modeling_openpangu_vl.py

# coding=utf-8
# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
# Copyright 2022 EleutherAI 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 dataclasses import dataclass
from typing import Any, Callable, Optional, Union

import torch
import torch.nn as nn
import torch.nn.functional as F
import torch_npu
from einops import rearrange

from transformers.cache_utils import Cache
from transformers.generation import GenerationMixin
from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
from transformers.modeling_layers import GradientCheckpointingLayer
from transformers.modeling_outputs import ModelOutput
from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
from transformers.processing_utils import Unpack
from transformers.utils import LossKwargs, auto_docstring, can_return_tuple, is_torchdynamo_compiling, logging

from .configuration_openpangu_vl import OpenPanguVLConfig as OpenPanguConfig
from .configuration_openpangu_vl import OpenPanguVLTextConfig, OpenPanguVLVisionConfig
from .modeling_openpangu_embedded import PanguEmbeddedConfig, PanguEmbeddedMLP, PanguEmbeddedModel, PanguEmbeddedRMSNorm
from .imageprocessor_openpangu_vl import rescale_and_normalize
if "910" in torch.npu.get_device_name():
    NPU_ATTN_INFR = True
    print("[INFO] torch_npu detected. Using NPU fused infer attention.")
else:
    NPU_ATTN_INFR = False


logger = logging.get_logger(__name__)


class OpenPanguVLMLP(PanguEmbeddedMLP):
    pass


class OpenPanguVisionPatchEmbed(nn.Module):
    def __init__(
        self,
        patch_size: int = 14,
        temporal_patch_size: int = 2,
        in_channels: int = 3,
        embed_dim: int = 1152,
    ) -> None:
        super().__init__()
        self.patch_size = patch_size
        self.temporal_patch_size = temporal_patch_size
        self.in_channels = in_channels
        self.embed_dim = embed_dim

        kernel_size = [temporal_patch_size, patch_size, patch_size]
        self.proj = nn.Conv3d(in_channels, embed_dim, kernel_size=kernel_size, stride=kernel_size, bias=False)
        self.input_size = self.patch_size * self.patch_size * in_channels * self.temporal_patch_size

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        if hidden_states.shape[-1] != self.input_size:
            hidden_states = torch.cat([hidden_states.reshape(-1, self.patch_size * self.patch_size), \
            hidden_states.reshape(-1, self.patch_size * self.patch_size)], dim=-1).reshape(-1, self.input_size)
        target_dtype = self.proj.weight.dtype
        hidden_states = hidden_states.view(
            -1,
            self.in_channels,
            self.temporal_patch_size,
            self.patch_size,
            self.patch_size,
        )
        hidden_states = self.proj(hidden_states.to(dtype=target_dtype)).view(-1, self.embed_dim)
        return hidden_states


class OpenPanguVLPatchEmbed(OpenPanguVisionPatchEmbed):
    pass


class OpenPanguVisionRotaryEmbedding(nn.Module):
    def __init__(self, dim: int, theta: float = 10000.0) -> None:
        super().__init__()
        inv_freq = 1.0 / (theta ** (torch.arange(0, dim, 2, dtype=torch.float) / dim))
        self.register_buffer("inv_freq", inv_freq, persistent=False)

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


class OpenPanguRMSNorm(PanguEmbeddedRMSNorm):
    pass


class OpenPanguVLPatchMerger(nn.Module):
    def __init__(self, dim: int, context_dim: int, spatial_merge_size: int = 2) -> None:
        super().__init__()
        self.hidden_size = context_dim * (spatial_merge_size**2)
        self.ln_q = OpenPanguRMSNorm(context_dim, eps=1e-6)
        self.mlp = nn.Sequential(
            nn.Linear(self.hidden_size, self.hidden_size),
            nn.GELU(),
            nn.Linear(self.hidden_size, dim),
        )

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        x = self.mlp(self.ln_q(x).view(-1, self.hidden_size))
        return x


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


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


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


class OpenPanguVLVisionAttention(nn.Module):
    def __init__(self, config: OpenPanguVLVisionConfig) -> None:
        super().__init__()
        self.dim = config.hidden_size
        self.num_heads = config.num_heads
        self.head_dim = self.dim // self.num_heads
        self.num_key_value_groups = 1  # needed for eager attention
        self.qkv = nn.Linear(self.dim, self.dim * 3, bias=True)
        self.proj = nn.Linear(self.dim, self.dim)
        self.scaling = self.head_dim**-0.5
        self.config = config
        self.attention_dropout = 0.0
        self.is_causal = False

    def forward(
        self,
        hidden_states: torch.Tensor,
        cu_seqlens: torch.Tensor,
        rotary_pos_emb: Optional[torch.Tensor] = None,
        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
        attention_mask: Optional[torch.Tensor] = None,
        **kwargs,
    ) -> torch.Tensor:
        seq_length = hidden_states.shape[0]
        query_states, key_states, value_states = (
            self.qkv(hidden_states).reshape(seq_length, 3, self.num_heads, -1).permute(1, 0, 2, 3).unbind(0)
        )
        if position_embeddings is None:
            logger.warning_once(
                "The attention layers in this model are transitioning from computing the RoPE embeddings internally "
                "through `rotary_pos_emb` (2D tensor of RoPE theta values), to using externally computed "
                "`position_embeddings` (Tuple of tensors, containing cos and sin). In v4.54 `rotary_pos_emb` will be "
                "removed and `position_embeddings` will be mandatory."
            )
            emb = torch.cat((rotary_pos_emb, rotary_pos_emb), dim=-1)
            cos = emb.cos()
            sin = emb.sin()
        else:
            cos, sin = position_embeddings
        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)
        max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max().item()

        attention_interface: Callable = eager_attention_forward
        if self.config._attn_implementation != "eager":
            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]

        if not self.training and NPU_ATTN_INFR:
            if isinstance(cu_seqlens, torch.Tensor):
                cu_seqlens = cu_seqlens.tolist()

            q, k, v = [rearrange(x, "b n s d -> (b s) n d") for x in [query_states, key_states, value_states]]
            attn_output = torch_npu.npu_fusion_attention(
                q,
                k,
                v,
                self.num_heads,
                "TND",
                pse=None,
                padding_mask=None,
                atten_mask=None,
                scale=self.scaling,
                pre_tockens=1048576,
                next_tockens=0,
                keep_prob=1.0,
                inner_precise=0,
                sparse_mode=0,
                actual_seq_qlen=cu_seqlens,
                actual_seq_kvlen=cu_seqlens,
            )[0]
        else:
            attn_output, _ = attention_interface(
                self,
                query_states,
                key_states,
                value_states,
                attention_mask=attention_mask,
                dropout=0.0 if not self.training else self.attention_dropout,
                scaling=self.scaling,
                cu_seq_lens_q=cu_seqlens,  # pass cu seq lens for FA2
                cu_seq_lens_k=cu_seqlens,
                max_length_q=max_seqlen,
                max_length_k=max_seqlen,
                is_causal=False,
                **kwargs,
            )

        attn_output = attn_output.reshape(seq_length, -1).contiguous()
        attn_output = self.proj(attn_output)
        return attn_output


class OpenPanguVLVisionBlock(GradientCheckpointingLayer):
    def __init__(self, config, attn_implementation: str = "sdpa") -> None:
        super().__init__()
        self.norm1 = OpenPanguRMSNorm(config.hidden_size, eps=1e-6)
        self.norm2 = OpenPanguRMSNorm(config.hidden_size, eps=1e-6)
        self.attn = OpenPanguVLVisionAttention(config=config)
        self.mlp = OpenPanguVLMLP(config, bias=True)

    def forward(
        self,
        hidden_states: torch.Tensor,
        cu_seqlens: torch.Tensor,
        rotary_pos_emb: Optional[torch.Tensor] = None,
        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
        attention_mask: Optional[torch.Tensor] = None,
        **kwargs,
    ) -> torch.Tensor:
        hidden_states = hidden_states + self.attn(
            self.norm1(hidden_states),
            cu_seqlens=cu_seqlens,
            rotary_pos_emb=rotary_pos_emb,
            position_embeddings=position_embeddings,
            attention_mask=attention_mask,
            **kwargs,
        )
        hidden_states = hidden_states + self.mlp(self.norm2(hidden_states))
        return hidden_states


@auto_docstring
class OpenPanguPreTrainedModel(PreTrainedModel):
    config_class = OpenPanguConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["OpenPanguVLDecoderLayer", "OpenPanguVLVisionBlock"]
    _skip_keys_device_placement = "past_key_values"
    _supports_flash_attn_2 = True
    _supports_sdpa = True
    _supports_cache_class = True
    _supports_static_cache = True
    _supports_attention_backend = True

    def _init_weights(self, module):
        std = self.config.get_text_config().initializer_range
        if isinstance(module, (nn.Linear, nn.Conv3d)):
            module.weight.data.normal_(mean=0.0, std=std)
            if module.bias is not None:
                module.bias.data.zero_()
        elif isinstance(module, nn.Embedding):
            module.weight.data.normal_(mean=0.0, std=std)
            if module.padding_idx is not None:
                module.weight.data[module.padding_idx].zero_()
        elif isinstance(module, OpenPanguRMSNorm):
            module.weight.data.fill_(1.0)


class OpenPanguVisionTransformerPretrainedModel(OpenPanguPreTrainedModel):
    config_class = OpenPanguVLVisionConfig
    _no_split_modules = ["OpenPanguVLVisionBlock"]

    def __init__(self, config, *inputs, **kwargs) -> None:
        super().__init__(config, *inputs, **kwargs)
        self.spatial_merge_size = config.spatial_merge_size
        self.patch_size = config.patch_size
        self.fullatt_block_indexes = config.fullatt_block_indexes
        self.window_size = config.window_size
        self.spatial_merge_unit = self.spatial_merge_size * self.spatial_merge_size
        self.patch_embed = OpenPanguVLPatchEmbed(
            patch_size=config.patch_size,
            temporal_patch_size=config.temporal_patch_size,
            in_channels=config.in_channels,
            embed_dim=config.hidden_size,
        )
        head_dim = config.hidden_size // config.num_heads
        self.rotary_pos_emb = OpenPanguVisionRotaryEmbedding(head_dim // 2)
        self.blocks = nn.ModuleList([OpenPanguVLVisionBlock(config) for _ in range(config.depth)])
        self.select_layer = getattr(config, "mm_unit_vision_select_layer", [-1, -3])
        self.select_index = [config.depth + i for i in self.select_layer]
        self.select_index = self.select_index[::-1]
        self.select_layer = [-1 * (i + 1) for i in range(len(self.select_index))]
        self.merger = nn.ModuleList(
            [
                OpenPanguVLPatchMerger(
                    dim=config.out_hidden_size,
                    context_dim=config.hidden_size,
                    spatial_merge_size=config.spatial_merge_size,
                )
                for i in range(len(self.select_layer))
            ]
        )
        self.gradient_checkpointing = False
        self.take_indices = self.select_index
        self.final_layernorm = OpenPanguRMSNorm(config.hidden_size, eps=1e-6)

    def rot_pos_emb(self, grid_thw):
        pos_ids = []
        for t, h, w in grid_thw:
            hpos_ids = torch.arange(h).unsqueeze(1).expand(-1, w)
            hpos_ids = hpos_ids.reshape(
                h // self.spatial_merge_size,
                self.spatial_merge_size,
                w // self.spatial_merge_size,
                self.spatial_merge_size,
            )
            hpos_ids = hpos_ids.permute(0, 2, 1, 3)
            hpos_ids = hpos_ids.flatten()

            wpos_ids = torch.arange(w).unsqueeze(0).expand(h, -1)
            wpos_ids = wpos_ids.reshape(
                h // self.spatial_merge_size,
                self.spatial_merge_size,
                w // self.spatial_merge_size,
                self.spatial_merge_size,
            )
            wpos_ids = wpos_ids.permute(0, 2, 1, 3)
            wpos_ids = wpos_ids.flatten()
            pos_ids.append(torch.stack([hpos_ids, wpos_ids], dim=-1).repeat(t, 1))
        pos_ids = torch.cat(pos_ids, dim=0)
        max_grid_size = grid_thw[:, 1:].max()
        rotary_pos_emb_full = self.rotary_pos_emb(max_grid_size)
        rotary_pos_emb = rotary_pos_emb_full[pos_ids].flatten(1)
        return rotary_pos_emb

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

        for grid_t, grid_h, grid_w in grid_thw:
            llm_grid_h, llm_grid_w = (
                grid_h // self.spatial_merge_size,
                grid_w // self.spatial_merge_size,
            )
            index = torch.arange(grid_t * llm_grid_h * llm_grid_w).reshape(grid_t, llm_grid_h, llm_grid_w)
            pad_h = vit_merger_window_size - llm_grid_h % vit_merger_window_size
            pad_w = vit_merger_window_size - llm_grid_w % vit_merger_window_size
            num_windows_h = (llm_grid_h + pad_h) // vit_merger_window_size
            num_windows_w = (llm_grid_w + pad_w) // vit_merger_window_size
            index_padded = F.pad(index, (0, pad_w, 0, pad_h), "constant", -100)
            index_padded = index_padded.reshape(
                grid_t,
                num_windows_h,
                vit_merger_window_size,
                num_windows_w,
                vit_merger_window_size,
            )
            index_padded = index_padded.permute(0, 1, 3, 2, 4).reshape(
                grid_t,
                num_windows_h * num_windows_w,
                vit_merger_window_size,
                vit_merger_window_size,
            )
            seqlens = (index_padded != -100).sum([2, 3]).reshape(-1)
            index_padded = index_padded.reshape(-1)
            index_new = index_padded[index_padded != -100]
            window_index.append(index_new + window_index_id)
            cu_seqlens_tmp = seqlens.cumsum(0) * self.spatial_merge_unit + cu_window_seqlens[-1]
            cu_window_seqlens.extend(cu_seqlens_tmp.tolist())
            window_index_id += (grid_t * llm_grid_h * llm_grid_w).item()
        window_index = torch.cat(window_index, dim=0)

        return window_index, cu_window_seqlens

    def _prepare_attention_mask(self, inputs_tensor: torch.Tensor, cu_seqlens: torch.Tensor) -> torch.Tensor:
        # Flash Attention 2 doesn't need a 4D mask and relies on `cu_seqlens/max_seqlen`
        # NOTE: the created attention masl only approximates the ragged FA2 attention by
        # allowing bidirectional attention within `cu_seqlens` blocks, and not attending between
        # blocks. Though it will not be a 100% match for FA2's `varlen` path
        if self.config._attn_implementation == "flash_attention_2":
            return None

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

    def forward(self, hidden_states: torch.Tensor, grid_thw: torch.Tensor, **kwargs) -> torch.Tensor:
        """
        Args:
            hidden_states (`torch.Tensor` of shape `(seq_len, hidden_size)`):
                The final hidden states of the model.
            grid_thw (`torch.Tensor` of shape `(num_images_or_videos, 3)`):
                The temporal, height and width of feature shape of each image in LLM.

        Returns:
            `torch.Tensor`: hidden_states.
        """
        hidden_states = self.patch_embed(hidden_states)
        rotary_pos_emb = self.rot_pos_emb(grid_thw)
        window_index, cu_window_seqlens = self.get_window_index(grid_thw)
        cu_window_seqlens = torch.tensor(
            cu_window_seqlens,
            device=hidden_states.device,
            dtype=grid_thw.dtype if torch.jit.is_tracing() else torch.int32,
        )
        cu_window_seqlens = torch.unique_consecutive(cu_window_seqlens)

        seq_len, _ = hidden_states.size()
        hidden_states = hidden_states.reshape(seq_len // self.spatial_merge_unit, self.spatial_merge_unit, -1)
        hidden_states = hidden_states[window_index, :, :]
        hidden_states = hidden_states.reshape(seq_len, -1)
        rotary_pos_emb = rotary_pos_emb.reshape(seq_len // self.spatial_merge_unit, self.spatial_merge_unit, -1)
        rotary_pos_emb = rotary_pos_emb[window_index, :, :]
        rotary_pos_emb = rotary_pos_emb.reshape(seq_len, -1)
        emb = torch.cat((rotary_pos_emb, rotary_pos_emb), dim=-1)
        position_embeddings = (emb.cos(), emb.sin())

        cu_seqlens = torch.repeat_interleave(grid_thw[:, 1] * grid_thw[:, 2], 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=grid_thw.dtype if torch.jit.is_tracing() else torch.int32,
        )
        cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0)
        intermediates = []
        for layer_num, blk in enumerate(self.blocks):
            if layer_num in self.fullatt_block_indexes:
                cu_seqlens_now = cu_seqlens
            else:
                cu_seqlens_now = cu_window_seqlens

            attention_mask = self._prepare_attention_mask(hidden_states, cu_seqlens_now)
            hidden_states = blk(
                hidden_states,
                cu_seqlens=cu_seqlens_now,
                position_embeddings=position_embeddings,
                attention_mask=attention_mask,
                **kwargs,
            )
            if layer_num in self.take_indices:
                ln_hs = self.final_layernorm(hidden_states)
                intermediates.append(ln_hs)

        image_embeddings_list = []
        for idx, sl in enumerate(self.select_layer):
            image_embeddings_list.append(self.merger[idx](intermediates[sl]))
        hidden_states = sum(image_embeddings_list)

        reverse_indices = torch.argsort(window_index)
        hidden_states = hidden_states[reverse_indices, :]

        return hidden_states


@dataclass
@auto_docstring(
    custom_intro="""
    Base class for Llava outputs, with hidden states and attentions.
    """
)
class OpenPanguVLModelOutputWithPast(ModelOutput):
    r"""
    past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
        Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
        `(batch_size, num_heads, sequence_length, embed_size_per_head)`)

        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: torch.FloatTensor = None
    past_key_values: Optional[list[torch.FloatTensor]] = None
    hidden_states: Optional[tuple[torch.FloatTensor]] = None
    attentions: Optional[tuple[torch.FloatTensor]] = None
    rope_deltas: Optional[torch.LongTensor] = None


class OpenPanguVLRotaryEmbedding(nn.Module):
    def __init__(self, config: OpenPanguVLTextConfig, device=None):
        super().__init__()
        # BC: "rope_type" was originally "type"
        if hasattr(config, "rope_scaling") and config.rope_scaling is not None:
            self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
            self.mrope_interleaved = config.rope_scaling.get("mrope_interleaved", False)
        else:
            self.rope_type = "default"
        self.max_seq_len_cached = config.max_position_embeddings
        self.original_max_seq_len = config.max_position_embeddings

        self.config = config
        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]

        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
        self.register_buffer("inv_freq", inv_freq, persistent=False)
        self.original_inv_freq = self.inv_freq

        mrope_section = config.rope_scaling.get("mrope_section", None)
        self.mrope_section = mrope_section
        if self.mrope_interleaved:
            if not self.mrope_section:
                raise AssertionError("when you use interleave mrope, mrope_section cannot be None.")

            # Generate interleaved indices
            if len(mrope_section) == 2:
                h_num, w_num = mrope_section[0], mrope_section[1]
                mrope_dim = self.get_mrope_interleaved_id_list(h_num, w_num, 0)
            elif len(mrope_section) == 3:
                t_num, h_num, w_num = mrope_section[0], mrope_section[1], mrope_section[2]
                mrope_dim = self.get_mrope_interleaved_id_list(t_num, h_num, w_num, force_last=True)
            else:
                raise AssertionError("Cannot support the length of mrope section is not 2 or 3.")
            mrope_dim = mrope_dim * 2
            self.mrope_dim = mrope_dim

    @torch.no_grad()
    @dynamic_rope_update  # power user: used with advanced RoPE types (e.g. dynamic rope)
    def forward(self, x, position_ids):
        # In contrast to other models, OpenPanguVL 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 torch.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)
            # mrope interleaved
            if self.mrope_interleaved:
                mrope_section_3d = [1] * len(self.mrope_dim)
                mrope_dim = self.mrope_dim
                emb = torch.cat([m[mrope_dim[i]] for i, m in enumerate(emb.split(mrope_section_3d, dim=-1))], dim=-1)

            cos = emb.cos() * self.attention_scaling
            sin = emb.sin() * self.attention_scaling
            # normal mrope
            if not self.mrope_interleaved and self.mrope_section:
                mrope_section = self.mrope_section * 2
                cos = torch.cat([m[i % 3] for i, m in enumerate(cos.split(mrope_section, dim=-1))], dim=-1)
                sin = torch.cat([m[i % 3] for i, m in enumerate(sin.split(mrope_section, dim=-1))], dim=-1)

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

    @staticmethod
    def get_mrope_interleaved_id_list(a: int, b: int, c: int, force_last: bool = False) -> list[int]:
        """
        Generate an interleaved list of indices for multi-modal rotary embedding.

        Args:
            a: Number of indices for first modality
            b: Number of indices for second modality
            c: Number of indices for third modality
            force_last: Whether to force the last element to be from the first modality

        Returns:
            List of interleaved indices
        """
        if force_last:
            a -= 1

        counts = {0: a, 1: b, 2: c}
        placed = dict.fromkeys(counts, 0)
        rem = counts.copy()
        seq: list[int] = []
        last = None

        total = a + b + c
        for _ in range(total):
            # Candidates: remaining > 0 and ≠ last
            cands = [k for k in rem if rem[k] > 0 and k != last]
            if not cands:
                # If only last remains, relax the condition
                cands = [k for k in rem if rem[k] > 0]

            # Select the rarest candidate
            try:
                best = min(cands, key=lambda k: (placed[k] / counts[k], k))
            except KeyError:
                best = 0

            seq.append(best)
            placed[best] += 1
            rem[best] -= 1
            last = best

        if force_last:
            seq.append(0)

        return seq


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 OpenPanguVLAttention(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: OpenPanguVLTextConfig, 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 = 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 = config.attention_dropout
        self.rope_scaling = config.rope_scaling
        self.scaling = self.head_dim**-0.5

        if (self.head_dim * self.num_heads) != self.hidden_size:
            raise ValueError(
                f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
                f" and `num_heads`: {self.num_heads})."
            )
        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=True)
        self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=True)
        self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=True)
        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
        self.sliding_window = config.sliding_window if config.layer_types[layer_idx] == "sliding_attention" else None
        self.rotary_emb = OpenPanguVLRotaryEmbedding(config=config)

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_value: 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,  # necessary, but kept here for BC
        **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.rope_scaling["mrope_section"]
        )

        if past_key_value is not None:
            cache_kwargs = {
                "sin": sin,
                "cos": cos,
                "cache_position": cache_position,
            }  # Specific to RoPE models
            key_states, value_states = past_key_value.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,
            **kwargs,
        )

        attn_output = attn_output.reshape(bsz, q_len, -1).contiguous()
        attn_output = self.o_proj(attn_output)
        return attn_output, attn_weights, past_key_value


class OpenPanguVLDecoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: OpenPanguVLTextConfig, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size

        if config.use_sliding_window and config._attn_implementation != "flash_attention_2":
            logger.warning_once(
                f"Sliding Window Attention is enabled but not implemented for `{config._attn_implementation}`; "
                "unexpected results may be encountered."
            )
        self.self_attn = OpenPanguVLAttention(config, layer_idx)
        self.mlp = OpenPanguVLMLP(config)
        self.input_layernorm = OpenPanguRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = OpenPanguRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.attention_type = config.layer_types[layer_idx]

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_value: Optional[tuple[torch.Tensor]] = None,
        output_attentions: Optional[bool] = False,
        use_cache: Optional[bool] = False,
        cache_position: Optional[torch.LongTensor] = None,
        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> tuple[torch.FloatTensor, Optional[tuple[torch.FloatTensor, torch.FloatTensor]]]:
        """
        Args:
            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
            attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
                `(batch, sequence_length)` where padding elements are indicated by 0.
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
            use_cache (`bool`, *optional*):
                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
                (see `past_key_values`).
            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
            cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
                Indices depicting the position of the input sequence tokens in the sequence.
            position_embeddings (`tuple[torch.FloatTensor, torch.FloatTensor]`, *optional*):
                Tuple containing the cosine and sine positional embeddings of shape `(batch_size, seq_len, head_dim)`,
                with `head_dim` being the embedding dimension of each attention head.
            kwargs (`dict`, *optional*):
                Arbitrary kwargs to be ignored, used for FSDP and other methods that injects code
                into the model
        """

        residual = hidden_states

        hidden_states = self.input_layernorm(hidden_states)

        # Self Attention
        hidden_states, self_attn_weights, present_key_value = self.self_attn(
            hidden_states=hidden_states,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_value=past_key_value,
            output_attentions=output_attentions,
            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

        outputs = (hidden_states,)

        if output_attentions:
            outputs += (self_attn_weights,)

        if use_cache:
            outputs += (present_key_value,)

        return outputs


class KwargsForCausalLM(FlashAttentionKwargs, LossKwargs): ...


class ProjectionSingle(nn.Module):
    def __init__(self, i_hidden_size: int, t_hidden_size: int):
        super().__init__()
        self.act = F.silu
        self.fc1 = nn.Linear(i_hidden_size, t_hidden_size, bias=True)  # bias 默认为 True

    def forward(self, hidden_states):
        x = self.act(hidden_states)
        return self.fc1(x)


@auto_docstring
class OpenPanguVLTextModel(PanguEmbeddedModel):
    def __init__(self, config: PanguEmbeddedConfig):
        super().__init__(config)
        self.rotary_emb = OpenPanguVLRotaryEmbedding(config=config)


@auto_docstring
class OpenPanguVLModel(OpenPanguPreTrainedModel):
    base_model_prefix = ""
    _checkpoint_conversion_mapping = {"^model": "language_model"}
    config_class = OpenPanguConfig
    _no_split_modules = ["OpenPanguVLDecoderLayer", "OpenPanguVLVisionBlock"]

    def __init__(self, config):
        super().__init__(config)
        self.visual = OpenPanguVisionTransformerPretrainedModel._from_config(config.vision_config)
        self.language_model = OpenPanguVLTextModel(config.text_config)

        self.rope_deltas = None  # cache rope_deltas here

        self.visual.vision_projection = ProjectionSingle(config.vision_config.out_hidden_size, config.hidden_size)

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

    def _parse_preprocess_params(self, vision_config):
        self.channel = vision_config.in_channels
        self.patch_size = vision_config.patch_size
        from transformers import AutoProcessor
        processor = AutoProcessor.from_pretrained(self.config.name_or_path, trust_remote_code=True, local_files_only=True)
        self.do_rescale = processor.image_processor.do_rescale
        self.rescale_factor = processor.image_processor.rescale_factor
        self.do_normalize = processor.image_processor.do_normalize
        self.image_mean = tuple(processor.image_processor.image_mean)
        self.image_std = tuple(processor.image_processor.image_std)

    def get_input_embeddings(self):
        return self.language_model.get_input_embeddings()

    def set_input_embeddings(self, value):
        self.language_model.set_input_embeddings(value)

    def set_decoder(self, decoder):
        self.language_model = decoder

    def get_decoder(self):
        return self.language_model

    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,
        second_per_grid_ts: Optional[torch.Tensor] = 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:
                Temporal (Time): 3 patches, representing different segments of the video in time.
                Height: 2 patches, dividing each frame vertically.
                Width: 2 patches, dividing each frame horizontally.
                We also have some important parameters:
                fps (Frames Per Second): The video's frame rate, set to 1.
                    This means one frame is processed each second.
                tokens_per_second: This is a crucial parameter. It dictates how many "time-steps" or "temporal tokens"
                    are conceptually packed into a one-second interval of the video. In this case, we have 25 tokens
                    per second. So each second of the video will be represented with 25 separate time points.
                    It essentially defines the temporal granularity.
                temporal_patch_size: The number of frames that compose one temporal patch. Here, it's 2 frames.
                interval: The step size for the temporal position IDs, calculated as
                    tokens_per_second * temporal_patch_size / fps. In this case, 25 * 2 / 1 = 50.
                    This means that each temporal patch will be have a difference of 50 in the temporal position IDs.
                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, 50, 50, 50, 50, 100, 100, 100, 100]
                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: [101, 102, 103, 104, 105]
                text height position_ids: [101, 102, 103, 104, 105]
                text width position_ids: [101, 102, 103, 104, 105]
                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.
            second_per_grid_ts (`torch.Tensor` of shape `(num_videos)`, *optional*):
                The time interval (in seconds) for each grid along the temporal dimension in the 3D position IDs.
            attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
                Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:

                - 1 for tokens that are **not masked**,
                - 0 for tokens that are **masked**.

        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
        vision_end_token_id = self.config.vision_end_token_id
        tokens_per_second = getattr(self.config, "tokens_per_second", 1.0)
        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,
            )
            attention_mask = attention_mask.to(total_input_ids.device)
            for i, input_ids in enumerate(total_input_ids):
                input_ids = input_ids[attention_mask[i] == 1]
                input_tokens = input_ids.tolist()
                src_item = input_tokens
                video_idx = 0
                image_idx = 0
                new_src_item: list[int] = []
                llm_pos_ids_list: list[torch.Tensor] = []

                idx = 0
                while idx < len(src_item):
                    new_src_item_len = len(new_src_item)
                    start_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0
                    if src_item[idx] not in [video_token_id, image_token_id]:
                        new_src_item.append(src_item[idx])
                        llm_pos_ids = torch.tensor([start_idx], dtype=torch.long).expand(3, -1)
                        llm_pos_ids_list.append(llm_pos_ids.to(position_ids.device))
                    elif src_item[idx] == image_token_id:
                        grid_t = image_grid_thw[image_idx][0]
                        grid_hs = image_grid_thw[:, 1]
                        grid_ws = image_grid_thw[:, 2]
                        t_index = (torch.arange(grid_t) * 1 * tokens_per_second).long()
                        llm_pos_ids = self._get_llm_pos_ids_for_vision(
                            start_idx, image_idx, spatial_merge_size, t_index, grid_hs, grid_ws
                        )
                        llm_pos_ids_list.append(llm_pos_ids.to(position_ids.device))
                        vision_seqlen = image_grid_thw[image_idx].prod() // (spatial_merge_size**2)
                        new_src_item.extend([image_token_id] * vision_seqlen)
                        image_idx += 1
                    else:
                        # src_item[idx] == video_token_id
                        # Get the grid information of the current video
                        T = video_grid_thw[video_idx][0].item()
                        H = video_grid_thw[video_idx][1].item()
                        W = video_grid_thw[video_idx][2].item()
                        llm_H = H // spatial_merge_size
                        llm_W = W // spatial_merge_size
                        tokens_per_frame = llm_H * llm_W
                        # Get timestamps (one t value per frame)
                        t_index_all = (torch.arange(T)).long()
                        # Calculate the current starting position
                        start_pos = llm_pos_ids_list[-1].max().item() + 1 if llm_pos_ids_list else 0
                        current_pos = start_pos
                        # frame by frame processing
                        final_frame_time = T - 1 # Record the order of the last frame
                        for t in range(T):
                            # 1. Calculate the left placeholder position of the first frame, skip
                            if t != 0:
                                new_src_item.append(vision_start_token_id) # For looping, count
                                bot_pos = torch.full((3, 1), current_pos, dtype=torch.long)
                                llm_pos_ids_list.append(bot_pos.to(position_ids.device))
                                current_pos += 1
                            # 2. Video tokens for frame t
                            # Construct a single frame of (t, h, w)
                            grid_h = torch.arange(llm_H).view(-1, 1).expand(-1, llm_W).flatten()
                            grid_w = torch.arange(llm_W).view(1, -1).expand(llm_H, -1).flatten()
                            # Here we don't add current_pos to h/w, just keep the original (t, h, w)
                            frame_pos = torch.stack(
                                [
                                    torch.full_like(grid_h, 0, dtype=torch.long),      # t
                                    grid_h,                                            # h
                                    grid_w                                             # w
                                ]
                            )  # shape: (3, tokens_per_frame)
                            frame_pos_with_offset = frame_pos + current_pos # Current frame position offset
                            new_src_item.extend([video_token_id] * tokens_per_frame) # For looping, count
                            llm_pos_ids_list.append(frame_pos_with_offset.to(position_ids.device))
                            current_pos += max(llm_H, llm_W)
                            # 3. Calculate the right placeholder position of the last frame and skip it
                            if t != final_frame_time:
                                new_src_item.append(vision_end_token_id) # For looping, count
                                eot_pos = torch.full((3, 1), current_pos, dtype=torch.long)
                                llm_pos_ids_list.append(eot_pos.to(position_ids.device))
                                current_pos += 1
                        video_idx += 1
                    # move to the next token
                    idx += len(new_src_item) - new_src_item_len
                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_delta = llm_positions.max() + 1 - len(src_item)
                mrope_position_deltas.append(mrope_position_delta)
            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_llm_pos_ids_for_vision(
        self,
        start_idx: int,
        vision_idx: int,
        spatial_merge_size: int,
        t_index: list[int],
        grid_hs: torch.Tensor,
        grid_ws: torch.Tensor,
    ) -> torch.Tensor:
        llm_pos_ids_list = []
        llm_grid_h = grid_hs[vision_idx] // spatial_merge_size
        llm_grid_w = grid_ws[vision_idx] // spatial_merge_size
        h_index = (
            torch.arange(llm_grid_h)
            .to(llm_grid_h.device)
            .view(1, -1, 1)
            .expand(len(t_index), -1, llm_grid_w)
            .flatten()
        )
        w_index = (
            torch.arange(llm_grid_w)
            .to(llm_grid_h.device)
            .view(1, 1, -1)
            .expand(len(t_index), llm_grid_h, -1)
            .flatten()
        )
        t_index_tensor = (
            torch.Tensor(t_index)
            .to(llm_grid_h.device)
            .view(-1, 1)
            .expand(-1, llm_grid_h * llm_grid_w)
            .long()
            .flatten()
        )
        _llm_pos_ids = torch.stack([t_index_tensor, h_index, w_index])
        llm_pos_ids_list.append(_llm_pos_ids + start_idx)
        llm_pos_ids = torch.cat(llm_pos_ids_list, dim=1)
        return llm_pos_ids

    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)

        video_embeds = self.visual.vision_projection(video_embeds)

        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)
        # rescale and normalize
        pixel_values = pixel_values.reshape(-1, self.channel, self.patch_size, self.patch_size)
        pixel_values = rescale_and_normalize(pixel_values, self.do_rescale, self.rescale_factor, self.do_normalize,
                                                self.image_mean, self.image_std)
        pixel_values = pixel_values.reshape(-1, self.channel * self.patch_size * self.patch_size)
        image_embeds = self.visual(pixel_values, grid_thw=image_grid_thw)

        image_embeds = self.visual.vision_projection(image_embeds)

        split_sizes = (image_grid_thw.prod(-1) // self.visual.spatial_merge_size**2).tolist()
        image_embeds = torch.split(image_embeds, split_sizes)
        return image_embeds

    @auto_docstring
    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,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        pixel_values: Optional[torch.Tensor] = None,
        pixel_values_videos: Optional[torch.FloatTensor] = None,
        image_grid_thw: Optional[torch.LongTensor] = None,
        video_grid_thw: Optional[torch.LongTensor] = None,
        rope_deltas: Optional[torch.LongTensor] = None,
        cache_position: Optional[torch.LongTensor] = None,
        second_per_grid_ts: Optional[torch.Tensor] = None,
        **kwargs: Unpack[KwargsForCausalLM],
    ) -> Union[tuple, OpenPanguVLModelOutputWithPast]:
        r"""
        pixel_values_videos (`torch.FloatTensor` of shape `(seq_length,
            num_channels * temporal_size * image_size * image_size)):
            The tensors corresponding to the input videos. Pixel values can be obtained using
            [`AutoImageProcessor`]. See [`OpenPanguVLImageProcessor.__call__`] for details. [`OpenPanguVLProcessor`] uses
            [`OpenPanguVLImageProcessor`] for processing videos.
        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.
        rope_deltas (`torch.LongTensor` of shape `(batch_size, )`, *optional*):
            The rope index difference between sequence length and multimodal rope.
        second_per_grid_ts (`torch.Tensor` of shape `(num_videos)`, *optional*):
            The time interval (in seconds) for each grid along the temporal dimension in the 3D position IDs.
        """

        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        if inputs_embeds is None:
            inputs_embeds = self.get_input_embeddings()(input_ids)
            if pixel_values is not None:
                image_embeds = self.get_image_features(pixel_values, image_grid_thw)
                image_embeds = torch.cat(image_embeds, dim=0)
                n_image_tokens = (input_ids == self.config.image_token_id).sum()
                n_image_features = image_embeds.shape[0]
                if not is_torchdynamo_compiling() and n_image_tokens != n_image_features:
                    raise ValueError(
                        "Image features and image tokens do not match: "
                        f"tokens: {n_image_tokens}, features {n_image_features}"
                    )

                mask = input_ids == self.config.image_token_id
                mask_unsqueezed = mask.unsqueeze(-1)
                mask_expanded = mask_unsqueezed.expand_as(inputs_embeds)
                image_mask = mask_expanded.to(inputs_embeds.device)

                image_embeds = image_embeds.to(inputs_embeds.device, inputs_embeds.dtype)
                inputs_embeds = inputs_embeds.masked_scatter(image_mask, image_embeds)

            if pixel_values_videos is not None:
                video_embeds = self.get_video_features(pixel_values_videos, video_grid_thw)
                video_embeds = torch.cat(video_embeds, dim=0)
                n_video_tokens = (input_ids == self.config.video_token_id).sum()
                n_video_features = video_embeds.shape[0]
                if not is_torchdynamo_compiling() and n_video_tokens != n_video_features:
                    raise ValueError(
                        "Video features and video tokens do not match: "
                        f"tokens: {n_video_tokens}, features {n_video_features}"
                    )

                mask = input_ids == self.config.video_token_id
                mask_unsqueezed = mask.unsqueeze(-1)
                mask_expanded = mask_unsqueezed.expand_as(inputs_embeds)
                video_mask = mask_expanded.to(inputs_embeds.device)

                video_embeds = video_embeds.to(inputs_embeds.device, inputs_embeds.dtype)
                inputs_embeds = inputs_embeds.masked_scatter(video_mask, video_embeds)

        if position_ids is None:
            attention_mask_tensor = (
                attention_mask if not isinstance(attention_mask, dict) else attention_mask["full_attention"]
            )
            if attention_mask_tensor is not None and attention_mask_tensor.ndim == 4:
                attention_mask_tensor = torch.diagonal(attention_mask_tensor[:, 0], dim1=1, dim2=2)
                attention_mask_tensor = attention_mask_tensor / torch.finfo(attention_mask_tensor.dtype).min
                attention_mask_tensor = (1.0 - attention_mask_tensor).int()

            # 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
            prefill_compiled_stage = is_torchdynamo_compiling() and (
                (input_ids is not None and input_ids.shape[1] != 1)
                or (inputs_embeds is not None and inputs_embeds.shape[1] != 1)
            )
            prefill_noncompiled_stage = not is_torchdynamo_compiling() and (
                (cache_position is not None and cache_position[0] == 0)
                or (past_key_values is None or past_key_values.get_seq_length() == 0)
            )
            if (prefill_compiled_stage or prefill_noncompiled_stage) or self.rope_deltas is None:
                position_ids, rope_deltas = self.get_rope_index(
                    input_ids,
                    image_grid_thw,
                    video_grid_thw,
                    second_per_grid_ts=second_per_grid_ts,
                    attention_mask=attention_mask_tensor,
                )
                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
                delta = (
                    (cache_position[0] + self.rope_deltas).to(inputs_embeds.device)
                    if cache_position is not None
                    else 0
                )
                position_ids = torch.arange(seq_length, device=inputs_embeds.device)
                position_ids = position_ids.view(1, -1).expand(batch_size, -1)
                if cache_position is not None:  # otherwise `deltas` is an int `0`
                    delta = delta.repeat_interleave(batch_size // delta.shape[0], dim=0)
                position_ids = position_ids.add(delta)
                position_ids = position_ids.unsqueeze(0).expand(3, -1, -1)

        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,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=True,
            cache_position=cache_position,
            **kwargs,
        )

        output = OpenPanguVLModelOutputWithPast(
            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 if return_dict else output.to_tuple()


@dataclass
@auto_docstring(
    custom_intro="""
    Base class for OpenPanguVL causal language model (or autoregressive) outputs.
    """
)
class OpenPanguVLCausalLMOutputWithPast(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 (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
        Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
        `(batch_size, num_heads, sequence_length, embed_size_per_head)`)

        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[list[torch.FloatTensor]] = None
    hidden_states: Optional[tuple[torch.FloatTensor]] = None
    attentions: Optional[tuple[torch.FloatTensor]] = None
    rope_deltas: Optional[torch.LongTensor] = None


class OpenPanguVL(OpenPanguPreTrainedModel, GenerationMixin):
    _checkpoint_conversion_mapping = {
        "^visual": "model.visual",
        r"^model(?!\.(language_model|visual))": "model.language_model",
    }
    _tied_weights_keys = ["lm_head.weight"]

    def __init__(self, config):
        super().__init__(config)
        self.model = OpenPanguVLModel(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_output_embeddings(self):
        return self.lm_head

    def set_output_embeddings(self, new_embeddings):
        self.lm_head = new_embeddings

    def set_decoder(self, decoder):
        self.model.set_decoder(decoder)

    def get_decoder(self):
        return self.model.get_decoder()

    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)

    # Make modules available throught conditional class for BC
    @property
    def language_model(self):
        return self.model.language_model

    @property
    def visual(self):
        return self.model.visual

    @can_return_tuple
    @auto_docstring
    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,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        pixel_values: Optional[torch.Tensor] = None,
        pixel_values_videos: Optional[torch.FloatTensor] = None,
        image_grid_thw: Optional[torch.LongTensor] = None,
        video_grid_thw: Optional[torch.LongTensor] = None,
        rope_deltas: Optional[torch.LongTensor] = None,
        cache_position: Optional[torch.LongTensor] = None,
        second_per_grid_ts: Optional[torch.Tensor] = None,
        **kwargs: Unpack[KwargsForCausalLM],
    ) -> Union[tuple, OpenPanguVLCausalLMOutputWithPast]:
        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]`.
        pixel_values_videos (`torch.FloatTensor` of shape `(seq_length,
            num_channels * temporal_size * image_size * image_size)):
            The tensors corresponding to the input videos. Pixel values can be obtained using
            [`AutoImageProcessor`]. See [`OpenPanguVLImageProcessor.__call__`] for details. [`OpenPanguVLProcessor`] uses
            [`OpenPanguVLImageProcessor`] for processing videos.
        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.
        rope_deltas (`torch.LongTensor` of shape `(batch_size, )`, *optional*):
            The rope index difference between sequence length and multimodal rope.
        second_per_grid_ts (`torch.Tensor` of shape `(num_videos)`, *optional*):
            The time interval (in seconds) for each grid along the temporal dimension in the 3D position IDs.

        Example:

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

        >>> model = OpenPanguVLForConditionalGeneration.from_pretrained("Pangu/Pangu_7B_V5_VL_HF_vllm_ascend")
        >>> processor = AutoProcessor.from_pretrained("Pangu/Pangu_7B_V5_VL_HF_vllm_ascend")

        >>> messages = [
            {
                "role": "user",
                "content": [
                    {"type": "image"},
                    {"type": "text", "text": "What is shown in this image?"},
                ],
            },
        ]
        >>> image = Image.open(requests.get(url, stream=True).raw)

        >>> text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
        >>> inputs = processor(text=[text], images=[image], vision_infos=[vision_infos])

        >>> # Generate
        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
        "The image shows a street scene with a red stop sign in the foreground.
        In the background, there is a large red gate with Chinese characters ..."
        ```"""

        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )

        outputs = self.model(
            input_ids=input_ids,
            pixel_values=pixel_values,
            pixel_values_videos=pixel_values_videos,
            image_grid_thw=image_grid_thw,
            video_grid_thw=video_grid_thw,
            second_per_grid_ts=second_per_grid_ts,
            position_ids=position_ids,
            attention_mask=attention_mask,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=True,
            cache_position=cache_position,
            **kwargs,
        )

        hidden_states = outputs[0]
        logits = self.lm_head(hidden_states)

        loss = None
        if labels is not None:
            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size)

        return OpenPanguVLCausalLMOutputWithPast(
            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,
        second_per_grid_ts=None,
        **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,
            second_per_grid_ts=second_per_grid_ts,
            use_cache=use_cache,
            **kwargs,
        )

        # OpenPangu-VL position_ids are prepareed with rope_deltas in forward
        model_inputs["position_ids"] = None

        if cache_position[0] != 0:
            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],
    ) -> 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

        vision_start_mask = input_ids == vision_start_token_id
        vision_first_mask = torch.roll(vision_start_mask, shifts=1, dims=1)
        image_mask = input_ids == image_token_id
        video_mask = input_ids == video_token_id
        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)

            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":
                    if not isinstance(dict_to_expand[key], list):
                        raise TypeError(
                            f"Expected value for key '{key}' to be a list,but got {type(dict_to_expand[key])} instead."
                        )
                    tensor = torch.tensor(dict_to_expand[key])
                    lengths = list(video_nums)
                    tensor = _repeat_interleave_samples(tensor, lengths=lengths, repeat_times=expand_size)
                    dict_to_expand[key] = tensor.tolist()
            return dict_to_expand

        def _expand_dict_for_generation(dict_to_expand):
            for key in dict_to_expand:
                if key != "cache_position":
                    if (
                        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

        # input_ids is required for expanding visual inputs
        # If input_ids is unavailable, visual inputs will not be used;
        # therefore, there is no need to expand visual inputs.
        if input_ids is not None and input_ids.numel() != 0:
            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__ = ["OpenPanguVL", "OpenPanguVLModel", "OpenPanguPreTrainedModel"]