply 1D rotary position embedding. The three rotary position index (temporal,
+        height and width) of text embedding is always the same, so the text embedding rotary position embedding has no
+        difference with modern LLMs.
+
+    Args:
+        q (`torch.Tensor`): The query tensor.
+        k (`torch.Tensor`): The key tensor.
+        cos (`torch.Tensor`): The cosine part of the rotary embedding.
+        sin (`torch.Tensor`): The sine part of the rotary embedding.
+        position_ids (`torch.Tensor`):
+            The position indices of the tokens corresponding to the query and key tensors. For example, this can be
+            used to pass offsetted position ids when working with a KV-cache.
+        mrope_section(`List(int)`):
+            Multimodal rope section is for channel dimension of temporal, height and width in rope calculation.
+        unsqueeze_dim (`int`, *optional*, defaults to 1):
+            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+    Returns:
+        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
+    """
+    mrope_section = mrope_section * 2
+    cos = torch.cat([m[i % 3] for i, m in enumerate(cos.split(mrope_section, dim=-1))], dim=-1).unsqueeze(
+        unsqueeze_dim
+    )
+    sin = torch.cat([m[i % 3] for i, m in enumerate(sin.split(mrope_section, dim=-1))], dim=-1).unsqueeze(
+        unsqueeze_dim
+    )
+
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+class PaddleOCRAttention(nn.Module):
+    """
+    Multi-headed attention from 'Attention Is All You Need' paper. Modified to use sliding window attention: Longformer
+    and "Generating Long Sequences with Sparse Transformers".
+    """
+
+    def __init__(self, config: PaddleOCRVLConfig, layer_idx: Optional[int] = None):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        if layer_idx is None:
+            logger.warning_once(
+                f"Instantiating {self.__class__.__name__} without passing `layer_idx` is not recommended and will "
+                "to errors during the forward call, if caching is used. Please make sure to provide a `layer_idx` "
+                "when creating this class."
+            )
+
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = getattr(config, "head_dim", self.hidden_size // self.num_heads)
+        self.num_key_value_heads = config.num_key_value_heads
+        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
+        self.is_causal = True
+
+        self.attention_dropout = 0.0
+        self.rope_parameters = config.rope_parameters
+        self.scaling = self.head_dim**-0.5
+        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=config.use_bias)
+        self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.use_bias)
+        self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.use_bias)
+        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=config.use_bias)
+        self.layer_type = config.layer_types[layer_idx] if hasattr(config, "layer_types") else None
+        self.sliding_window = config.sliding_window if self.layer_type == "sliding_attention" else None
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ) -> tuple[torch.Tensor, Optional[torch.Tensor], Optional[tuple[torch.Tensor]]]:
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        query_states = query_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2)
+
+        cos, sin = position_embeddings
+        query_states, key_states = apply_multimodal_rotary_pos_emb(
+            query_states, key_states, cos, sin, self.config.rope_parameters["mrope_section"]
+        )
+
+        if past_key_values is not None:
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}  # Specific to RoPE models
+            key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        attention_interface: Callable = eager_attention_forward
+        if self.config._attn_implementation != "eager":
+            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+
+        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,
+            sliding_window=self.sliding_window,
+            position_ids=position_ids,  # pass positions for FA2
+            **kwargs,
+        )
+
+        attn_output = attn_output.reshape(bsz, q_len, -1).contiguous()
+        attn_output = self.o_proj(attn_output)
+        return attn_output, attn_weights
+
+
+@use_kernel_forward_from_hub("RMSNorm")
+class PaddleOCRRMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        PaddleOCRRMSNorm is equivalent to T5LayerNorm
+        """
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+        return self.weight * hidden_states.to(input_dtype)
+
+    def extra_repr(self):
+        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
+
+
+class PaddleOCRDecoderLayer(GradientCheckpointingLayer):
+    def __init__(self, config: PaddleOCRTextConfig, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+
+        self.self_attn = PaddleOCRAttention(config=config, layer_idx=layer_idx)
+
+        self.mlp = PaddleOCRMLP(config)
+        self.input_layernorm = PaddleOCRRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = PaddleOCRRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        use_cache: Optional[bool] = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> torch.Tensor:
+        residual = hidden_states
+        hidden_states = self.input_layernorm(hidden_states)
+        # Self Attention
+        hidden_states, _ = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            position_embeddings=position_embeddings,
+            **kwargs,
+        )
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+        return hidden_states
+
+
+@auto_docstring
+class PaddleOCRVLPreTrainedModel(PreTrainedModel):
+    config: PaddleOCRVLConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["PaddleOCRDecoderLayer"]
+    _skip_keys_device_placement = ["past_key_values"]
+    _supports_flash_attn = True
+    _supports_sdpa = True
+    _supports_flex_attn = True
+
+    _can_compile_fullgraph = True
+    _supports_attention_backend = True
+
+    _can_record_outputs = {
+        "hidden_states": PaddleOCRDecoderLayer,
+        "attentions": PaddleOCRAttention,
+    }
+
+    def _init_weights(self, module):
+        super()._init_weights(module)
+        if isinstance(module, PaddleOCRVisionEmbeddings):
+            init.copy_(module.position_ids, torch.arange(module.position_ids.shape[-1]).expand((1, -1)))
+        elif isinstance(module, PaddleOCRVisionRotaryEmbedding):
+            inv_freq = 1.0 / (module.theta ** (torch.arange(0, module.dim, 2, dtype=torch.float) / module.dim))
+            init.copy_(module.inv_freq, inv_freq)
+
+
+@auto_docstring
+class PaddleOCRTextModel(PaddleOCRVLPreTrainedModel):
+    def __init__(self, config: PaddleOCRTextConfig):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
+        self.layers = nn.ModuleList(
+            [PaddleOCRDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+        )
+        self.norm = PaddleOCRRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.rotary_emb = PaddleOCRRotaryEmbedding(config=config)
+        self.gradient_checkpointing = False
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @check_model_inputs
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> BaseModelOutputWithPast:
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+        if inputs_embeds is None:
+            inputs_embeds: torch.Tensor = self.embed_tokens(input_ids)
+
+        if use_cache and past_key_values is None:
+            past_key_values = DynamicCache(config=self.config)
+
+        if cache_position is None:
+            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+            cache_position: torch.Tensor = (
+                torch.arange(inputs_embeds.shape[1], device=inputs_embeds.device) + past_seen_tokens
+            )
+
+        if position_ids is None:
+            position_ids = cache_position.view(1, 1, -1).expand(3, inputs_embeds.shape[0], -1)
+        elif position_ids.ndim == 2:
+            position_ids = position_ids[None, ...].expand(3, position_ids.shape[0], -1)
+
+        if position_ids.ndim == 3 and position_ids.shape[0] == 4:
+            text_position_ids = position_ids[0]
+            position_ids = position_ids[1:]
+        else:
+            text_position_ids = None
+
+        causal_mask = create_causal_mask(
+            config=self.config,
+            input_embeds=inputs_embeds,
+            attention_mask=attention_mask,
+            cache_position=cache_position,
+            past_key_values=past_key_values,
+            position_ids=text_position_ids,
+        )
+
+        hidden_states = inputs_embeds
+        position_embeddings = self.rotary_emb(hidden_states, position_ids=position_ids)
+
+        for decoder_layer in self.layers[: self.config.num_hidden_layers]:
+            hidden_states = decoder_layer(
+                hidden_states,
+                attention_mask=causal_mask,
+                position_embeddings=position_embeddings,
+                position_ids=text_position_ids,
+                past_key_values=past_key_values,
+                use_cache=use_cache,
+                cache_position=cache_position,
+                **kwargs,
+            )
+
+        hidden_states = self.norm(hidden_states)
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values,
+        )
+
+
+class PaddleOCRVisionModel(PaddleOCRVLPreTrainedModel):
+    config: PaddleOCRVisionConfig
+    main_input_name = "pixel_values"
+    input_modalities = "image"
+
+    def __init__(self, config: PaddleOCRVisionConfig):
+        super().__init__(config)
+
+        self.vision_model = PaddleOCRVisionTransformer(config)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def forward(
+        self,
+        pixel_values: torch.FloatTensor,
+        cu_seqlens: torch.Tensor,
+        image_grid_thw: Optional[list[Union[tuple[int, int, int], list[tuple[int, int, int]]]]] = None,
+        **kwargs,
+    ) -> BaseModelOutputWithPooling:
+        """
+        Args:
+            pixel_values (`torch.FloatTensor` of shape `(batch_size, sequence_length, image_channels, patch_size, patch_size)`):
+                The tensors corresponding to the input images.
+            cu_seqlens (`torch.Tensor` of shape `(num_images + 1,)`):
+                The cumulative sequence lengths of each image or video feature.
+            image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
+                The temporal, height and width of feature shape of each image in LLM.
+        """
+        return self.vision_model(
+            pixel_values=pixel_values,
+            cu_seqlens=cu_seqlens,
+            image_grid_thw=image_grid_thw,
+        )
+
+
+class PaddleOCRVisionEmbeddings(nn.Module):
+    def __init__(self, config: PaddleOCRVisionConfig):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.image_size = config.image_size
+        self.patch_size = config.patch_size
+
+        self.patch_embedding = nn.Conv2d(
+            in_channels=config.num_channels,
+            out_channels=self.embed_dim,
+            kernel_size=self.patch_size,
+            stride=self.patch_size,
+            padding="valid",
+        )
+
+        self.num_patches = (self.image_size // self.patch_size) ** 2
+        self.num_positions = self.num_patches
+        self.position_embedding = nn.Embedding(self.num_positions, self.embed_dim)
+        self.register_buffer("position_ids", torch.arange(self.num_positions).expand((1, -1)), persistent=False)
+
+    def interpolate_pos_encoding(self, embeddings: torch.Tensor, height: int, width: int) -> torch.Tensor:
+        """
+        This method allows to interpolate the pre-trained position encodings, to be able to use the model on higher resolution
+        images. This method is also adapted to support torch.jit tracing and no class embeddings.
+
+        Adapted from:
+        - https://github.com/facebookresearch/dino/blob/de9ee3df6cf39fac952ab558447af1fa1365362a/vision_transformer.py#L174-L194, and
+        - https://github.com/facebookresearch/dinov2/blob/e1277af2ba9496fbadf7aec6eba56e8d882d1e35/dinov2/models/vision_transformer.py#L179-L211
+        """
+        num_positions = self.position_embedding.weight.shape[0]
+
+        patch_pos_embed = self.position_embedding.weight.unsqueeze(0)
+
+        dim = embeddings.shape[-1]
+
+        sqrt_num_positions = torch_int(num_positions**0.5)
+        patch_pos_embed = patch_pos_embed.reshape(1, sqrt_num_positions, sqrt_num_positions, dim)
+        patch_pos_embed = patch_pos_embed.permute(0, 3, 1, 2)
+
+        patch_pos_embed = nn.functional.interpolate(
+            patch_pos_embed,
+            size=(height, width),
+            mode="bilinear",
+            align_corners=False,
+        )
+
+        patch_pos_embed = patch_pos_embed.permute(0, 2, 3, 1).view(1, -1, dim)
+        return patch_pos_embed
+
+    def forward(
+        self,
+        pixel_values: torch.FloatTensor,
+        image_grid_thw: Optional[list[Union[tuple[int, int, int], list[tuple[int, int, int]]]]] = None,
+    ) -> torch.Tensor:
+        """
+        Args:
+            pixel_values (`torch.FloatTensor` of shape `(batch_size, sequence_length, image_channels, patch_size, patch_size)`):
+                The tensors corresponding to the input images.
+            image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
+                The temporal, height and width of feature shape of each image in LLM.
+        """
+        batch_size, squence_len, channel, height, width = pixel_values.shape
+        target_dtype = self.patch_embedding.weight.dtype
+        pixel_values = pixel_values.reshape(batch_size * squence_len, channel, height, width)
+        patch_embeds = self.patch_embedding(pixel_values.to(dtype=target_dtype))  # shape = [*, width, grid, grid]
+        embeddings = patch_embeds.flatten(-2).squeeze(-1)
+        embeddings = embeddings.reshape(batch_size, squence_len, -1)
+
+        start = 0
+        embeddings = embeddings.squeeze(0)
+        tmp_embeddings = []
+        for image_grid in image_grid_thw:
+            t, h, w = image_grid
+            end = start + t * h * w
+            image_embeddings = embeddings[start:end, :]
+            position_embedding = self.interpolate_pos_encoding(image_embeddings, h, w).squeeze(0).repeat(t, 1)
+            image_embeddings = image_embeddings + position_embedding
+            tmp_embeddings.append(image_embeddings)
+            start = end
+        embeddings = torch.concat(tmp_embeddings, dim=0)
+
+        return embeddings
+
+
+def apply_rotary_pos_emb_vision(
+    q: torch.Tensor, k: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor
+) -> tuple[torch.Tensor, torch.Tensor]:
+    orig_q_dtype = q.dtype
+    orig_k_dtype = k.dtype
+    q, k = q.float(), k.float()
+    cos, sin = cos.unsqueeze(-2).float(), sin.unsqueeze(-2).float()
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    q_embed = q_embed.to(orig_q_dtype)
+    k_embed = k_embed.to(orig_k_dtype)
+    return q_embed, k_embed
+
+
+class PaddleOCRVisionAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config: PaddleOCRVisionConfig):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.embed_dim // self.num_heads
+        if self.head_dim * self.num_heads != self.embed_dim:
+            raise ValueError(
+                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:"
+                f" {self.num_heads})."
+            )
+        self.is_causal = False
+
+        self.k_proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.v_proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.q_proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.out_proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.num_key_value_groups = 1
+        self.scaling = self.head_dim**-0.5
+        self.attention_dropout = config.attention_dropout
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        cu_seqlens: torch.Tensor,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor],
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple[torch.Tensor, Optional[torch.Tensor]]:
+        """
+        Args:
+            hidden_states (`torch.Tensor`):
+                Input to the layer of shape `(seq_len, embed_dim)`.
+            cu_seqlens (`torch.Tensor` of shape `(num_images_or_videos + 1,)`):
+                The cumulative sequence lengths of each image or video feature.
+            position_embeddings (`tuple(torch.Tensor, torch.Tensor)` of shape `(num_patches, head_dim // 2)`):
+                The cosine and sine position embeddings for vision attention.
+        """
+        seq_length = hidden_states.shape[0]
+        query_states = self.q_proj(hidden_states).view(seq_length, self.num_heads, self.head_dim)
+        key_states = self.k_proj(hidden_states).view(seq_length, self.num_heads, self.head_dim)
+        value_states = self.v_proj(hidden_states).view(seq_length, self.num_heads, self.head_dim)
+
+        cos, sin = position_embeddings
+        query_states, key_states = apply_rotary_pos_emb_vision(query_states, key_states, cos, sin)
+
+        query_states = query_states.transpose(0, 1).unsqueeze(0)
+        key_states = key_states.transpose(0, 1).unsqueeze(0)
+        value_states = value_states.transpose(0, 1).unsqueeze(0)
+
+        attention_interface: Callable = eager_attention_forward
+        if self.config._attn_implementation != "eager":
+            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+
+        if self.config._attn_implementation == "flash_attention_2":
+            # Flash Attention 2: Use cu_seqlens for variable length attention
+            max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max()
+            attn_output, attn_weights = attention_interface(
+                self,
+                query_states,
+                key_states,
+                value_states,
+                attention_mask=None,
+                scaling=self.scaling,
+                dropout=0.0 if not self.training else self.attention_dropout,
+                cu_seq_lens_q=cu_seqlens,
+                cu_seq_lens_k=cu_seqlens,
+                max_length_q=max_seqlen,
+                max_length_k=max_seqlen,
+                is_causal=False,
+                **kwargs,
+            )
+        else:
+            # Other implementations: Process each chunk separately
+            lengths = cu_seqlens[1:] - cu_seqlens[:-1]
+            splits = [
+                torch.split(tensor, lengths.tolist(), dim=2) for tensor in (query_states, key_states, value_states)
+            ]
+
+            attn_outputs, attn_weights = [], []
+            for q, k, v in zip(*splits):
+                attn_output, attn_weight = attention_interface(
+                    self,
+                    q,
+                    k,
+                    v,
+                    attention_mask=None,
+                    scaling=self.scaling,
+                    dropout=0.0 if not self.training else self.attention_dropout,
+                    is_causal=False,
+                    **kwargs,
+                )
+                attn_outputs.append(attn_output)
+                attn_weights.append(attn_weight)
+
+            attn_output = torch.cat(attn_outputs, dim=1)
+
+        attn_output = attn_output.reshape(seq_length, -1).contiguous()
+        attn_output = self.out_proj(attn_output)
+
+        return attn_output, attn_weights
+
+
+class PaddleOCRVisionMLP(nn.Module):
+    def __init__(self, config: PaddleOCRVisionConfig):
+        super().__init__()
+        self.config = config
+        self.activation_fn = ACT2FN[config.hidden_act]
+        self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
+        self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.fc1(hidden_states)
+        hidden_states = self.activation_fn(hidden_states)
+        hidden_states = self.fc2(hidden_states)
+        return hidden_states
+
+
+class PaddleOCRVisionEncoderLayer(GradientCheckpointingLayer):
+    def __init__(self, config: PaddleOCRVisionConfig):
+        super().__init__()
+        self.embed_dim = config.hidden_size
+        self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+        self.self_attn = PaddleOCRVisionAttention(config=config)
+        self.layer_norm2 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+        self.mlp = PaddleOCRVisionMLP(config=config)
+
+    @auto_docstring
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        cu_seqlens: torch.Tensor,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor],
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> torch.Tensor:
+        r"""
+        cu_seqlens (`torch.Tensor` of shape `(num_images_or_videos + 1,)`):
+            The cumulative sequence lengths of each image or video feature.
+        position_embeddings (`tuple(torch.Tensor, torch.Tensor)` of shape `(num_patches, head_dim // 2)`):
+            The cosine and sine position embeddings for vision attention.
+        """
+        residual = hidden_states
+
+        hidden_states = self.layer_norm1(hidden_states)
+        hidden_states, _ = self.self_attn(
+            hidden_states,
+            cu_seqlens=cu_seqlens,
+            position_embeddings=position_embeddings,
+            **kwargs,
+        )
+        hidden_states = residual + hidden_states
+
+        residual = hidden_states
+        hidden_states = self.layer_norm2(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        return hidden_states
+
+
+class PaddleOCRVisionEncoder(nn.Module):
+    """
+    Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a
+    [`PaddleOCRVisionEncoderLayer`].
+
+    Args:
+        config: PaddleOCRVisionConfig
+    """
+
+    def __init__(self, config: PaddleOCRVisionConfig):
+        super().__init__()
+        self.config = config
+        self.layers = nn.ModuleList([PaddleOCRVisionEncoderLayer(config) for _ in range(config.num_hidden_layers)])
+        self.gradient_checkpointing = False
+        embed_dim = config.hidden_size
+        num_heads = config.num_attention_heads
+        head_dim = embed_dim // num_heads
+        self.rotary_pos_emb = PaddleOCRVisionRotaryEmbedding(head_dim // 2)
+
+    # Ignore copy
+    @can_return_tuple
+    @auto_docstring
+    def forward(
+        self,
+        inputs_embeds: torch.FloatTensor,
+        cu_seqlens: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        image_grid_thw: Optional[list[Union[tuple[int, int, int], list[tuple[int, int, int]]]]] = None,
+    ) -> BaseModelOutput:
+        r"""
+        inputs_embeds (`torch.FloatTensor` of shape `(sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+            than the model's internal embedding lookup matrix.
+        cu_seqlens (`torch.Tensor` of shape `(num_images + 1,)`):
+            The cumulative sequence lengths of each image or video feature.
+        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            The attention_mask used in forward function shape [batch_size X sequence_length] if not None.
+        image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
+            The temporal, height and width of feature shape of each image in LLM.
+        """
+        device = inputs_embeds.device
+        hidden_states = inputs_embeds
+        attention_mask = create_bidirectional_mask(
+            config=self.config,
+            input_embeds=inputs_embeds,
+            attention_mask=attention_mask,
+        )
+        split_hids = []
+        split_wids = []
+        for t, h, w in image_grid_thw:
+            image_pids = torch.arange(t * h * w, device=device) % (h * w)
+            sample_hids = image_pids // w
+            sample_wids = image_pids % w
+            split_hids.append(sample_hids)
+            split_wids.append(sample_wids)
+        width_position_ids = torch.concat(split_wids, dim=0)
+        height_position_ids = torch.concat(split_hids, dim=0)
+
+        pids = torch.stack([height_position_ids, width_position_ids], dim=-1)
+        max_grid_size = pids.max() + 1
+        rotary_embeddings_max_grid = self.rotary_pos_emb(max_grid_size)
+        rotary_embeddings = rotary_embeddings_max_grid[pids].flatten(1)
+        rotary_embeddings = rotary_embeddings.repeat(1, 2)
+        position_embeddings = (rotary_embeddings.cos(), rotary_embeddings.sin())
+
+        for encoder_layer in self.layers:
+            hidden_states = encoder_layer(
+                hidden_states,
+                cu_seqlens=cu_seqlens,
+                position_embeddings=position_embeddings,
+            )
+
+        return BaseModelOutput(
+            last_hidden_state=hidden_states,
+        )
+
+
+class PaddleOCRVisionTransformer(PaddleOCRVLPreTrainedModel):
+    def __init__(self, config: PaddleOCRVisionConfig):
+        super().__init__(config)
+        self.config = config
+        embed_dim = config.hidden_size
+
+        self.embeddings = PaddleOCRVisionEmbeddings(config)
+        self.encoder = PaddleOCRVisionEncoder(config)
+        self.post_layernorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
+
+        self.post_init()
+
+    def forward(
+        self,
+        pixel_values: torch.FloatTensor,
+        cu_seqlens: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        image_grid_thw: Optional[list[Union[tuple[int, int, int], list[tuple[int, int, int]]]]] = None,
+        **kwargs,
+    ) -> BaseModelOutputWithPooling:
+        """
+        Args:
+            pixel_values (`torch.FloatTensor` of shape `(batch_size, sequence_length, patch_size * patch_size * image_channels)`):
+                The tensors corresponding to the input images.
+            cu_seqlens (`torch.Tensor` of shape `(num_images + 1,)`):
+                The cumulative sequence lengths of each image or video feature.
+            attention_mask (`torch.Tensor`, *optional*):
+                The attention_mask used in forward function shape [batch_size X sequence_length] if not None.
+            image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
+                The temporal, height and width of feature shape of each image in LLM.
+        """
+        hidden_states = self.embeddings(pixel_values, image_grid_thw=image_grid_thw)
+
+        encoder_outputs: BaseModelOutput = self.encoder(
+            inputs_embeds=hidden_states,
+            cu_seqlens=cu_seqlens,
+            attention_mask=attention_mask,
+            image_grid_thw=image_grid_thw,
+        )
+
+        last_hidden_state = encoder_outputs.last_hidden_state
+        last_hidden_state = self.post_layernorm(last_hidden_state)
+
+        return BaseModelOutputWithPooling(
+            last_hidden_state=last_hidden_state,
+            pooler_output=None,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )
+
+
+@dataclass
+@auto_docstring(
+    custom_intro="""
+    Base class for Llava outputs, with hidden states and attentions.
+    """
+)
+class PaddleOCRVLModelOutputWithPast(ModelOutput):
+    r"""
+    past_key_values (`Cache`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+        It is a [`~cache_utils.Cache`] instance. For more details, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache).
+
+        Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
+        `past_key_values` input) to speed up sequential decoding.
+    rope_deltas (`torch.LongTensor` of shape `(batch_size, )`, *optional*):
+        The rope index difference between sequence length and multimodal rope.
+    """
+
+    last_hidden_state: Optional[torch.FloatTensor] = None
+    past_key_values: Optional[Cache] = None
+    hidden_states: Optional[tuple[torch.FloatTensor]] = None
+    attentions: Optional[tuple[torch.FloatTensor]] = None
+    rope_deltas: Optional[torch.LongTensor] = None
+
+
+@dataclass
+@auto_docstring(
+    custom_intro="""
+    Base class for PaddleOCRVL causal language model (or autoregressive) outputs.
+    """
+)
+class PaddleOCRVLCausalLMOutputWithPast(ModelOutput):
+    r"""
+    loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
+        Language modeling loss (for next-token prediction).
+    logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
+        Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+    past_key_values (`Cache`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+        It is a [`~cache_utils.Cache`] instance. For more details, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache).
+
+        Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
+        `past_key_values` input) to speed up sequential decoding.
+    rope_deltas (`torch.LongTensor` of shape `(batch_size, )`, *optional*):
+        The rope index difference between sequence length and multimodal rope.
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    logits: Optional[torch.FloatTensor] = None
+    past_key_values: Optional[Cache] = None
+    hidden_states: Optional[tuple[torch.FloatTensor]] = None
+    attentions: Optional[tuple[torch.FloatTensor]] = None
+    rope_deltas: Optional[torch.LongTensor] = None
+
+
+@auto_docstring
+class PaddleOCRVLModel(PaddleOCRVLPreTrainedModel):
+    base_model_prefix = "model"
+    _checkpoint_conversion_mapping = {"^model": "language_model"}
+    # Reference: fix gemma3 grad acc #37208
+    accepts_loss_kwargs = False
+    _keys_to_ignore_on_load_unexpected = ["packing_position_embedding", "vision_model.head"]
+
+    def __init__(self, config: PaddleOCRVLConfig):
+        super().__init__(config)
+        self.visual = PaddleOCRVisionModel._from_config(config.vision_config)
+        self.language_model = PaddleOCRTextModel._from_config(config.text_config)
+        self.rope_deltas = None
+        self.projector = PaddleOCRProjector(config)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.language_model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.language_model.embed_tokens = value
+
+    def get_rope_index(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        image_grid_thw: Optional[torch.LongTensor] = None,
+        video_grid_thw: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        """
+        Calculate the 3D rope index based on image and video's temporal, height and width in LLM.
+
+        Explanation:
+            Each embedding sequence contains vision embedding and text embedding or just contains text embedding.
+
+            For pure text embedding sequence, the rotary position embedding has no difference with modern LLMs.
+            Examples:
+                input_ids: [T T T T T], here T is for text.
+                temporal position_ids: [0, 1, 2, 3, 4]
+                height position_ids: [0, 1, 2, 3, 4]
+                width position_ids: [0, 1, 2, 3, 4]
+
+            For vision and text embedding sequence, we calculate 3D rotary position embedding for vision part
+            and 1D rotary position embedding for text part.
+            Examples:
+                Assume we have a video input with 3 temporal patches, 2 height patches and 2 width patches.
+                input_ids: [V V V V V V V V V V V V T T T T T], here V is for vision.
+                vision temporal position_ids: [0, 0, 0, 0, 1, 1, 1, 1, 2, 2, 2, 2]
+                vision height position_ids: [0, 0, 1, 1, 0, 0, 1, 1, 0, 0, 1, 1]
+                vision width position_ids: [0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1]
+                text temporal position_ids: [3, 4, 5, 6, 7]
+                text height position_ids: [3, 4, 5, 6, 7]
+                text width position_ids: [3, 4, 5, 6, 7]
+                Here we calculate the text start position_ids as the max vision position_ids plus 1.
+
+        Args:
+            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+                Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+                it.
+            image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
+                The temporal, height and width of feature shape of each image in LLM.
+            video_grid_thw (`torch.LongTensor` of shape `(num_videos, 3)`, *optional*):
+                The temporal, height and width of feature shape of each video in LLM.
+            attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+                - 1 for tokens that are **not masked**,
+                - 0 for tokens that are **masked**.
+
+        Returns:
+            position_ids (`torch.LongTensor` of shape `(3, batch_size, sequence_length)`)
+            mrope_position_deltas (`torch.Tensor` of shape `(batch_size)`)
+        """
+        spatial_merge_size = self.config.vision_config.spatial_merge_size
+        image_token_id = self.config.image_token_id
+        video_token_id = self.config.video_token_id
+        vision_start_token_id = self.config.vision_start_token_id
+        mrope_position_deltas = []
+        if input_ids is not None and (image_grid_thw is not None or video_grid_thw is not None):
+            total_input_ids = input_ids
+            if attention_mask is None:
+                attention_mask = torch.ones_like(total_input_ids)
+            position_ids = torch.ones(
+                3, input_ids.shape[0], input_ids.shape[1], dtype=input_ids.dtype, device=input_ids.device
+            )
+            image_index, video_index = 0, 0
+            for i, input_ids in enumerate(total_input_ids):
+                input_ids = input_ids[attention_mask[i].to(input_ids.device) == 1]
+                image_nums, video_nums = 0, 0
+                vision_start_indices = torch.argwhere(input_ids == vision_start_token_id).squeeze(1)
+                vision_tokens = input_ids[vision_start_indices + 1]
+                image_nums = (vision_tokens == image_token_id).sum()
+                video_nums = (vision_tokens == video_token_id).sum()
+                input_tokens = input_ids.tolist()
+                llm_pos_ids_list: list = []
+                st = 0
+                remain_images, remain_videos = image_nums, video_nums
+                for _ in range(image_nums + video_nums):
+                    if image_token_id in input_tokens and remain_images > 0:
+                        ed_image = input_tokens.index(image_token_id, st)
+                    else:
+                        ed_image = len(input_tokens) + 1
+                    if video_token_id in input_tokens and remain_videos > 0:
+                        ed_video = input_tokens.index(video_token_id, st)
+                    else:
+                        ed_video = len(input_tokens) + 1
+                    if ed_image < ed_video:
+                        t, h, w = (
+                            image_grid_thw[image_index][0],
+                            image_grid_thw[image_index][1],
+                            image_grid_thw[image_index][2],
+                        )
+                        image_index += 1
+                        remain_images -= 1
+                        ed = ed_image
+                    else:
+                        t, h, w = (
+                            video_grid_thw[video_index][0],
+                            video_grid_thw[video_index][1],
+                            video_grid_thw[video_index][2],
+                        )
+                        video_index += 1
+                        remain_videos -= 1
+                        ed = ed_video
+                    llm_grid_t, llm_grid_h, llm_grid_w = (
+                        t.item(),
+                        h.item() // spatial_merge_size,
+                        w.item() // spatial_merge_size,
+                    )
+                    text_len = ed - st
+
+                    st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0
+                    llm_pos_ids_list.append(torch.arange(text_len).view(1, -1).expand(3, -1) + st_idx)
+
+                    t_index = torch.arange(llm_grid_t).view(-1, 1).expand(-1, llm_grid_h * llm_grid_w).flatten()
+                    h_index = torch.arange(llm_grid_h).view(1, -1, 1).expand(llm_grid_t, -1, llm_grid_w).flatten()
+                    w_index = torch.arange(llm_grid_w).view(1, 1, -1).expand(llm_grid_t, llm_grid_h, -1).flatten()
+                    llm_pos_ids_list.append(torch.stack([t_index, h_index, w_index]) + text_len + st_idx)
+                    st = ed + llm_grid_t * llm_grid_h * llm_grid_w
+
+                if st < len(input_tokens):
+                    st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0
+                    text_len = len(input_tokens) - st
+                    llm_pos_ids_list.append(torch.arange(text_len).view(1, -1).expand(3, -1) + st_idx)
+
+                llm_positions = torch.cat(llm_pos_ids_list, dim=1).reshape(3, -1)
+                position_ids[..., i, attention_mask[i] == 1] = llm_positions.to(position_ids.device)
+                mrope_position_deltas.append(llm_positions.max() + 1 - len(total_input_ids[i]))
+            mrope_position_deltas = torch.tensor(mrope_position_deltas, device=input_ids.device).unsqueeze(1)
+            return position_ids, mrope_position_deltas
+        else:
+            if attention_mask is not None:
+                position_ids = attention_mask.long().cumsum(-1) - 1
+                position_ids.masked_fill_(attention_mask == 0, 1)
+                position_ids = position_ids.unsqueeze(0).expand(3, -1, -1).to(attention_mask.device)
+                max_position_ids = position_ids.max(0, keepdim=False)[0].max(-1, keepdim=True)[0]
+                mrope_position_deltas = max_position_ids + 1 - attention_mask.shape[-1]
+            else:
+                position_ids = (
+                    torch.arange(input_ids.shape[1], device=input_ids.device)
+                    .view(1, 1, -1)
+                    .expand(3, input_ids.shape[0], -1)
+                )
+                mrope_position_deltas = torch.zeros(
+                    [input_ids.shape[0], 1],
+                    device=input_ids.device,
+                    dtype=input_ids.dtype,
+                )
+
+            return position_ids, mrope_position_deltas
+
+    def get_video_features(
+        self, pixel_values_videos: torch.FloatTensor, video_grid_thw: Optional[torch.LongTensor] = None
+    ):
+        """
+        Encodes videos into continuous embeddings that can be forwarded to the language model.
+
+        Args:
+            pixel_values_videos (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)`):
+                The tensors corresponding to the input videos.
+            video_grid_thw (`torch.LongTensor` of shape `(num_videos, 3)`, *optional*):
+                The temporal, height and width of feature shape of each video in LLM.
+        """
+        pixel_values_videos = pixel_values_videos.type(self.visual.dtype)
+        video_embeds = self.visual(pixel_values_videos, grid_thw=video_grid_thw)
+        split_sizes = (video_grid_thw.prod(-1) // self.visual.spatial_merge_size**2).tolist()
+        video_embeds = torch.split(video_embeds, split_sizes)
+        return video_embeds
+
+    def get_image_features(self, pixel_values: torch.FloatTensor, image_grid_thw: Optional[torch.LongTensor] = None):
+        """
+        Encodes images into continuous embeddings that can be forwarded to the language model.
+
+        Args:
+            pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)`):
+                The tensors corresponding to the input images.
+            image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
+                The temporal, height and width of feature shape of each image in LLM.
+        """
+        pixel_values = pixel_values.type(self.visual.dtype).unsqueeze(0)
+        cu_seqlens = torch.repeat_interleave(image_grid_thw[:, 1] * image_grid_thw[:, 2], image_grid_thw[:, 0]).cumsum(
+            dim=0,
+            # Select dtype based on the following factors:
+            #  - FA2 requires that cu_seqlens_q must have dtype int32
+            #  - torch.onnx.export requires that cu_seqlens_q must have same dtype as grid_thw
+            # See https://github.com/huggingface/transformers/pull/34852 for more information
+            dtype=image_grid_thw.dtype if torch.jit.is_tracing() else torch.int32,
+        )
+        cu_seqlens = torch.nn.functional.pad(cu_seqlens, (1, 0), value=0)
+        vision_outputs = self.visual(
+            pixel_values=pixel_values,
+            image_grid_thw=image_grid_thw,
+            cu_seqlens=cu_seqlens,
+        )
+        image_embeds = vision_outputs.last_hidden_state
+        image_embeds = self.projector(image_embeds, image_grid_thw)
+        return image_embeds
+
+    def get_placeholder_mask(
+        self, input_ids: torch.LongTensor, inputs_embeds: torch.FloatTensor, image_features: torch.FloatTensor
+    ):
+        """
+        Obtains multimodal placeholder mask from `input_ids` or `inputs_embeds`, and checks that the placeholder token count is
+        equal to the length of multimodal features. If the lengths are different, an error is raised.
+        """
+        if input_ids is None:
+            special_image_mask = inputs_embeds == self.get_input_embeddings()(
+                torch.tensor(self.config.image_token_id, dtype=torch.long, device=inputs_embeds.device)
+            )
+            special_image_mask = special_image_mask.all(-1)
+        else:
+            special_image_mask = input_ids == self.config.image_token_id
+
+        n_image_tokens = special_image_mask.sum()
+        special_image_mask = special_image_mask.unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device)
+        n_image_features = image_features.shape[0] * image_features.shape[1]
+        if inputs_embeds[special_image_mask].numel() != image_features.numel():
+            raise ValueError(
+                f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}"
+            )
+        return special_image_mask
+
+    @can_return_tuple
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[list[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        pixel_values: Optional[torch.Tensor] = None,
+        image_grid_thw: Optional[torch.LongTensor] = None,
+        rope_deltas: Optional[torch.LongTensor] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs,
+    ) -> Union[tuple, PaddleOCRVLModelOutputWithPast]:
+        r"""
+        image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
+            The temporal, height and width of feature shape of each image in LLM.
+        rope_deltas (`torch.LongTensor` of shape `(batch_size, )`, *optional*):
+            The rope index difference between sequence length and multimodal rope.
+        """
+        if inputs_embeds is None:
+            inputs_embeds = self.language_model.embed_tokens(input_ids)
+
+        if pixel_values is not None:
+            image_embeds = self.get_image_features(pixel_values, image_grid_thw).to(
+                inputs_embeds.device, inputs_embeds.dtype
+            )
+            image_mask = self.get_placeholder_mask(input_ids, inputs_embeds=inputs_embeds, image_features=image_embeds)
+            inputs_embeds = inputs_embeds.masked_scatter(image_mask, image_embeds)
+
+        if position_ids is None:
+            past_key_values_length = 0 if past_key_values is None else past_key_values.get_seq_length()
+            if self.rope_deltas is None or past_key_values_length == 0:
+                position_ids, rope_deltas = self.get_rope_index(
+                    input_ids=input_ids,
+                    image_grid_thw=image_grid_thw,
+                    attention_mask=attention_mask,
+                )
+                self.rope_deltas = rope_deltas
+            # then use the prev pre-calculated rope-deltas to get the correct position ids
+            else:
+                batch_size, seq_length, _ = inputs_embeds.shape
+                position_ids = torch.arange(seq_length, device=inputs_embeds.device)
+                position_ids = position_ids.view(1, 1, -1).expand(3, batch_size, -1)
+                delta = (past_key_values_length + self.rope_deltas).to(inputs_embeds.device)
+                delta = delta.repeat_interleave(batch_size // delta.shape[0], dim=0)
+                position_ids = position_ids + delta.to(position_ids.device)
+
+        outputs = self.language_model(
+            input_ids=None,
+            position_ids=position_ids,
+            attention_mask=attention_mask,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            **kwargs,
+        )
+
+        output = PaddleOCRVLModelOutputWithPast(
+            last_hidden_state=outputs.last_hidden_state,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            rope_deltas=self.rope_deltas,
+        )
+
+        return output
+
+
+class PaddleOCRVLForConditionalGeneration(PaddleOCRVLPreTrainedModel, GenerationMixin):
+    _checkpoint_conversion_mapping = {
+        "^visual": "model.visual",
+        "^mlp_AR": "model.projector",
+        r"^model(?!(\.visual|\.projector|\.language_model))": "model.language_model",
+    }
+    _tied_weights_keys = {"lm_head.weight": "model.language_model.embed_tokens.weight"}
+    _keys_to_ignore_on_load_unexpected = ["packing_position_embedding", "vision_model.head"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = PaddleOCRVLModel(config)
+        self.lm_head = nn.Linear(config.text_config.hidden_size, config.text_config.vocab_size, bias=False)
+
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.get_input_embeddings()
+
+    def set_input_embeddings(self, value):
+        self.model.set_input_embeddings(value)
+
+    def get_video_features(
+        self, pixel_values_videos: torch.FloatTensor, video_grid_thw: Optional[torch.LongTensor] = None
+    ):
+        return self.model.get_video_features(pixel_values_videos, video_grid_thw)
+
+    def get_image_features(self, pixel_values: torch.FloatTensor, image_grid_thw: Optional[torch.LongTensor] = None):
+        return self.model.get_image_features(pixel_values, image_grid_thw)
+
+    @can_return_tuple
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        pixel_values: Optional[torch.Tensor] = None,
+        image_grid_thw: Optional[torch.LongTensor] = None,
+        rope_deltas: Optional[torch.LongTensor] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        logits_to_keep: Union[int, torch.Tensor] = 0,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> Union[tuple, PaddleOCRVLCausalLMOutputWithPast]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+        image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
+            The temporal, height and width of feature shape of each image in LLM.
+        rope_deltas (`torch.LongTensor` of shape `(batch_size, )`, *optional*):
+            The rope index difference between sequence length and multimodal rope.
+
+        Example:
+
+        ```python
+        >>> from transformers import AutoProcessor, PaddleOCRVLForConditionalGeneration
+
+        >>> model = PaddleOCRVLForConditionalGeneration.from_pretrained("PaddlePaddle/PaddleOCR-VL", dtype="bfloat16")
+        >>> processor = AutoProcessor.from_pretrained("PaddlePaddle/PaddleOCR-VL")
+
+        >>> messages = [
+            {
+                "role": "user",
+                "content": [
+                    {
+                        "type": "image",
+                        "image": "https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/ocr_demo.jpg",
+                    },
+                    {"type": "text", "text": "OCR:"},
+                ],
+            }
+        ]
+
+        >>> inputs = processor.apply_chat_template(
+            messages,
+            tokenize=True,
+            add_generation_prompt=True,
+            return_dict=True,
+            return_tensors="pt"
+        ).to(model.device)
+
+        >>> # Generate
+        >>> generated_ids = model.generate(**inputs, max_new_tokens=1024)
+        >>> generated_ids_trimmed = [out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)]
+        >>> output_text = processor.batch_decode(generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+        >>> print(output_text)
+        ```
+        """
+        outputs: PaddleOCRVLModelOutputWithPast = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            image_grid_thw=image_grid_thw,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            pixel_values=pixel_values,
+            rope_deltas=rope_deltas,
+            cache_position=cache_position,
+            **kwargs,
+        )
+        hidden_states = outputs.last_hidden_state
+
+        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
+        logits = self.lm_head(hidden_states[:, slice_indices, :])
+
+        loss = None
+        if labels is not None:
+            loss = self.loss_function(
+                logits=logits, labels=labels, vocab_size=self.config.text_config.vocab_size, **kwargs
+            )
+
+        return PaddleOCRVLCausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            rope_deltas=outputs.rope_deltas,
+        )
+
+    def prepare_inputs_for_generation(
+        self,
+        input_ids,
+        past_key_values=None,
+        attention_mask=None,
+        inputs_embeds=None,
+        cache_position=None,
+        position_ids=None,
+        use_cache=True,
+        pixel_values=None,
+        pixel_values_videos=None,
+        image_grid_thw=None,
+        video_grid_thw=None,
+        is_first_iteration=False,
+        **kwargs,
+    ):
+        # Overwritten -- in specific circumstances we don't want to forward image inputs to the model
+
+        model_inputs = super().prepare_inputs_for_generation(
+            input_ids,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+            inputs_embeds=inputs_embeds,
+            cache_position=cache_position,
+            position_ids=position_ids,
+            pixel_values=pixel_values,
+            pixel_values_videos=pixel_values_videos,
+            image_grid_thw=image_grid_thw,
+            video_grid_thw=video_grid_thw,
+            use_cache=use_cache,
+            is_first_iteration=is_first_iteration,
+            **kwargs,
+        )
+
+        # Qwen2-VL position_ids are prepareed with rope_deltas in forward
+        if position_ids is None:
+            # Calculate RoPE index once per generation in the pre-fill stage only.
+            # When compiling, we can't check tensor values thus we check only input length
+            # It is safe to assume that `length!=1` means we're in pre-fill because compiled
+            # models currently cannot do asssisted decoding
+            if model_inputs["cache_position"][0] == 0 or self.model.rope_deltas is None:
+                vision_positions, rope_deltas = self.model.get_rope_index(
+                    model_inputs.get("input_ids", None),
+                    image_grid_thw=image_grid_thw,
+                    video_grid_thw=video_grid_thw,
+                    attention_mask=attention_mask,
+                )
+                self.model.rope_deltas = rope_deltas
+            # then use the prev pre-calculated rope-deltas to get the correct position ids
+            elif "position_ids" in model_inputs:
+                batch_size, seq_length = model_inputs["position_ids"].shape
+                device = model_inputs["position_ids"].device
+                position_ids = torch.arange(seq_length, device=device)
+                position_ids = position_ids.view(1, 1, -1).expand(3, batch_size, -1)
+                delta = cache_position[0] + self.model.rope_deltas
+                delta = delta.repeat_interleave(batch_size // delta.shape[0], dim=0)
+                vision_positions = position_ids + delta.expand_as(position_ids)
+
+            # Concatenate "text + vision" positions into [4, bs, seq-len]
+            text_positions = model_inputs["position_ids"][None, ...]
+            model_inputs["position_ids"] = torch.cat([text_positions, vision_positions], dim=0)
+
+        if not is_first_iteration and use_cache:
+            model_inputs["pixel_values"] = None
+            model_inputs["pixel_values_videos"] = None
+
+        return model_inputs
+
+    def _get_image_nums_and_video_nums(
+        self,
+        input_ids: Optional[torch.LongTensor],
+        inputs_embeds: Optional[torch.Tensor] = None,
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        """
+        Get the number of images and videos for each sample to calculate the separation length of the sample tensor.
+        These parameters are not passed through the processor to avoid unpredictable impacts from interface modifications.
+
+        Args:
+            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+                Indices of input sequence tokens in the vocabulary.
+
+        Returns:
+            image_nums (`torch.LongTensor` of shape `(batch_size, num_images_sample)`)
+            video_nums (`torch.LongTensor` of shape `(batch_size, num_videos_sample)`)
+        """
+        image_token_id = self.config.image_token_id
+        video_token_id = self.config.video_token_id
+        vision_start_token_id = self.config.vision_start_token_id
+
+        if inputs_embeds is not None:
+            vision_start_mask = (
+                inputs_embeds
+                == self.get_input_embeddings()(
+                    torch.tensor(vision_start_token_id, dtype=torch.long, device=inputs_embeds.device)
+                )
+            )[..., 0]
+            image_mask = (
+                inputs_embeds
+                == self.get_input_embeddings()(
+                    torch.tensor(image_token_id, dtype=torch.long, device=inputs_embeds.device)
+                )
+            )[..., 0]
+            video_mask = (
+                inputs_embeds
+                == self.get_input_embeddings()(
+                    torch.tensor(video_token_id, dtype=torch.long, device=inputs_embeds.device)
+                )
+            )[..., 0]
+        else:
+            vision_start_mask = input_ids == vision_start_token_id
+            image_mask = input_ids == image_token_id
+            video_mask = input_ids == video_token_id
+
+        vision_first_mask = torch.roll(vision_start_mask, shifts=1, dims=1)
+        image_nums = torch.sum(vision_first_mask & image_mask, dim=1)
+        video_nums = torch.sum(vision_first_mask & video_mask, dim=1)
+
+        return image_nums, video_nums
+
+    def _expand_inputs_for_generation(
+        self,
+        expand_size: int = 1,
+        is_encoder_decoder: bool = False,
+        input_ids: Optional[torch.LongTensor] = None,
+        **model_kwargs,
+    ) -> tuple[torch.LongTensor, dict[str, Any]]:
+        # Overwritten -- Support for expanding tensors without a batch size dimension
+        # e.g., pixel_values, image_grid_thw, pixel_values_videos, video_grid_thw, second_per_grid_t
+        # pixel_values.shape[0] is sum(seqlen_images for samples)
+        # image_grid_thw.shape[0] is sum(num_images for samples)
+
+        if expand_size == 1:
+            return input_ids, model_kwargs
+
+        visual_keys = ["pixel_values", "image_grid_thw", "pixel_values_videos", "video_grid_thw", "second_per_grid_ts"]
+
+        def _expand_dict_for_generation_visual(dict_to_expand):
+            image_grid_thw = model_kwargs.get("image_grid_thw", None)
+            video_grid_thw = model_kwargs.get("video_grid_thw", None)
+            image_nums, video_nums = self._get_image_nums_and_video_nums(
+                input_ids, inputs_embeds=model_kwargs.get("inputs_embeds", None)
+            )
+
+            def _repeat_interleave_samples(x, lengths, repeat_times):
+                samples = torch.split(x, lengths)
+                repeat_args = [repeat_times] + [1] * (x.dim() - 1)
+                result = torch.cat([sample.repeat(*repeat_args) for sample in samples], dim=0)
+                return result
+
+            for key in dict_to_expand:
+                if key == "pixel_values":
+                    # split images into samples
+                    samples = torch.split(image_grid_thw, list(image_nums))
+                    # compute the sequence length of images for each sample
+                    lengths = [torch.prod(sample, dim=1).sum() for sample in samples]
+                    dict_to_expand[key] = _repeat_interleave_samples(
+                        dict_to_expand[key], lengths=lengths, repeat_times=expand_size
+                    )
+                elif key == "image_grid_thw":
+                    # get the num of images for each sample
+                    lengths = list(image_nums)
+                    dict_to_expand[key] = _repeat_interleave_samples(
+                        dict_to_expand[key], lengths=lengths, repeat_times=expand_size
+                    )
+                elif key == "pixel_values_videos":
+                    samples = torch.split(video_grid_thw, list(video_nums))
+                    lengths = [torch.prod(sample, dim=1).sum() for sample in samples]
+                    dict_to_expand[key] = _repeat_interleave_samples(
+                        dict_to_expand[key], lengths=lengths, repeat_times=expand_size
+                    )
+                elif key == "video_grid_thw":
+                    lengths = list(video_nums)
+                    dict_to_expand[key] = _repeat_interleave_samples(
+                        dict_to_expand[key], lengths=lengths, repeat_times=expand_size
+                    )
+                elif key == "second_per_grid_ts":
+                    dict_to_expand[key] = _repeat_interleave_samples(
+                        dict_to_expand[key], lengths=list(video_nums), repeat_times=expand_size
+                    )
+            return dict_to_expand
+
+        def _expand_dict_for_generation(dict_to_expand):
+            for key in dict_to_expand:
+                if (
+                    key != "cache_position"
+                    and dict_to_expand[key] is not None
+                    and isinstance(dict_to_expand[key], torch.Tensor)
+                    and key not in visual_keys
+                ):
+                    dict_to_expand[key] = dict_to_expand[key].repeat_interleave(expand_size, dim=0)
+            return dict_to_expand
+
+        model_kwargs = _expand_dict_for_generation_visual(model_kwargs)
+
+        if input_ids is not None:
+            input_ids = input_ids.repeat_interleave(expand_size, dim=0)
+
+        model_kwargs = _expand_dict_for_generation(model_kwargs)
+
+        if is_encoder_decoder:
+            if model_kwargs.get("encoder_outputs") is None:
+                raise ValueError("If `is_encoder_decoder` is True, make sure that `encoder_outputs` is defined.")
+            model_kwargs["encoder_outputs"] = _expand_dict_for_generation(model_kwargs["encoder_outputs"])
+
+        return input_ids, model_kwargs
+
+
+__all__ = [
+    "PaddleOCRVLForConditionalGeneration",
+    "PaddleOCRVLModel",
+    "PaddleOCRVLPreTrainedModel",
+    "PaddleOCRVisionTransformer",
+    "PaddleOCRTextModel",
+    "PaddleOCRVisionModel",
+]