src/llamafactory/model/onellmpp/modeling_qwen2_5_omni.py

#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
#           This file was automatically generated from src/transformers/models/qwen2_5_omni/modular_qwen2_5_omni.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_qwen2_5_omni.py file directly. One of our CI enforces this.
#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
# coding=utf-8
# Copyright 2025 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved.
#
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

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

import numpy as np
import torch
import torch.nn.functional as F
from torch import nn
from torch.nn import Parameter

from transformers.activations import ACT2FN
from transformers.cache_utils import Cache, DynamicCache
from transformers.generation import GenerationMixin
from transformers.masking_utils import create_causal_mask, create_sliding_window_causal_mask
from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
from transformers.modeling_layers import GradientCheckpointingLayer
from transformers.modeling_outputs import BaseModelOutput, BaseModelOutputWithPast, 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 TransformersKwargs, auto_docstring, check_torch_load_is_safe, logging
from transformers.utils.deprecation import deprecate_kwarg
from transformers.utils.hub import cached_file
from transformers.models.qwen2.modeling_qwen2 import Qwen2RMSNorm
from .configuration_qwen2_5_omni import (
    Qwen2_5OmniAudioEncoderConfig,
    Qwen2_5OmniBigVGANConfig,
    Qwen2_5OmniConfig,
    Qwen2_5OmniDiTConfig,
    Qwen2_5OmniTalkerConfig,
    Qwen2_5OmniTextConfig,
    Qwen2_5OmniThinkerConfig,
    Qwen2_5OmniToken2WavConfig,
    Qwen2_5OmniVisionEncoderConfig,
)


logger = logging.get_logger(__name__)


@auto_docstring
class Qwen2_5OmniPreTrainedModel(PreTrainedModel):
    config: Qwen2_5OmniConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["Qwen2_5OmniDecoderLayer", "Qwen2_5OmniVisionBlock"]
    _skip_keys_device_placement = "past_key_values"
    _supports_flash_attn = True
    _supports_sdpa = True
    _can_compile_fullgraph = False
    _supports_attention_backend = True


class Qwen2_5OmniPreTrainedModelForConditionalGeneration(Qwen2_5OmniPreTrainedModel):
    def _prepare_4d_causal_attention_mask_with_cache_position(
        self,
        attention_mask: torch.Tensor,
        sequence_length: int,
        target_length: int,
        dtype: torch.dtype,
        device: torch.device,
        min_dtype: float,
        cache_position: torch.Tensor,
        batch_size: int,
    ):
        """
        Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
        `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.

        Args:
            attention_mask (`torch.Tensor`):
                A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`.
            sequence_length (`int`):
                The sequence length being processed.
            target_length (`int`):
                The target length: when generating with static cache, the mask should be as long as the static cache, to account for the 0 padding, the part of the cache that is not filled yet.
            dtype (`torch.dtype`):
                The dtype to use for the 4D attention mask.
            device (`torch.device`):
                The device to place the 4D attention mask on.
            min_dtype (`float`):
                The minimum value representable with the dtype `dtype`.
            cache_position (`torch.Tensor`):
                Indices depicting the position of the input sequence tokens in the sequence.
            batch_size (`torch.Tensor`):
                Batch size.
        """
        if attention_mask is not None and attention_mask.dim() == 4:
            # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
            causal_mask = attention_mask
        else:
            causal_mask = torch.full(
                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
            )
            if sequence_length != 1:
                causal_mask = torch.triu(causal_mask, diagonal=1)
            causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
            causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
            if attention_mask is not None:
                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
                mask_length = attention_mask.shape[-1]
                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
                padding_mask = padding_mask == 0
                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
                    padding_mask, min_dtype
                )

        return causal_mask

    def get_llm_pos_ids_for_vision(
        self,
        start_idx: int,
        vision_idx: int,
        spatial_merge_size: int,
        t_index: list[int],
        grid_hs: list[int],
        grid_ws: list[int],
    ):
        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).view(1, -1, 1).expand(len(t_index), -1, llm_grid_w).flatten()
        w_index = torch.arange(llm_grid_w).view(1, 1, -1).expand(len(t_index), llm_grid_h, -1).flatten()
        t_index = torch.Tensor(t_index).view(-1, 1).expand(-1, llm_grid_h * llm_grid_w).flatten().long()
        _llm_pos_ids = torch.stack([t_index, h_index, w_index])
        llm_pos_ids_list.append(_llm_pos_ids + start_idx)  # + 1 ) # 12.09 by malinhan
        llm_pos_ids = torch.cat(llm_pos_ids_list, dim=1)
        return llm_pos_ids

    def get_chunked_index(
        self, token_indices: torch.Tensor, tokens_per_chunk: int, remove_index: int
    ) -> list[tuple[int, int]]:
        """
        Splits token index list into chunks based on token value ranges.

        Given a list of token indices, returns a list of (start, end) index tuples representing
        slices of the list where the token values fall within successive ranges of `t_ntoken_per_chunk`.

        For example, if `t_ntoken_per_chunk` is 1000, the function will create chunks such that:
        - the first chunk contains token values < 1000,
        - the second chunk contains values >= 1000 and < 2000, and so on.

        Parameters:
            token_indices (`torch.Tensor` of shape `(seq_len, )`): A monotonically increasing list of
                                token index values.
            t_ntoken_per_chunk (`int`): Number of tokens per chunk (used as the chunk size threshold).
            remove_index (`int`) An index id to subtract from `token_indices` before chunking

        Returns:
            `list[tuple[int, int]]`: A list of tuples, each representing the start (inclusive)
                                and end (exclusive) indices of a chunk in `token_indices`.
        """

        def _iter():
            i, start_idx = 0, 0  # skip bos token
            current_chunk = 1
            while i < len(token_indices):  # skip eos token
                if token_indices[i] - remove_index >= current_chunk * tokens_per_chunk:
                    yield (start_idx, i)
                    start_idx = i
                    current_chunk += 1
                i += 1
            yield (start_idx, len(token_indices))

        return list(_iter())

    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,
        use_audio_in_video: bool = False,
        audio_seqlens: Optional[torch.LongTensor] = None,
        second_per_grids: 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.
            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**.
            use_audio_in_video (`bool`, *optional*):
                 If set to `True`, use the audio in video.
            audio_seqlens (`torch.LongTensor` of shape `(num_audios)`, *optional*):
                The length of feature shape of each audio in LLM.
            second_per_grids (`torch.LongTensor` of shape `(num_videos)`, *optional*):
                The time interval (in seconds) for each grid along the temporal dimension in the 3D position IDs.

        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.spatial_merge_size
        image_token_id = self.config.image_token_id
        video_token_id = self.config.video_token_id
        audio_token_id = self.config.audio_token_id
        vision_start_token_id = self.config.vision_start_token_id
        audio_start_token_id = self.config.audio_start_token_id
        position_id_per_seconds = self.config.position_id_per_seconds
        seconds_per_chunk = self.config.seconds_per_chunk

        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 not None:
                attention_mask = attention_mask == 1
            position_ids = torch.ones(
                3,
                input_ids.shape[0],
                input_ids.shape[1],
                dtype=input_ids.dtype,
                device=input_ids.device,
            )
            image_idx, video_idx, audio_idx = 0, 0, 0
            for i, input_ids in enumerate(total_input_ids):
                if attention_mask is not None:
                    input_ids = input_ids[attention_mask[i]]
                image_nums, video_nums, audio_nums = 0, 0, 0
                vision_start_indices = torch.argwhere(input_ids == vision_start_token_id).squeeze(1)
                vision_tokens = input_ids[vision_start_indices + 1]
                audio_nums = torch.sum(input_ids == audio_start_token_id)
                image_nums = (vision_tokens == image_token_id).sum()
                video_nums = (
                    (vision_tokens == audio_start_token_id).sum()
                    if use_audio_in_video
                    else (vision_tokens == video_token_id).sum()
                )
                input_tokens = input_ids.tolist()
                llm_pos_ids_list: list = []
                st = 0
                remain_images, remain_videos, remain_audios = image_nums, video_nums, audio_nums
                multimodal_nums = (
                    image_nums + audio_nums if use_audio_in_video else image_nums + video_nums + audio_nums
                )
                for _ in range(multimodal_nums):
                    st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0
                    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 audio_token_id in input_tokens and remain_audios > 0:
                        ed_audio = input_tokens.index(audio_token_id, st)
                    else:
                        ed_audio = len(input_tokens) + 1
                    min_ed = min(ed_image, ed_video, ed_audio)
                    if min_ed == ed_audio:
                        text_len = min_ed - st - 1
                        if text_len != 0:
                            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)

                        st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0
                        bos_len = 1
                        llm_pos_ids_list.append(torch.arange(bos_len).view(1, -1).expand(3, -1) + st_idx)

                        st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0
                        audio_len = ((audio_seqlens[audio_idx] - 1) // 2 + 1 - 2) // 2 + 1
                        llm_pos_ids = torch.arange(audio_len).view(1, -1).expand(3, -1) + st_idx
                        llm_pos_ids_list.append(llm_pos_ids)

                        st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0
                        eos_len = 1
                        llm_pos_ids_list.append(torch.arange(eos_len).view(1, -1).expand(3, -1) + st_idx)

                        st += text_len + bos_len + audio_len + eos_len
                        audio_idx += 1
                        remain_audios -= 1

                    elif min_ed == ed_image:
                        text_len = min_ed - st - 1
                        if text_len != 0:
                            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)

                        st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0
                        bos_len = 1
                        llm_pos_ids_list.append(torch.arange(bos_len).view(1, -1).expand(3, -1) + st_idx)

                        st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0
                        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 * position_id_per_seconds).long()
                        llm_pos_ids = self.get_llm_pos_ids_for_vision(
                            st_idx, image_idx, spatial_merge_size, t_index, grid_hs, grid_ws
                        )
                        image_len = image_grid_thw[image_idx].prod() // (spatial_merge_size**2)
                        llm_pos_ids_list.append(llm_pos_ids)

                        st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0
                        eos_len = 1
                        llm_pos_ids_list.append(torch.arange(eos_len).view(1, -1).expand(3, -1) + st_idx)

                        st += text_len + bos_len + image_len + eos_len
                        image_idx += 1
                        remain_images -= 1

                    elif min_ed == ed_video and not use_audio_in_video:
                        text_len = min_ed - st - 1
                        if text_len != 0:
                            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)

                        st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0
                        bos_len = 1
                        llm_pos_ids_list.append(torch.arange(bos_len).view(1, -1).expand(3, -1) + st_idx)

                        st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0
                        grid_t = video_grid_thw[video_idx][0]
                        grid_hs = video_grid_thw[:, 1]
                        grid_ws = video_grid_thw[:, 2]
                        t_index = (
                            torch.arange(grid_t) * second_per_grids[video_idx].cpu().float() * position_id_per_seconds
                        ).long()
                        llm_pos_ids = self.get_llm_pos_ids_for_vision(
                            st_idx, video_idx, spatial_merge_size, t_index, grid_hs, grid_ws
                        )
                        video_len = video_grid_thw[video_idx].prod() // (spatial_merge_size**2)
                        llm_pos_ids_list.append(llm_pos_ids)

                        st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0
                        eos_len = 1
                        llm_pos_ids_list.append(torch.arange(eos_len).view(1, -1).expand(3, -1) + st_idx)

                        st += text_len + bos_len + video_len + eos_len
                        video_idx += 1
                        remain_videos -= 1

                    elif min_ed == ed_video and use_audio_in_video:
                        text_len = min_ed - st - 2
                        if text_len != 0:
                            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)

                        st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0
                        bos_len = 1
                        llm_pos_ids_list.append(torch.arange(bos_len).view(1, -1).expand(3, -1) + st_idx)
                        llm_pos_ids_list.append(torch.arange(bos_len).view(1, -1).expand(3, -1) + st_idx)

                        st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0
                        audio_len = ((audio_seqlens[audio_idx] - 1) // 2 + 1 - 2) // 2 + 1
                        audio_llm_pos_ids = torch.arange(audio_len).view(1, -1).expand(3, -1) + st_idx
                        grid_t = video_grid_thw[video_idx][0]
                        grid_hs = video_grid_thw[:, 1]
                        grid_ws = video_grid_thw[:, 2]

                        t_index = (
                            torch.arange(grid_t) * second_per_grids[video_idx].cpu().float() * position_id_per_seconds
                        ).long()
                        video_llm_pos_ids = self.get_llm_pos_ids_for_vision(
                            st_idx, video_idx, spatial_merge_size, t_index, grid_hs, grid_ws
                        )

                        t_ntoken_per_chunk = int(position_id_per_seconds * seconds_per_chunk)
                        video_chunk_indexes = self.get_chunked_index(video_llm_pos_ids[0], t_ntoken_per_chunk, st_idx)
                        audio_chunk_indexes = self.get_chunked_index(audio_llm_pos_ids[0], t_ntoken_per_chunk, st_idx)
                        sub_len = 0
                        for j in range(max(len(video_chunk_indexes), len(audio_chunk_indexes))):
                            video_chunk_index = video_chunk_indexes[j] if j < len(video_chunk_indexes) else None
                            audio_chunk_index = audio_chunk_indexes[j] if j < len(audio_chunk_indexes) else None
                            if video_chunk_index is not None:
                                sub_len += video_chunk_index[1] - video_chunk_index[0]

                                llm_pos_ids_list.append(
                                    video_llm_pos_ids[:, video_chunk_index[0] : video_chunk_index[1]]
                                )
                            if audio_chunk_index is not None:
                                sub_len += audio_chunk_index[1] - audio_chunk_index[0]

                                llm_pos_ids_list.append(
                                    audio_llm_pos_ids[:, audio_chunk_index[0] : audio_chunk_index[1]]
                                )
                        video_len = video_grid_thw[video_idx].prod() // (spatial_merge_size**2)

                        st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0
                        eos_len = 1
                        llm_pos_ids_list.append(torch.arange(eos_len).view(1, -1).expand(3, -1) + st_idx)
                        llm_pos_ids_list.append(torch.arange(eos_len).view(1, -1).expand(3, -1) + st_idx)

                        st += text_len + bos_len * 2 + audio_len + video_len + eos_len * 2

                        audio_idx += 1
                        video_idx += 1
                        remain_videos -= 1
                        remain_audios -= 1

                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)

                if attention_mask is not None:
                    position_ids[..., i, attention_mask[i]] = llm_positions.to(position_ids.device)
                else:
                    position_ids[..., i, :] = llm_positions.to(position_ids.device)
                mrope_position_deltas.append(llm_positions.max() + 1 - len(input_ids))
            mrope_position_deltas = torch.tensor(mrope_position_deltas).unsqueeze(1).to(device=input_ids.device)

            return position_ids, mrope_position_deltas
        else:
            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 - torch.sum(attention_mask, dim=-1, keepdim=True)

            return position_ids, mrope_position_deltas


############################
#      Start Thinker       #
############################


@dataclass
@auto_docstring(
    custom_intro="""
    Base class for Qwen2.5OmniThinker causal language model (or autoregressive) outputs.
    """
)
class Qwen2_5OmniThinkerCausalLMOutputWithPast(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)`, *optional*):
        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


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 Qwen2_5OmniAudioAttention(nn.Module):
    """Multi-headed attention from 'Attention Is All You Need' paper"""

    def __init__(
        self,
        config: Qwen2_5OmniAudioEncoderConfig,
    ):
        super().__init__()
        self.embed_dim = config.d_model
        self.num_heads = config.encoder_attention_heads
        self.dropout = config.attention_dropout
        self.head_dim = self.embed_dim // self.num_heads
        self.num_key_value_groups = 1  # needed for eager attention
        self.config = config

        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}"
                f" and `num_heads`: {self.num_heads})."
            )
        self.scaling = self.head_dim**-0.5
        self.attention_dropout = 0.0
        self.is_decoder = False
        self.is_causal = False

        self.k_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=False)
        self.v_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=True)
        self.q_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=True)
        self.out_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=True)

    def forward(
        self,
        hidden_states: torch.Tensor,
        cu_seqlens: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        **kwargs,
    ) -> tuple[torch.Tensor, Optional[torch.Tensor], Optional[tuple[torch.Tensor]]]:
        """Input shape: Batch x Time x Channel"""

        seq_length, _ = hidden_states.size()

        query_states = self.q_proj(hidden_states).reshape(seq_length, self.num_heads, -1)
        key_states = self.k_proj(hidden_states).reshape(seq_length, self.num_heads, -1)
        value_states = self.v_proj(hidden_states).reshape(seq_length, self.num_heads, -1)

        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()

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

        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.out_proj(attn_output)

        return attn_output


class Qwen2_5OmniAudioEncoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: Qwen2_5OmniAudioEncoderConfig):
        super().__init__()
        self.embed_dim = config.d_model
        self.self_attn = Qwen2_5OmniAudioAttention(config)
        self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
        self.dropout = config.dropout
        self.activation_fn = ACT2FN[config.activation_function]
        self.activation_dropout = config.activation_dropout
        self.fc1 = nn.Linear(self.embed_dim, config.encoder_ffn_dim)
        self.fc2 = nn.Linear(config.encoder_ffn_dim, self.embed_dim)
        self.final_layer_norm = nn.LayerNorm(self.embed_dim)

    def forward(
        self,
        hidden_states: torch.Tensor,
        cu_seqlens: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        **kwargs,
    ) -> torch.Tensor:
        """
        Args:
            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
            attention_mask (`torch.FloatTensor`): attention mask of size
                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
            layer_head_mask (`torch.FloatTensor`): mask for attention heads in a given layer of size
                `(encoder_attention_heads,)`.
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
        """
        residual = hidden_states
        hidden_states = self.self_attn_layer_norm(hidden_states)
        hidden_states = self.self_attn(
            hidden_states=hidden_states,
            cu_seqlens=cu_seqlens,
            attention_mask=attention_mask,
            **kwargs,
        )
        hidden_states = residual + hidden_states
        residual = hidden_states
        hidden_states = self.final_layer_norm(hidden_states)
        hidden_states = self.fc1(hidden_states)
        hidden_states = self.activation_fn(hidden_states)
        hidden_states = self.fc2(hidden_states)
        hidden_states = residual + hidden_states

        if hidden_states.dtype == torch.float16:
            clamp_value = torch.finfo(hidden_states.dtype).max - 1000
            hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)

        outputs = (hidden_states,)

        return outputs


class SinusoidsPositionEmbedding(nn.Module):
    def __init__(self, length, channels, max_timescale=10000):
        super().__init__()
        if channels % 2 != 0:
            raise ValueError("SinusoidsPositionEmbedding needs even channels input")
        log_timescale_increment = np.log(max_timescale) / (channels // 2 - 1)
        inv_timescales = torch.exp(-log_timescale_increment * torch.arange(channels // 2).float())
        scaled_time = torch.arange(length)[:, np.newaxis] * inv_timescales[np.newaxis, :]
        self.register_buffer(
            "positional_embedding",
            torch.cat([torch.sin(scaled_time), torch.cos(scaled_time)], dim=1),
            persistent=False,
        )

    def forward(self, seqlen: int):
        return self.positional_embedding[:seqlen, :]


@auto_docstring(
    custom_intro="""
    Transformer encoder consisting of *config.encoder_layers* self attention layers. Each layer is a
    [`Qwen2_5OmniAudioEncoderLayer`].
    """
)
class Qwen2_5OmniAudioEncoder(Qwen2_5OmniPreTrainedModel):
    config: Qwen2_5OmniAudioEncoderConfig
    main_input_name = "input_features"
    _no_split_modules = ["Qwen2_5OmniAudioEncoderLayer"]
    _supports_sdpa = True

    def __init__(self, config: Qwen2_5OmniAudioEncoderConfig):
        super().__init__(config)
        self.dropout = config.dropout

        embed_dim = config.d_model
        self.num_mel_bins = config.num_mel_bins
        self.max_source_positions = config.max_source_positions
        self.embed_scale = math.sqrt(embed_dim) if config.scale_embedding else 1.0
        self.n_window = config.n_window
        self.conv1 = nn.Conv1d(self.num_mel_bins, embed_dim, kernel_size=3, padding=1)
        self.conv2 = nn.Conv1d(embed_dim, embed_dim, kernel_size=3, stride=2, padding=1)
        self.positional_embedding = SinusoidsPositionEmbedding(self.max_source_positions, embed_dim)
        self.audio_bos_eos_token = nn.Embedding(2, config.output_dim)
        self.layers = nn.ModuleList([Qwen2_5OmniAudioEncoderLayer(config) for _ in range(config.encoder_layers)])
        self.ln_post = nn.LayerNorm(config.d_model)
        self.avg_pooler = nn.AvgPool1d(2, stride=2)
        self.proj = nn.Linear(config.d_model, config.output_dim)
        self.gradient_checkpointing = False
        # Initialize weights and apply final processing
        self.post_init()

    def _freeze_parameters(self):
        for param in self.parameters():
            param.requires_grad = False
        self._requires_grad = False

    def get_input_embeddings(self) -> nn.Module:
        return self.conv1

    def set_input_embeddings(self, value: nn.Module):
        self.conv1 = value

    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

    @auto_docstring
    def forward(
        self,
        input_features,
        feature_lens=None,
        aftercnn_lens=None,
        **kwargs,
    ):
        r"""
        feature_lens (`torch.LongTensor` of shape `(batch_size,)`):
            mel length
        aftercnn_lens (`torch.LongTensor` of shape `(batch_size,)`):
            mel length after cnn
        """
        chunk_num = torch.ceil(feature_lens / (self.n_window * 2)).long()

        chunk_lengths = torch.tensor(
            [self.n_window * 2] * chunk_num.sum(),
            dtype=torch.long,
            device=feature_lens.device,
        )
        tail_chunk_index = F.pad(chunk_num, (1, 0), value=-1).cumsum(0)[1:]
        chunk_lengths[tail_chunk_index] = feature_lens % (self.n_window * 2)
        chunk_lengths = torch.where(chunk_lengths == 0, self.n_window * 2, chunk_lengths)

        chunk_list = input_features.split(chunk_lengths.tolist(), dim=1)
        padded_feature, padded_mask, padded_mask_after_cnn = self.padded_and_mask_function(
            chunk_list, chunk_lengths, padding_value=0, padding_side="right"
        )
        padded_embed = nn.functional.gelu(self.conv1(padded_feature)) * padded_mask
        padded_embed = nn.functional.gelu(self.conv2(padded_embed)).transpose(1, 2)

        padded_embed = padded_embed + self.positional_embedding.positional_embedding[
            : padded_embed.shape[1], :
        ].unsqueeze(0).to(padded_embed.dtype)
        hidden_states = padded_embed[padded_mask_after_cnn]
        cu_seqlens = torch.cat(
            (
                torch.zeros(1, device=padded_mask_after_cnn.device, dtype=torch.int32),
                padded_mask_after_cnn.sum(1).cumsum(0),
            )
        ).to(torch.int32)
        attention_mask = self._prepare_attention_mask(hidden_states, cu_seqlens)

        for encoder_layer in self.layers:
            layer_outputs = encoder_layer(
                hidden_states,
                cu_seqlens=cu_seqlens,
                attention_mask=attention_mask,
                **kwargs,
            )
            hidden_states = layer_outputs[0]

        hidden_states_list = hidden_states.split(aftercnn_lens.tolist(), dim=0)
        token_audio_list = []
        for each_audio_states in hidden_states_list:
            each_audio_states = self.avg_pooler(each_audio_states.transpose(0, 1)).transpose_(0, 1)
            each_audio_states = self.ln_post(each_audio_states)
            each_audio_states = self.proj(each_audio_states)
            token_audio_list.append(each_audio_states)
        token_audio = torch.cat(token_audio_list, dim=0)
        return BaseModelOutput(last_hidden_state=token_audio)

    def padded_and_mask_function(self, tensor_list, tensor_len, padding_value=0, padding_side="right"):
        """
        Pads a sequence of tensors to their maximum length on indicated `padding_side`.
        Then prepares a mask so that pad tokens are not attended to.
        """
        max_len = tensor_len.max()
        dim = tensor_list[0].shape[0]
        padded_tensor = torch.full(
            size=(len(tensor_list), dim, max_len),
            fill_value=padding_value,
            dtype=self.dtype,
            device=tensor_list[0].device,
        )

        batch_mask = torch.zeros(
            (len(tensor_len), max_len),
            dtype=torch.long,
            device=padded_tensor.device,
        )
        for i, length in enumerate(tensor_len):
            batch_mask[i, :length] = 1
            padded_tensor[i, :, :length] = tensor_list[i]

        feature_lens_after_cnn = (tensor_len - 1) // 2 + 1
        max_len_after_cnn = feature_lens_after_cnn.max()
        batch_mask_after_cnn = torch.zeros(
            (len(tensor_len), max_len_after_cnn),
            dtype=torch.long,
            device=padded_tensor.device,
        )
        for i, length in enumerate(feature_lens_after_cnn):
            batch_mask_after_cnn[i, :length] = 1
        return (
            padded_tensor,
            batch_mask.unsqueeze(1),
            batch_mask_after_cnn.bool(),
        )

    # Ignore copy
    def _get_feat_extract_output_lengths(self, input_lengths: torch.LongTensor):
        """
        Computes the output length of the convolutional layers and the output length of the audio encoder
        """
        input_lengths = (input_lengths - 1) // 2 + 1
        output_lengths = (input_lengths - 2) // 2 + 1
        return input_lengths, output_lengths


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(tensor: torch.Tensor, freqs: torch.Tensor) -> torch.Tensor:
    orig_dtype = tensor.dtype
    tensor = tensor.float()
    cos = freqs.cos()
    sin = freqs.sin()
    cos = cos.unsqueeze(1).repeat(1, 1, 2).unsqueeze(0).float()
    sin = sin.unsqueeze(1).repeat(1, 1, 2).unsqueeze(0).float()
    output = (tensor * cos) + (rotate_half(tensor) * sin)
    output = output.to(orig_dtype)
    return output


class Qwen2_5OmniVisionAttention(nn.Module):
    def __init__(self, config: Qwen2_5OmniVisionEncoderConfig = None) -> None:
        super().__init__()
        self.dim = config.hidden_size
        self.num_heads = config.num_heads
        self.head_dim = self.dim // self.num_heads
        self.q = nn.Linear(self.dim, self.dim, bias=True)
        self.k = nn.Linear(self.dim, self.dim, bias=True)
        self.v = nn.Linear(self.dim, self.dim, bias=True)
        self.proj = nn.Linear(self.dim, self.dim)
        self.scaling = self.head_dim**-0.5
        self.num_key_value_groups = 1  # needed for eager attention
        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,
        **kwargs,
    ) -> torch.Tensor:
        seq_length = hidden_states.shape[0]
        query_states = self.q(hidden_states).reshape(seq_length, self.num_heads, -1)
        key_states = self.k(hidden_states).reshape(seq_length, self.num_heads, -1)
        value_states = self.v(hidden_states).reshape(seq_length, self.num_heads, -1)
        query_states = apply_rotary_pos_emb_vision(query_states.unsqueeze(0), rotary_pos_emb).squeeze(0)
        key_states = apply_rotary_pos_emb_vision(key_states.unsqueeze(0), rotary_pos_emb).squeeze(0)

        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, _ = 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 = [
                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,
                )[0]
                for q, k, v in zip(*splits)
            ]
            attn_output = torch.cat(attn_outputs, dim=1)

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


class Qwen2_5OmniMLP(nn.Module):
    def __init__(self, config, bias: bool = False):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size
        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=bias)
        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=bias)
        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=bias)
        self.act_fn = ACT2FN[config.hidden_act]

    def forward(self, hidden_state):
        return self.down_proj(self.act_fn(self.gate_proj(hidden_state)) * self.up_proj(hidden_state))


class Qwen2_5OmniVisionBlock(GradientCheckpointingLayer):
    def __init__(self, config: Qwen2_5OmniVisionEncoderConfig) -> None:
        super().__init__()
        self.norm1 = Qwen2RMSNorm(config.hidden_size, eps=1e-6)
        self.norm2 = Qwen2RMSNorm(config.hidden_size, eps=1e-6)
        self.attn = Qwen2_5OmniVisionAttention(config=config)
        self.mlp = Qwen2_5OmniMLP(config, bias=True)

    def forward(
        self,
        hidden_states: torch.Tensor,
        cu_seqlens: torch.Tensor,
        rotary_pos_emb: 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,
            **kwargs,
        )
        hidden_states = hidden_states + self.mlp(self.norm2(hidden_states))
        return hidden_states


class Qwen2_5_VisionPatchEmbed(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)

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        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 Qwen2_5_VisionRotaryEmbedding(nn.Module):
    inv_freq: torch.Tensor  # fix linting for `register_buffer`

    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 Qwen2_5OmniPatchMerger(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 = Qwen2RMSNorm(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


class Qwen2_5OmniVisionEncoder(Qwen2_5OmniPreTrainedModel):
    config: Qwen2_5OmniVisionEncoderConfig
    _no_split_modules = ["Qwen2_5OmniVisionBlock"]

    def __init__(self, config: Qwen2_5OmniVisionEncoderConfig, *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 = Qwen2_5_VisionPatchEmbed(
            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 = Qwen2_5_VisionRotaryEmbedding(head_dim // 2)
        self.blocks = nn.ModuleList([Qwen2_5OmniVisionBlock(config) for _ in range(config.depth)])
        self.merger = Qwen2_5OmniPatchMerger(
            dim=config.out_hidden_size,
            context_dim=config.hidden_size,
            spatial_merge_size=config.spatial_merge_size,
        )
        self.gradient_checkpointing = False

    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 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)

        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)

        # Modification here
        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

            hidden_states = blk(
                hidden_states,
                cu_seqlens=cu_seqlens_now,
                rotary_pos_emb=rotary_pos_emb,
                **kwargs,
            )
        hidden_states = self.merger(hidden_states)
        reverse_indices = torch.argsort(window_index)
        hidden_states = hidden_states[reverse_indices, :]

        return hidden_states


class Qwen2_5OmniRotaryEmbedding(nn.Module):
    inv_freq: torch.Tensor  # fix linting for `register_buffer`

    def __init__(self, config: Qwen2_5OmniThinkerConfig, 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"))
        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

    @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, Qwen2_5Omni 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)
            cos = emb.cos() * self.attention_scaling
            sin = emb.sin() * self.attention_scaling

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


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 Qwen2_5OmniAttention(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: Qwen2_5OmniConfig, 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 = config.attention_dropout
        self.rope_scaling = config.rope_scaling
        self.scaling = self.head_dim**-0.5

        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 = Qwen2_5OmniRotaryEmbedding(config=config)

    @deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")
    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,  # 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_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


class Qwen2MLP(nn.Module):
    def __init__(self, config, bias: bool = False):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size
        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=bias)
        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=bias)
        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=bias)
        self.act_fn = ACT2FN[config.hidden_act]

    def forward(self, hidden_state):
        return self.down_proj(self.act_fn(self.gate_proj(hidden_state)) * self.up_proj(hidden_state))


class Qwen2_5OmniDecoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: Qwen2_5OmniTextConfig, 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 = Qwen2_5OmniAttention(config, layer_idx)

        self.mlp = Qwen2MLP(config)
        self.input_layernorm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.attention_type = config.layer_types[layer_idx]

    @deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")
    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: 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_values (`Cache`, *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 = self.self_attn(
            hidden_states=hidden_states,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            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,)

        return outputs


@auto_docstring
class Qwen2_5OmniThinkerTextModel(Qwen2_5OmniPreTrainedModel):
    config: Qwen2_5OmniTextConfig
    _no_split_modules = ["Qwen2_5OmniDecoderLayer"]

    def __init__(self, config: Qwen2_5OmniTextConfig):
        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(
            [Qwen2_5OmniDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self._attn_implementation = config._attn_implementation
        self.norm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.rotary_emb = Qwen2_5OmniRotaryEmbedding(config=config)
        self.has_sliding_layers = "sliding_attention" in self.config.layer_types

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

    @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,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> Union[tuple, BaseModelOutputWithPast]:
        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
        )
        use_cache = use_cache if use_cache is not None else self.config.use_cache

        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        if (input_ids is None) ^ (inputs_embeds is not None):
            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")

        if self.gradient_checkpointing and self.training:
            if use_cache:
                logger.warning_once(
                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                )
                use_cache = False

        # torch.jit.trace() doesn't support cache objects in the output
        if use_cache and past_key_values is None and not torch.jit.is_tracing():
            past_key_values = DynamicCache(config=self.config)

        if inputs_embeds is None:
            inputs_embeds = self.embed_tokens(input_ids)

        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.arange(
                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
            )

        # the hard coded `3` is for temporal, height and width.
        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)

        # NOTE: we need to pass text position ids for packing. Qwen2-VL uses 3D positions
        # where each dim indicates visual spatial positions for temporal/height/width grids.
        # There are two scenarios when FA2-like packed masking might be activated.
        # 1. User specifically passed packed `position_ids` and no attention mask.
        #    In this case we expect the useer to create correct position ids for all 3 grids
        #    and prepend text-only position ids to it. The final tensor will be [4, bs, seq-len]
        # 2. User runs forward with no attention mask and no position ids. In this case, position ids
        #    are prepared by the model (`get_rope_index`) as `[4, bs, seq-len]` tensor. Text-only positions are
        #    prepended by us when creating positions so that the mask is constructed correctly. NOTE: failing to pass
        #    text-only positions will cause incorrect mask construction, do not change `prepare_input_for_generation`
        if position_ids.ndim == 3 and position_ids.shape[0] == 4:
            text_position_ids = position_ids[0]
            position_ids = position_ids[1:]
        else:
            # If inputs are not packed (usual 3D positions), do not prepare mask from position_ids
            text_position_ids = None

        # It may already have been prepared by e.g. `generate`
        if not isinstance(causal_mask_mapping := attention_mask, dict):
            # Prepare mask arguments
            mask_kwargs = {
                "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,
            }
            # Create the masks
            causal_mask_mapping = {
                "full_attention": create_causal_mask(**mask_kwargs),
            }
            # The sliding window alternating layers are not always activated depending on the config
            if self.has_sliding_layers:
                causal_mask_mapping["sliding_attention"] = create_sliding_window_causal_mask(**mask_kwargs)

        hidden_states = inputs_embeds

        # create position embeddings to be shared across the decoder layers
        position_embeddings = self.rotary_emb(hidden_states, position_ids)

        # decoder layers
        all_hidden_states = () if output_hidden_states else None
        all_self_attns = () if output_attentions else None

        for decoder_layer in self.layers:
            if output_hidden_states:
                all_hidden_states += (hidden_states,)

            layer_outputs = decoder_layer(
                hidden_states,
                attention_mask=causal_mask_mapping[decoder_layer.attention_type],
                position_ids=text_position_ids,
                past_key_values=past_key_values,
                output_attentions=output_attentions,
                use_cache=use_cache,
                cache_position=cache_position,
                position_embeddings=position_embeddings,
                **kwargs,
            )

            hidden_states = layer_outputs[0]

            if output_attentions:
                all_self_attns += (layer_outputs[1],)

        hidden_states = self.norm(hidden_states)

        # add hidden states from the last decoder layer
        if output_hidden_states:
            all_hidden_states += (hidden_states,)

        if not return_dict:
            return tuple(
                v for v in [hidden_states, past_key_values, all_hidden_states, all_self_attns] if v is not None
            )
        return BaseModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=past_key_values,
            hidden_states=all_hidden_states,
            attentions=all_self_attns,
        )


@auto_docstring(
    custom_intro="""
    The Qwen2.5OmniThinker model which consists of a audio backbone and a language model.
    """
)
class Qwen2_5OmniThinkerForConditionalGeneration(Qwen2_5OmniPreTrainedModelForConditionalGeneration, GenerationMixin):
    config: Qwen2_5OmniThinkerConfig
    base_model_prefix = "thinker"
    _tied_weights_keys = ["model.embed_tokens.weight", "lm_head.weight"]
    _no_split_modules = ["Qwen2_5OmniAudioEncoder", "Qwen2_5OmniVisionEncoder"]

    def __init__(self, config: Qwen2_5OmniThinkerConfig):
        super().__init__(config)
        self.audio_tower = Qwen2_5OmniAudioEncoder._from_config(config.audio_config)
        self.visual = Qwen2_5OmniVisionEncoder._from_config(config.vision_config)
        self.vocab_size = config.text_config.vocab_size
        self.model = Qwen2_5OmniThinkerTextModel._from_config(config.text_config)
        self.lm_head = nn.Linear(config.text_config.hidden_size, config.text_config.vocab_size, bias=False)
        self.pad_token_id = self.config.pad_token_id if self.config.pad_token_id is not None else -1
        self.spatial_merge_size = config.vision_config.spatial_merge_size
        self.rope_deltas = None
        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
    ):
        """
        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)
        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)
        image_embeds = self.visual(pixel_values, grid_thw=image_grid_thw)
        return image_embeds

    def get_audio_features(
        self,
        input_features: torch.FloatTensor,
        feature_attention_mask: Optional[torch.LongTensor] = None,
        audio_feature_lengths: Optional[torch.LongTensor] = None,
    ):
        """
        Encodes audios into continuous embeddings that can be forwarded to the language model.

        Args:
            input_features (`torch.FloatTensor`):
                The tensors corresponding to the input audios.
            feature_attention_mask (`torch.LongTensor`, *optional*):
                Mask to avoid performing attention on padding feature indices. Mask values selected in `[0, 1]`:
            audio_feature_lengths (`torch.LongTensor` of shape `(num_audios)`, *optional*):
                The length of feature shape of each audio in LLM.
        """
        if feature_attention_mask is not None:
            audio_feature_lengths = torch.sum(feature_attention_mask, dim=1)
            input_features = input_features.permute(0, 2, 1)[feature_attention_mask.bool()].permute(1, 0)
        else:
            audio_feature_lengths = None

        audio_feat_lengths, audio_output_lengths = self.audio_tower._get_feat_extract_output_lengths(
            audio_feature_lengths if audio_feature_lengths is not None else feature_attention_mask.sum(-1)
        )
        feature_lens = audio_feature_lengths if audio_feature_lengths is not None else feature_attention_mask.sum(-1)
        audio_outputs = self.audio_tower(
            input_features,
            feature_lens=feature_lens,
            aftercnn_lens=audio_feat_lengths,
        )
        audio_features = audio_outputs.last_hidden_state

        if audio_features.shape[0] != sum(audio_output_lengths.tolist()):
            raise ValueError("length of audio_features should match audio_output_lengths")

        return audio_features

    def get_placeholder_mask(
        self,
        input_ids: torch.LongTensor,
        inputs_embeds: torch.FloatTensor,
        image_features: Optional[torch.FloatTensor] = None,
        video_features: Optional[torch.FloatTensor] = None,
    ):
        """
        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)
            special_video_mask = inputs_embeds == self.get_input_embeddings()(
                torch.tensor(self.config.video_token_id, dtype=torch.long, device=inputs_embeds.device)
            )
            special_video_mask = special_video_mask.all(-1)
            special_audio_mask = (
                inputs_embeds
                == self.get_input_embeddings()(
                    torch.tensor(self.config.audio_token_id, dtype=torch.long, device=inputs_embeds.device)
                )
            ).all(-1)
        else:
            special_image_mask = input_ids == self.config.image_token_id
            special_video_mask = input_ids == self.config.video_token_id
            special_audio_mask = input_ids == self.config.audio_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)
        if image_features is not None and 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 {image_features.shape[0]}"
            )

        n_video_tokens = special_video_mask.sum()
        special_video_mask = special_video_mask.unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device)
        if video_features is not None and inputs_embeds[special_video_mask].numel() != video_features.numel():
            raise ValueError(
                f"Videos features and image tokens do not match: tokens: {n_video_tokens}, features {video_features.shape[0]}"
            )

        special_audio_mask = special_audio_mask.unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device)
        return special_image_mask, special_video_mask, special_audio_mask

    @auto_docstring
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        input_features: Optional[torch.FloatTensor] = None,
        pixel_values: Optional[torch.FloatTensor] = None,
        pixel_values_videos: Optional[torch.FloatTensor] = None,
        image_grid_thw: Optional[torch.LongTensor] = None,
        video_grid_thw: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        feature_attention_mask: Optional[torch.Tensor] = None,
        audio_feature_lengths: Optional[torch.LongTensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Cache] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        rope_deltas: Optional[torch.LongTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        use_audio_in_video: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
        video_second_per_grid: Optional[torch.LongTensor] = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> Union[tuple, Qwen2_5OmniThinkerCausalLMOutputWithPast]:
        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.
        video_grid_thw (`torch.LongTensor` of shape `(num_videos, 3)`, *optional*):
            The temporal, height and width of feature shape of each video in LLM.
        feature_attention_mask (`torch.Tensor` of shape `(batch_size, feature_sequence_length)`, *optional*):
            Mask to avoid performing attention on padding feature indices. Mask values selected in `[0, 1]`:

            - 1 for tokens that are **not masked**,
            - 0 for tokens that are **masked**.
        audio_feature_lengths (`torch.LongTensor` of shape `(num_audios)`, *optional*):
            The length of feature shape of each audio in LLM.
        rope_deltas (`torch.LongTensor` of shape `(batch_size, )`, *optional*):
            The rope index difference between sequence length and multimodal rope.
        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]`.
        use_audio_in_video (`bool`, *optional*):
            Whether or not use audio track in video, should same as the parameter in `process_audio_info`.
        video_second_per_grid (`torch.LongTensor` of shape `(num_videos)`, *optional*):
            Number of seconds per grid for each video, used for temporal feature mapping.

        Example:

        ```python
        >>> from io import BytesIO
        >>> from urllib.request import urlopen
        >>> import librosa
        >>> from qwen_vl_utils import process_vision_info
        >>> from transformers import Qwen2_5OmniProcessor, Qwen2_5OmniThinkerForConditionalGeneration

        >>> thinker = Qwen2_5OmniThinkerForConditionalGeneration.from_pretrained("Qwen/Qwen2.5-Omni-7B")
        >>> processor = Qwen2_5OmniProcessor.from_pretrained("Qwen/Qwen2.5-Omni-7B")

        >>> conversations = [
        >>>         {'role': 'system', 'content': 'You are a helpful voice chat bot, and please respond to me in a casual conversation manner using random voice.'},
        >>>         {"role": "user", "content": [
        >>>             {"type": "image", "image_url": "https://www.ilankelman.org/stopsigns/australia.jpg"},
        >>>             {"type": "audio", "audio_url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen2-Audio/audio/glass-breaking-151256.mp3"},
        >>>         ]},
        >>> ]

        >>> text = processor.apply_chat_template(conversation, add_generation_prompt=True, tokenize=False)
        >>> audios = [ librosa.load(BytesIO(urlopen( conversations[1]['content'][1]['audio_url'] ).read()), sr=self.processor.feature_extractor.sampling_rate) ]
        >>> images, videos = process_vision_info(conversations)
        >>> inputs = processor(text=text, audios=audios, images=images, videos=videos, return_tensors="pt", padding=True)

        >>> # Generate
        >>> inputs['use_audio_in_video'] = `True` or `False`
        >>> generation = thinker.generate(**inputs, max_new_tokens=2048)
        >>> generate_ids = generation[:, inputs.input_ids.size(1):]

        >>> response = processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
        ```"""

        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:
            # 1. Extract the input embeddings
            inputs_embeds = self.get_input_embeddings()(input_ids)

        # 2. Merge text , audios , image and video
        if input_features is not None:
            audio_features = self.get_audio_features(
                input_features,
                feature_attention_mask=feature_attention_mask,
                audio_feature_lengths=audio_feature_lengths,
            )
            audio_features = audio_features.to(inputs_embeds.device, inputs_embeds.dtype)
            _, _, audio_mask = self.get_placeholder_mask(input_ids, inputs_embeds=inputs_embeds)
            inputs_embeds = inputs_embeds.masked_scatter(audio_mask, audio_features)

        if pixel_values is not None:
            image_embeds = self.get_image_features(pixel_values, image_grid_thw)
            image_embeds = image_embeds.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 pixel_values_videos is not None:
            video_embeds = self.get_video_features(pixel_values_videos, video_grid_thw)
            video_embeds = video_embeds.to(inputs_embeds.device, inputs_embeds.dtype)
            _, video_mask, _ = self.get_placeholder_mask(
                input_ids, inputs_embeds=inputs_embeds, video_features=video_embeds
            )
            inputs_embeds = inputs_embeds.masked_scatter(video_mask, video_embeds)

        if feature_attention_mask is not None:
            audio_feature_lengths = torch.sum(feature_attention_mask, dim=1)
        else:
            audio_feature_lengths = None

        if attention_mask is not None and position_ids is None:
            if (
                cache_position is None
                or (cache_position is not None and cache_position[0] == 0)
                or self.rope_deltas is None
            ):
                delta0 = (1 - attention_mask).sum(dim=-1).unsqueeze(1)
                position_ids, rope_deltas = self.get_rope_index(
                    input_ids,
                    image_grid_thw,
                    video_grid_thw,
                    attention_mask,
                    use_audio_in_video,
                    audio_feature_lengths,
                    video_second_per_grid,
                )
                rope_deltas = rope_deltas - delta0
                self.rope_deltas = rope_deltas
            else:
                batch_size, seq_length = input_ids.shape
                delta = cache_position[0] + self.rope_deltas if cache_position is not None else 0
                position_ids = torch.arange(seq_length, device=input_ids.device)
                position_ids = position_ids.view(1, -1).expand(batch_size, -1)
                position_ids = position_ids.add(delta)
                position_ids = position_ids.unsqueeze(0).expand(3, -1, -1)

        outputs = self.model(
            attention_mask=attention_mask,
            position_ids=position_ids,
            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=return_dict,
            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.get_text_config().vocab_size
            )

        if not return_dict:
            output = (logits,) + outputs
            return (loss,) + output if loss is not None else output

        return Qwen2_5OmniThinkerCausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
            rope_deltas=self.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,
        input_features=None,
        feature_attention_mask=None,
        use_audio_in_video=False,
        video_second_per_grid=None,
        **kwargs,
    ):
        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,
            use_cache=use_cache,
            pixel_values=pixel_values,
            pixel_values_videos=pixel_values_videos,
            image_grid_thw=image_grid_thw,
            video_grid_thw=video_grid_thw,
            input_features=input_features,
            feature_attention_mask=feature_attention_mask,
            use_audio_in_video=use_audio_in_video,
            video_second_per_grid=video_second_per_grid,
            **kwargs,
        )

        model_inputs["position_ids"] = None

        if cache_position[0] != 0:
            model_inputs["pixel_values"] = None
            model_inputs["pixel_values_videos"] = None
            model_inputs["input_features"] = None

        return model_inputs


############################
#       Start Talker       #
############################


@dataclass
@auto_docstring(
    custom_intro="""
    Base class for Qwen2.5OmniTalker causal language model (or autoregressive) outputs.
    """
)
class Qwen2_5OmniTalkerCausalLMOutputWithPast(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.
    thinker_reply_part (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
        Hidden states from the thinker model that are used as input for the talker model. These represent the encoded
        response that the talker model will use to generate speech tokens.
    """

    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
    thinker_reply_part: Optional[torch.FloatTensor] = None


@auto_docstring
class Qwen2_5OmniTalkerModel(Qwen2_5OmniPreTrainedModel):
    config: Qwen2_5OmniTalkerConfig
    _no_split_modules = ["Qwen2_5OmniTalkerDecoderLayer"]

    def __init__(self, config: Qwen2_5OmniTalkerConfig):
        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.embedding_size, self.padding_idx)
        self.layers = nn.ModuleList(
            [Qwen2_5OmniDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self._attn_implementation = config._attn_implementation
        self.norm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.rotary_emb = Qwen2_5OmniRotaryEmbedding(config=config)
        self.has_sliding_layers = "sliding_attention" in self.config.layer_types

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

    @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,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> Union[tuple, BaseModelOutputWithPast]:
        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
        )
        use_cache = use_cache if use_cache is not None else self.config.use_cache

        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        if (input_ids is None) ^ (inputs_embeds is not None):
            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")

        if self.gradient_checkpointing and self.training:
            if use_cache:
                logger.warning_once(
                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                )
                use_cache = False

        # torch.jit.trace() doesn't support cache objects in the output
        if use_cache and past_key_values is None and not torch.jit.is_tracing():
            past_key_values = DynamicCache(config=self.config)

        if inputs_embeds is None:
            inputs_embeds = self.embed_tokens(input_ids)

        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.arange(
                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
            )

        # the hard coded `3` is for temporal, height and width.
        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)

        # NOTE: we need to pass text position ids for packing. Qwen2-VL uses 3D positions
        # where each dim indicates visual spatial positions for temporal/height/width grids.
        # There are two scenarios when FA2-like packed masking might be activated.
        # 1. User specifically passed packed `position_ids` and no attention mask.
        #    In this case we expect the useer to create correct position ids for all 3 grids
        #    and prepend text-only position ids to it. The final tensor will be [4, bs, seq-len]
        # 2. User runs forward with no attention mask and no position ids. In this case, position ids
        #    are prepared by the model (`get_rope_index`) as `[4, bs, seq-len]` tensor. Text-only positions are
        #    prepended by us when creating positions so that the mask is constructed correctly. NOTE: failing to pass
        #    text-only positions will cause incorrect mask construction, do not change `prepare_input_for_generation`
        if position_ids.ndim == 3 and position_ids.shape[0] == 4:
            text_position_ids = position_ids[0]
            position_ids = position_ids[1:]
        else:
            # If inputs are not packed (usual 3D positions), do not prepare mask from position_ids
            text_position_ids = None

        # It may already have been prepared by e.g. `generate`
        if not isinstance(causal_mask_mapping := attention_mask, dict):
            # Prepare mask arguments
            mask_kwargs = {
                "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,
            }
            # Create the masks
            causal_mask_mapping = {
                "full_attention": create_causal_mask(**mask_kwargs),
            }
            # The sliding window alternating layers are not always activated depending on the config
            if self.has_sliding_layers:
                causal_mask_mapping["sliding_attention"] = create_sliding_window_causal_mask(**mask_kwargs)

        hidden_states = inputs_embeds

        # create position embeddings to be shared across the decoder layers
        position_embeddings = self.rotary_emb(hidden_states, position_ids)

        # decoder layers
        all_hidden_states = () if output_hidden_states else None
        all_self_attns = () if output_attentions else None

        for decoder_layer in self.layers:
            if output_hidden_states:
                all_hidden_states += (hidden_states,)

            layer_outputs = decoder_layer(
                hidden_states,
                attention_mask=causal_mask_mapping[decoder_layer.attention_type],
                position_ids=text_position_ids,
                past_key_values=past_key_values,
                output_attentions=output_attentions,
                use_cache=use_cache,
                cache_position=cache_position,
                position_embeddings=position_embeddings,
                **kwargs,
            )

            hidden_states = layer_outputs[0]

            if output_attentions:
                all_self_attns += (layer_outputs[1],)

        hidden_states = self.norm(hidden_states)

        # add hidden states from the last decoder layer
        if output_hidden_states:
            all_hidden_states += (hidden_states,)

        if not return_dict:
            return tuple(
                v for v in [hidden_states, past_key_values, all_hidden_states, all_self_attns] if v is not None
            )
        return BaseModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=past_key_values,
            hidden_states=all_hidden_states,
            attentions=all_self_attns,
        )


class Qwen2_5OmniTalkerForConditionalGeneration(Qwen2_5OmniPreTrainedModelForConditionalGeneration, GenerationMixin):
    config: Qwen2_5OmniTalkerConfig
    base_model_prefix = "talker"

    def __init__(self, config: Qwen2_5OmniTalkerConfig):
        super().__init__(config)

        self.thinker_to_talker_proj = nn.Linear(config.embedding_size, config.hidden_size)

        self.model = Qwen2_5OmniTalkerModel(config)
        self.codebook_size = config.vocab_size
        self.codec_head = nn.Linear(config.hidden_size, self.codebook_size, bias=False)

        self.codec_bos_token = config.tts_codec_start_token_id
        self.codec_eos_token = config.tts_codec_end_token_id
        self.codec_pad_token = config.tts_codec_pad_token_id
        self.codec_mask_token = config.tts_codec_mask_token_id

        self.text_bos_token = config.tts_text_start_token_id
        self.text_eos_token = config.tts_text_end_token_id
        self.text_pad_token = config.tts_text_pad_token_id

        self.spatial_merge_size = self.config.spatial_merge_size
        self.rope_deltas = None

        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)

    @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,
        thinker_reply_part: Optional[torch.FloatTensor] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        rope_deltas: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
        input_text_ids: Optional[torch.LongTensor] = None,
        image_grid_thw: Optional[torch.LongTensor] = None,
        video_grid_thw: Optional[torch.LongTensor] = None,
        use_audio_in_video: Optional[bool] = None,
        audio_feature_lengths: Optional[torch.LongTensor] = None,
        video_second_per_grid: Optional[torch.LongTensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[tuple, Qwen2_5OmniTalkerCausalLMOutputWithPast]:
        r"""
        thinker_reply_part (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
            Hidden states from the thinker model's output that represent the text reply part to be processed.
        rope_deltas (`torch.LongTensor` of shape `(batch_size, )`, *optional*):
            The rope index difference between sequence length and multimodal rope.
        input_text_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Input token IDs for text-only content, used for position calculation in multimodal contexts.
        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.
        use_audio_in_video (`bool`, *optional*):
            Whether or not use audio track in video, should same as the parameter in `process_audio_info`.
        audio_feature_lengths (`torch.LongTensor` of shape `(num_audios)`, *optional*):
            The length of feature shape of each audio in LLM.
        video_second_per_grid (`torch.LongTensor` of shape `(num_videos)`, *optional*):
            Number of seconds per grid for each video, used for temporal feature mapping.

        Example:

        ```python
        >>> from io import BytesIO
        >>> from urllib.request import urlopen
        >>> import librosa
        >>> from transformers import AutoProcessor, Qwen2_5OmniTalkerForConditionalGeneration

        >>> model = Qwen2_5OmniTalkerForConditionalGeneration.from_pretrained("Qwen/Qwen2-Audio-7B")
        >>> processor = AutoProcessor.from_pretrained("Qwen/Qwen2-Audio-7B")

        >>> prompt = "<|audio_bos|><|AUDIO|><|audio_eos|>Generate the caption in English:"
        >>> url = "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen2-Audio/audio/glass-breaking-151256.mp3"
        >>> audio, _ = librosa.load(BytesIO(urlopen(url).read()), sr=self.processor.feature_extractor.sampling_rate)

        >>> inputs = processor(text=prompt, audios=audio, return_tensors="pt")

        >>> # Generate
        >>> generate_ids = model.generate(**inputs, max_length=30)
        >>> processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
        "Generate the caption in English: Glass is breaking."
        ```"""

        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 attention_mask is not None and position_ids is None:
            if (
                cache_position is None
                or (cache_position is not None and cache_position[0] == 0)
                or self.rope_deltas is None
            ):
                position_ids, rope_deltas = self.get_rope_index(
                    input_text_ids,
                    image_grid_thw,
                    video_grid_thw,
                    attention_mask,
                    use_audio_in_video,
                    audio_feature_lengths,
                    video_second_per_grid,
                )

                inputs_embeds[:, -1, :] += self.get_input_embeddings()(
                    torch.tensor([self.codec_bos_token], dtype=torch.long, device=inputs_embeds.device)
                )
                inputs_embeds[:, -2, :] += self.get_input_embeddings()(
                    torch.tensor([self.codec_pad_token], dtype=torch.long, device=inputs_embeds.device)
                )
                self.rope_deltas = rope_deltas

            else:
                batch_size, seq_length = input_ids.shape
                delta = cache_position[0] + self.rope_deltas if cache_position is not None else 0
                position_ids = torch.arange(seq_length, device=input_ids.device)
                position_ids = position_ids.view(1, -1).expand(batch_size, -1)
                position_ids = position_ids.add(delta)
                position_ids = position_ids.unsqueeze(0).expand(3, -1, -1)

        if inputs_embeds is None:
            # 1. Inference tokens after second token
            codec_embeds = self.get_input_embeddings()(input_ids)
            inputs_embeds = codec_embeds + thinker_reply_part[:, :1, :]
            if thinker_reply_part.shape[1] > 1:
                thinker_reply_part = thinker_reply_part[:, 1:, :]

        talker_lm_input = self.thinker_to_talker_proj(inputs_embeds)

        if attention_mask is not None:
            attention_mask = attention_mask.to(inputs_embeds.device)

        outputs = self.model(
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=talker_lm_input,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        hidden_states = outputs[0]
        logits = self.codec_head(hidden_states)
        logits = logits.float()

        loss = None

        if not return_dict:
            output = (logits,) + outputs[1:]
            return (loss,) + output if loss is not None else output

        return Qwen2_5OmniTalkerCausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=hidden_states,
            attentions=outputs.attentions,
            rope_deltas=self.rope_deltas,
            thinker_reply_part=thinker_reply_part,
        )

    def _get_initial_cache_position(self, seq_length, device, model_kwargs):
        # Talker needs to calculate cache_position with input_ids, so pop inputs_embeds temporarily
        inputs_embeds = model_kwargs.pop("inputs_embeds")
        model_kwargs = super()._get_initial_cache_position(seq_length, device, model_kwargs)
        model_kwargs["inputs_embeds"] = inputs_embeds
        return model_kwargs

    # prepare inputs for talker lm generation
    def prepare_inputs_for_generation(
        self,
        input_ids,
        input_text_ids,
        past_key_values=None,
        attention_mask=None,
        inputs_embeds=None,
        thinker_reply_part=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,
        input_audio_features=None,
        audio_feature_attention_mask=None,
        audio_feature_lengths=None,
        use_audio_in_video=False,
        video_second_per_grid=None,
        **kwargs,
    ):
        model_inputs = super().prepare_inputs_for_generation(
            input_ids,
            past_key_values,
            attention_mask,
            inputs_embeds,
            cache_position,
            use_cache=use_cache,
            thinker_reply_part=thinker_reply_part,
            input_text_ids=input_text_ids,
            image_grid_thw=image_grid_thw,
            video_grid_thw=video_grid_thw,
            use_audio_in_video=use_audio_in_video,
            audio_feature_lengths=audio_feature_lengths,
            video_second_per_grid=video_second_per_grid,
            **kwargs,
        )

        model_inputs["position_ids"] = None

        return model_inputs

    def _update_model_kwargs_for_generation(
        self,
        outputs: ModelOutput,
        model_kwargs: dict[str, Any],
        is_encoder_decoder: bool = False,
        num_new_tokens: int = 1,
    ) -> dict[str, Any]:
        model_kwargs = super()._update_model_kwargs_for_generation(
            outputs, model_kwargs, is_encoder_decoder, num_new_tokens
        )

        if getattr(outputs, "thinker_reply_part", None) is not None:
            model_kwargs["thinker_reply_part"] = outputs.thinker_reply_part

        return model_kwargs


############################
#      Start Token2Wav     #
############################


# Using custom RoPE, will use LlamaRotaryEmbedding next version
class Qwen2_5OmniDiTRotaryEmbedding(nn.Module):
    inv_freq: torch.Tensor  # fix linting for `register_buffer`

    def __init__(self, dim, base=10000):
        super().__init__()

        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float() / dim))
        self.register_buffer("inv_freq", inv_freq)

    def forward(self, x):
        batch_size, seq_len = x.shape[0], x.shape[1]
        t = torch.arange(seq_len, device=x.device)
        device_type = x.device.type
        device_type = device_type if device_type != "mps" else "cpu"
        with torch.autocast(device_type=device_type, enabled=False):
            freqs = t.unsqueeze(1).float() @ self.inv_freq.unsqueeze(0).float()
            freqs = torch.stack((freqs, freqs), dim=-1)
            freqs = freqs.reshape(*freqs.shape[:-2], -1)
            freqs = freqs.repeat(batch_size, *([1] * freqs.dim()))
            cos = freqs.cos()
            sin = freqs.sin()

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


class TimeDelayNetBlock(nn.Module):
    def __init__(
        self,
        in_channels,
        out_channels,
        kernel_size,
        dilation,
    ):
        super().__init__()
        self.conv = nn.Conv1d(
            in_channels=in_channels,
            out_channels=out_channels,
            kernel_size=kernel_size,
            dilation=dilation,
            padding="same",
            padding_mode="reflect",
        )
        self.activation = nn.ReLU()

    def forward(self, hidden_states: torch.Tensor):
        return self.activation(self.conv(hidden_states))


class Res2NetBlock(torch.nn.Module):
    def __init__(self, in_channels, out_channels, scale=8, kernel_size=3, dilation=1):
        super().__init__()

        in_channel = in_channels // scale
        hidden_channel = out_channels // scale

        self.blocks = nn.ModuleList(
            [
                TimeDelayNetBlock(
                    in_channel,
                    hidden_channel,
                    kernel_size=kernel_size,
                    dilation=dilation,
                )
                for i in range(scale - 1)
            ]
        )
        self.scale = scale

    def forward(self, hidden_states):
        outputs = []
        for i, hidden_part in enumerate(torch.chunk(hidden_states, self.scale, dim=1)):
            if i == 0:
                output_part = hidden_part
            elif i == 1:
                output_part = self.blocks[i - 1](hidden_part)
            else:
                output_part = self.blocks[i - 1](hidden_part + output_part)
            outputs.append(output_part)
        output = torch.cat(outputs, dim=1)
        return output


class SqueezeExcitationBlock(nn.Module):
    def __init__(self, in_channels, se_channels, out_channels):
        super().__init__()

        self.conv1 = nn.Conv1d(
            in_channels=in_channels,
            out_channels=se_channels,
            kernel_size=1,
            padding="same",
            padding_mode="reflect",
        )
        self.relu = nn.ReLU(inplace=True)
        self.conv2 = nn.Conv1d(
            in_channels=se_channels,
            out_channels=out_channels,
            kernel_size=1,
            padding="same",
            padding_mode="reflect",
        )
        self.sigmoid = nn.Sigmoid()

    def forward(self, hidden_states):
        hidden_states_mean = hidden_states.mean(dim=2, keepdim=True)

        hidden_states_mean = self.relu(self.conv1(hidden_states_mean))
        hidden_states_mean = self.sigmoid(self.conv2(hidden_states_mean))

        return hidden_states * hidden_states_mean


class AttentiveStatisticsPooling(nn.Module):
    """This class implements an attentive statistic pooling layer for each channel.
    It returns the concatenated mean and std of the input tensor.
    """

    def __init__(self, channels, attention_channels=128):
        super().__init__()

        self.eps = 1e-12
        self.tdnn = TimeDelayNetBlock(channels * 3, attention_channels, 1, 1)
        self.tanh = nn.Tanh()
        self.conv = nn.Conv1d(
            in_channels=attention_channels,
            out_channels=channels,
            kernel_size=1,
            padding="same",
            padding_mode="reflect",
        )

    def _length_to_mask(self, length, max_len=None, dtype=None, device=None):
        """Creates a binary mask for each sequence.

        Reference: https://discuss.pytorch.org/t/how-to-generate-variable-length-mask/23397/3

        Arguments
        ---------
        length : torch.LongTensor
            Containing the length of each sequence in the batch. Must be 1D.
        max_len : int
            Max length for the mask, also the size of the second dimension.
        dtype : torch.dtype, default: None
            The dtype of the generated mask.
        device: torch.device, default: None
            The device to put the mask variable.

        Returns
        -------
        mask : tensor
            The binary mask.
        """

        if max_len is None:
            max_len = length.max().long().item()  # using arange to generate mask
        mask = torch.arange(max_len, device=length.device, dtype=length.dtype).expand(
            len(length), max_len
        ) < length.unsqueeze(1)

        mask = torch.as_tensor(mask, dtype=dtype, device=device)
        return mask

    def _compute_statistics(self, x, m, dim=2):
        mean = (m * x).sum(dim)
        std = torch.sqrt((m * (x - mean.unsqueeze(dim)).pow(2)).sum(dim).clamp(self.eps))
        return mean, std

    def forward(self, hidden_states):
        seq_length = hidden_states.shape[-1]
        lengths = torch.ones(hidden_states.shape[0], device=hidden_states.device)

        # Make binary mask of shape [N, 1, L]
        mask = self._length_to_mask(
            lengths * seq_length, max_len=seq_length, dtype=hidden_states.dtype, device=hidden_states.device
        )
        mask = mask.unsqueeze(1)

        # Expand the temporal context of the pooling layer by allowing the
        # self-attention to look at global properties of the utterance.
        total = mask.sum(dim=2, keepdim=True)

        mean, std = self._compute_statistics(hidden_states, mask / total)
        mean = mean.unsqueeze(2).repeat(1, 1, seq_length)
        std = std.unsqueeze(2).repeat(1, 1, seq_length)
        attention = torch.cat([hidden_states, mean, std], dim=1)

        # Apply layers
        attention = self.conv(self.tanh(self.tdnn(attention)))

        # Filter out zero-paddings
        attention = attention.masked_fill(mask == 0, float("-inf"))

        attention = F.softmax(attention, dim=2)
        mean, std = self._compute_statistics(hidden_states, attention)
        # Append mean and std of the batch
        pooled_stats = torch.cat((mean, std), dim=1)
        pooled_stats = pooled_stats.unsqueeze(2)

        return pooled_stats


class SqueezeExcitationRes2NetBlock(nn.Module):
    """An implementation of building block in ECAPA-TDNN, i.e.,
    TDNN-Res2Net-TDNN-SqueezeExcitationBlock.
    """

    def __init__(
        self,
        in_channels,
        out_channels,
        res2net_scale=8,
        se_channels=128,
        kernel_size=1,
        dilation=1,
    ):
        super().__init__()
        self.out_channels = out_channels
        self.tdnn1 = TimeDelayNetBlock(
            in_channels,
            out_channels,
            kernel_size=1,
            dilation=1,
        )
        self.res2net_block = Res2NetBlock(out_channels, out_channels, res2net_scale, kernel_size, dilation)
        self.tdnn2 = TimeDelayNetBlock(
            out_channels,
            out_channels,
            kernel_size=1,
            dilation=1,
        )
        self.se_block = SqueezeExcitationBlock(out_channels, se_channels, out_channels)

    def forward(self, hidden_state):
        residual = hidden_state

        hidden_state = self.tdnn1(hidden_state)
        hidden_state = self.res2net_block(hidden_state)
        hidden_state = self.tdnn2(hidden_state)
        hidden_state = self.se_block(hidden_state)

        return hidden_state + residual


class ECAPA_TimeDelayNet(torch.nn.Module):
    """An implementation of the speaker embedding model in a paper.
    "ECAPA-TDNN: Emphasized Channel Attention, Propagation and Aggregation in
    TDNN Based Speaker Verification" (https://huggingface.co/papers/2005.07143).
    """

    def __init__(self, config: Qwen2_5OmniDiTConfig):
        super().__init__()
        if len(config.enc_channels) != len(config.enc_kernel_sizes) or len(config.enc_channels) != len(
            config.enc_dilations
        ):
            raise ValueError("enc_channels, enc_kernel_sizes and enc_dilations should have same length")
        self.channels = config.enc_channels
        self.blocks = nn.ModuleList()

        # The initial TDNN layer
        self.blocks.append(
            TimeDelayNetBlock(
                config.mel_dim,
                config.enc_channels[0],
                config.enc_kernel_sizes[0],
                config.enc_dilations[0],
            )
        )

        # SE-Res2Net layers
        for i in range(1, len(config.enc_channels) - 1):
            self.blocks.append(
                SqueezeExcitationRes2NetBlock(
                    config.enc_channels[i - 1],
                    config.enc_channels[i],
                    res2net_scale=config.enc_res2net_scale,
                    se_channels=config.enc_se_channels,
                    kernel_size=config.enc_kernel_sizes[i],
                    dilation=config.enc_dilations[i],
                )
            )

        # Multi-layer feature aggregation
        self.mfa = TimeDelayNetBlock(
            config.enc_channels[-1],
            config.enc_channels[-1],
            config.enc_kernel_sizes[-1],
            config.enc_dilations[-1],
        )

        # Attentive Statistical Pooling
        self.asp = AttentiveStatisticsPooling(
            config.enc_channels[-1],
            attention_channels=config.enc_attention_channels,
        )

        # Final linear transformation
        self.fc = nn.Conv1d(
            in_channels=config.enc_channels[-1] * 2,
            out_channels=config.enc_dim,
            kernel_size=1,
            padding="same",
            padding_mode="reflect",
        )

    def forward(self, hidden_states):
        # Minimize transpose for efficiency
        hidden_states = hidden_states.transpose(1, 2)

        hidden_states_list = []
        for layer in self.blocks:
            hidden_states = layer(hidden_states)
            hidden_states_list.append(hidden_states)

        # Multi-layer feature aggregation
        hidden_states = torch.cat(hidden_states_list[1:], dim=1)
        hidden_states = self.mfa(hidden_states)

        # Attentive Statistical Pooling
        hidden_states = self.asp(hidden_states)

        # Final linear transformation
        hidden_states = self.fc(hidden_states)

        hidden_states = hidden_states.squeeze(-1)
        return hidden_states


class DiTInputEmbedding(nn.Module):
    def __init__(self, config: Qwen2_5OmniDiTConfig):
        super().__init__()
        self.proj = nn.Linear(
            config.mel_dim + config.enc_dim + config.enc_emb_dim + config.emb_dim,
            config.hidden_size,
        )
        self.spk_encoder = ECAPA_TimeDelayNet(config)

    def forward(
        self,
        hidden_states: torch.Tensor,
        speaker_embedding: torch.Tensor,
        condition_vector: torch.Tensor,
        code_embed: torch.Tensor,
        drop_audio_cond: Optional[bool] = False,
        code_embed_uncond: Optional[bool] = None,
        apply_cfg: Optional[bool] = True,
    ):
        if apply_cfg:
            hidden_states = torch.cat([hidden_states, hidden_states], dim=0)
            speaker_embedding = torch.cat([speaker_embedding, torch.zeros_like(speaker_embedding)], dim=0)
            condition_vector = torch.cat([condition_vector, torch.zeros_like(condition_vector)], dim=0)
            code_embed = torch.cat([code_embed, code_embed_uncond], dim=0)
        elif drop_audio_cond:  # cfg for cond audio
            condition_vector = torch.zeros_like(condition_vector)
            speaker_embedding = torch.zeros_like(speaker_embedding)
        condition_vector = self.spk_encoder(condition_vector).unsqueeze(1).repeat(1, hidden_states.size(1), 1)
        hidden_states = self.proj(torch.cat((hidden_states, condition_vector, code_embed, speaker_embedding), dim=-1))

        return hidden_states


# Transformer backbone using DiT blocks
class DiTCodecEmbedding(nn.Module):
    def __init__(self, codec_num_embeds, codec_dim, repeats):
        super().__init__()
        self.repeats = repeats
        self.codec_embed = nn.Embedding(codec_num_embeds + 1, codec_dim)

    def forward(self, code, drop_code=False):
        if drop_code:
            code = torch.zeros_like(code)
        code_embed = self.codec_embed(code)

        code_embed = torch.repeat_interleave(code_embed, repeats=self.repeats, dim=1)
        return code_embed


# AdaLayerNormZero
# return with modulated x for attn input, and params for later mlp modulation
class Qwen2_5_OmniAdaLayerNormZero(nn.Module):
    def __init__(self, dim):
        super().__init__()

        self.silu = nn.SiLU()
        self.linear = nn.Linear(dim, dim * 6)

        self.norm = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-6)

    def forward(self, hidden_states, emb=None):
        emb = self.linear(self.silu(emb))
        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = torch.chunk(emb, 6, dim=1)

        hidden_states = self.norm(hidden_states) * (1 + scale_msa[:, None]) + shift_msa[:, None]
        return hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp


# AdaLayerNormZero for final layer
# return only with modulated x for attn input, cuz no more mlp modulation
class Qwen2_5_OmniAdaLayerNormZero_Final(nn.Module):
    def __init__(self, dim):
        super().__init__()

        self.silu = nn.SiLU()
        self.linear = nn.Linear(dim, dim * 2)

        self.norm = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-6)

    def forward(self, hidden_states, emb):
        emb = self.linear(self.silu(emb))
        scale, shift = torch.chunk(emb, 2, dim=1)

        hidden_states = self.norm(hidden_states) * (1 + scale)[:, None, :] + shift[:, None, :]
        return hidden_states


# FeedForward
class DiTMLP(nn.Module):
    def __init__(self, dim, mult=4, dropout=0.0):
        super().__init__()
        inner_dim = int(dim * mult)

        self.ff = nn.ModuleList(
            [
                nn.Linear(dim, inner_dim),
                nn.GELU(approximate="tanh"),
                nn.Dropout(dropout),
                nn.Linear(inner_dim, dim),
            ]
        )

    def forward(self, hidden_states):
        for layer in self.ff:
            hidden_states = layer(hidden_states)
        return hidden_states


# Modified from Llama with a different rotate function, will fixed in next release
def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
    """Applies Rotary Position Embedding to the query and key tensors.

    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`, *optional*):
            Deprecated and unused.
        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.
    """

    def rotate_half_codec(x):
        # x = rearrange(x, "... (d r) -> ... d r", r=2)
        x = x.reshape(*x.shape[:-1], -1, 2)
        x1, x2 = x.unbind(dim=-1)
        x = torch.stack((-x2, x1), dim=-1)
        return x.reshape(*x.shape[:-2], -1)

    cos = cos.unsqueeze(unsqueeze_dim)
    sin = sin.unsqueeze(unsqueeze_dim)
    q_embed = (q * cos) + (rotate_half_codec(q) * sin)
    k_embed = (k * cos) + (rotate_half_codec(k) * sin)
    return q_embed, k_embed


class DiTAttention(nn.Module):
    def __init__(self, config: Qwen2_5OmniDiTConfig):
        super().__init__()

        self.config = config
        self.dim = config.hidden_size
        self.heads = config.num_attention_heads
        self.inner_dim = config.head_dim * config.num_attention_heads
        self.dropout = config.dropout
        self.is_causal = False

        self.to_q = nn.Linear(config.hidden_size, self.inner_dim)
        self.to_k = nn.Linear(config.hidden_size, self.inner_dim)
        self.to_v = nn.Linear(config.hidden_size, self.inner_dim)

        self.to_out = nn.ModuleList([nn.Linear(self.inner_dim, config.hidden_size), nn.Dropout(config.dropout)])

    def forward(
        self,
        hidden_states,  # noised input x
        position_embeddings=None,  # rotary position embedding for x
        attention_mask=None,
    ) -> torch.Tensor:
        batch_size = hidden_states.shape[0]

        # `sample` projections.
        query = self.to_q(hidden_states)
        key = self.to_k(hidden_states)
        value = self.to_v(hidden_states)

        # attention
        inner_dim = key.shape[-1]
        head_dim = inner_dim // self.heads
        query = query.view(batch_size, -1, self.heads, head_dim).transpose(1, 2)
        key = key.view(batch_size, -1, self.heads, head_dim).transpose(1, 2)
        value = value.view(batch_size, -1, self.heads, head_dim).transpose(1, 2)

        # apply rotary position embedding
        # Due to training process, only first head is applied with RoPE, will be fixed at next release
        cos, sin = position_embeddings
        query[:, :1], key[:, :1] = apply_rotary_pos_emb(query[:, :1], key[:, :1], cos, sin)

        attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
        attention_weights, _ = attention_interface(
            self,
            query,
            key,
            value,
            attention_mask=attention_mask,
            is_causal=False,
        )

        # mask. e.g. inference got a batch with different target durations, mask out the padding
        attention_weights = attention_weights.reshape(batch_size, -1, self.heads * head_dim)
        attention_weights = attention_weights.to(query.dtype)

        # linear proj
        attention_output = self.to_out[0](attention_weights)
        attention_output = self.to_out[1](attention_output)

        return attention_output


# time step conditioning embedding
class SinusPositionEmbedding(nn.Module):
    def __init__(self, dim):
        super().__init__()
        self.dim = dim

    def forward(self, hidden_states, scale=1000):
        device = hidden_states.device
        half_dim = self.dim // 2
        emb = math.log(10000) / (half_dim - 1)
        emb = torch.exp(torch.arange(half_dim, device=device).float() * -emb)
        emb = scale * hidden_states.unsqueeze(1) * emb.unsqueeze(0)
        emb = torch.cat((emb.sin(), emb.cos()), dim=-1)
        return emb.type_as(hidden_states)


class DiTTimestepEmbedding(nn.Module):
    def __init__(self, dim, freq_embed_dim=256):
        super().__init__()
        self.time_embed = SinusPositionEmbedding(freq_embed_dim)
        self.time_mlp = nn.ModuleList([nn.Linear(freq_embed_dim, dim), nn.SiLU(), nn.Linear(dim, dim)])

    def forward(self, timestep):
        time_hidden = self.time_embed(timestep)
        time_hidden = time_hidden.to(timestep.dtype)
        for layer in self.time_mlp:
            time_hidden = layer(time_hidden)  # b d
        return time_hidden


class DiTDecoderLayer(nn.Module):
    def __init__(self, config: Qwen2_5OmniDiTConfig, look_ahead_block=0, look_backward_block=0):
        super().__init__()
        self.attn_norm = Qwen2_5_OmniAdaLayerNormZero(config.hidden_size)

        self.attn = DiTAttention(config)
        self.look_ahead_block = look_ahead_block
        self.look_backward_block = look_backward_block
        self.ff_norm = nn.LayerNorm(config.hidden_size, elementwise_affine=False, eps=1e-6)
        self.ff = DiTMLP(dim=config.hidden_size, mult=config.ff_mult, dropout=config.dropout)

    def forward(
        self, hidden_states, timestep, position_embeddings=None, block_diff=None
    ):  # x: noised input, t: time embedding
        # pre-norm & modulation for attention input
        norm, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.attn_norm(hidden_states, emb=timestep)

        # attention
        attn_output = self.attn(
            hidden_states=norm,
            position_embeddings=position_embeddings,
            attention_mask=(block_diff >= -float(self.look_backward_block))
            & (block_diff <= float(self.look_ahead_block)),
        )

        # process attention output for input x
        hidden_states = hidden_states + gate_msa.unsqueeze(1) * attn_output

        norm = self.ff_norm(hidden_states) * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
        ff_output = self.ff(norm)
        hidden_states = hidden_states + gate_mlp.unsqueeze(1) * ff_output

        return hidden_states


class SnakeBeta(nn.Module):
    """
    A modified Snake function which uses separate parameters for the magnitude of the periodic components
    Shape:
        - Input: (B, C, T)
        - Output: (B, C, T), same shape as the input
    Parameters:
        - alpha - trainable parameter that controls frequency
        - beta - trainable parameter that controls magnitude
    References:
        - This activation function is a modified version based on this paper by Liu Ziyin, Tilman Hartwig, Masahito Ueda:
        https://huggingface.co/papers/2006.08195
    """

    def __init__(self, in_features, alpha=1.0):
        super().__init__()
        self.in_features = in_features

        # initialize alpha
        self.alpha = Parameter(torch.zeros(in_features) * alpha)
        self.beta = Parameter(torch.zeros(in_features) * alpha)

        self.no_div_by_zero = 0.000000001

    def forward(self, hidden_states):
        """
        Forward pass of the function.
        Applies the function to the input elementwise.
        SnakeBeta ∶= x + 1/b * sin^2 (xa)
        """
        alpha = self.alpha.unsqueeze(0).unsqueeze(-1)  # line up with x to [B, C, T]
        beta = self.beta.unsqueeze(0).unsqueeze(-1)
        alpha = torch.exp(alpha)
        beta = torch.exp(beta)
        hidden_states = hidden_states + (1.0 / (beta + self.no_div_by_zero)) * torch.pow(
            torch.sin(hidden_states * alpha), 2
        )

        return hidden_states


def kaiser_sinc_filter1d(cutoff, half_width, kernel_size):
    """Generates a 1D Kaiser-windowed sinc filter.

    Args:
        cutoff (float): Normalized cutoff frequency (0 to 0.5).
        half_width (float): Transition bandwidth.
        kernel_size (int): Number of filter taps.

    Returns:
        torch.Tensor: A tensor of shape (1, 1, kernel_size) representing the filter.
    """
    is_even = kernel_size % 2 == 0
    half_size = kernel_size // 2

    # Compute Kaiser window parameters
    delta_f = 4 * half_width
    attenuation = 2.285 * (half_size - 1) * math.pi * delta_f + 7.95

    if attenuation > 50.0:
        beta = 0.1102 * (attenuation - 8.7)
    elif attenuation >= 21.0:
        beta = 0.5842 * (attenuation - 21) ** 0.4 + 0.07886 * (attenuation - 21.0)
    else:
        beta = 0.0

    kaiser_window = torch.kaiser_window(kernel_size, beta=beta, periodic=False, dtype=torch.float32)

    # Compute time indices
    if is_even:
        time_indices = torch.arange(-half_size, half_size) + 0.5
    else:
        time_indices = torch.arange(kernel_size) - half_size

    # Compute sinc filter
    if cutoff == 0:
        return torch.zeros((1, 1, kernel_size), dtype=torch.float32)  # Ensures correct shape

    sinc_filter = torch.sinc(2 * cutoff * time_indices)
    normalized_filter = 2 * cutoff * kaiser_window * sinc_filter

    # Normalize to ensure sum = 1 (avoid leakage of constant component)
    normalized_filter /= normalized_filter.sum()

    return normalized_filter.view(1, 1, kernel_size)


class UpSample1d(nn.Module):
    def __init__(self, ratio=2, kernel_size=None):
        super().__init__()
        self.ratio = ratio
        self.kernel_size = int(6 * ratio // 2) * 2 if kernel_size is None else kernel_size
        self.stride = ratio
        self.pad = self.kernel_size // ratio - 1
        self.pad_left = self.pad * self.stride + (self.kernel_size - self.stride) // 2
        self.pad_right = self.pad * self.stride + (self.kernel_size - self.stride + 1) // 2

        filter = kaiser_sinc_filter1d(cutoff=0.5 / ratio, half_width=0.6 / ratio, kernel_size=self.kernel_size)
        self.register_buffer("filter", filter, persistent=False)

    def forward(self, hidden_states):
        channels = hidden_states.shape[1]

        hidden_states = F.pad(hidden_states, (self.pad, self.pad), mode="replicate")
        hidden_states = self.ratio * F.conv_transpose1d(
            hidden_states, self.filter.expand(channels, -1, -1), stride=self.stride, groups=channels
        )
        hidden_states = hidden_states[..., self.pad_left : -self.pad_right]

        return hidden_states


class DownSample1d(nn.Module):
    def __init__(self, ratio=2, kernel_size=None):
        super().__init__()
        cutoff = 0.5 / ratio
        half_width = 0.6 / ratio

        if cutoff < 0.0:
            raise ValueError("Minimum cutoff must be larger than zero.")
        if cutoff > 0.5:
            raise ValueError("A cutoff above 0.5 does not make sense.")

        self.even = kernel_size % 2 == 0
        self.pad_left = kernel_size // 2 - int(self.even)
        self.pad_right = kernel_size // 2
        self.stride = ratio
        filter = kaiser_sinc_filter1d(cutoff, half_width, kernel_size)
        self.register_buffer("filter", filter, persistent=False)

    def forward(self, hidden_states):
        channels = hidden_states.shape[1]
        hidden_states = F.pad(hidden_states, (self.pad_left, self.pad_right), mode="replicate")
        out = F.conv1d(hidden_states, self.filter.expand(channels, -1, -1), stride=self.stride, groups=channels)
        return out


class TorchActivation1d(nn.Module):
    def __init__(
        self,
        activation,
        up_ratio: int = 2,
        down_ratio: int = 2,
        up_kernel_size: int = 12,
        down_kernel_size: int = 12,
    ):
        super().__init__()
        if not callable(activation):
            raise TypeError("Activation function must be callable")
        self.act = activation
        self.upsample = UpSample1d(up_ratio, up_kernel_size)
        self.downsample = DownSample1d(down_ratio, down_kernel_size)

    def forward(self, hidden_states):
        hidden_states = self.upsample(hidden_states)
        hidden_states = self.act(hidden_states)
        hidden_states = self.downsample(hidden_states)

        return hidden_states


class AMPBlock(torch.nn.Module):
    def __init__(
        self,
        channels,
        kernel_size=3,
        dilation=(1, 3, 5),
    ):
        super().__init__()

        self.convs1 = nn.ModuleList(
            [
                nn.Conv1d(
                    channels,
                    channels,
                    kernel_size,
                    1,
                    dilation=dilation[0],
                    padding=self._get_padding(kernel_size, dilation[0]),
                ),
                nn.Conv1d(
                    channels,
                    channels,
                    kernel_size,
                    1,
                    dilation=dilation[1],
                    padding=self._get_padding(kernel_size, dilation[1]),
                ),
                nn.Conv1d(
                    channels,
                    channels,
                    kernel_size,
                    1,
                    dilation=dilation[2],
                    padding=self._get_padding(kernel_size, dilation[2]),
                ),
            ]
        )

        self.convs2 = nn.ModuleList(
            [
                nn.Conv1d(
                    channels,
                    channels,
                    kernel_size,
                    1,
                    dilation=1,
                    padding=self._get_padding(kernel_size, 1),
                ),
                nn.Conv1d(
                    channels,
                    channels,
                    kernel_size,
                    1,
                    dilation=1,
                    padding=self._get_padding(kernel_size, 1),
                ),
                nn.Conv1d(
                    channels,
                    channels,
                    kernel_size,
                    1,
                    dilation=1,
                    padding=self._get_padding(kernel_size, 1),
                ),
            ]
        )

        self.num_layers = len(self.convs1) + len(self.convs2)  # total number of conv layers

        self.activations = nn.ModuleList(
            [TorchActivation1d(activation=SnakeBeta(channels)) for _ in range(self.num_layers)]
        )

    def _get_padding(self, kernel_size, dilation=1):
        return int((kernel_size * dilation - dilation) / 2)

    def forward(self, hidden_states):
        acts1, acts2 = self.activations[::2], self.activations[1::2]
        for conv1, conv2, act1, act2 in zip(self.convs1, self.convs2, acts1, acts2):
            residual = hidden_states
            hidden_states = act1(hidden_states)
            hidden_states = conv1(hidden_states)
            hidden_states = act2(hidden_states)
            hidden_states = conv2(hidden_states)
            hidden_states = residual + hidden_states

        return hidden_states


@auto_docstring(
    custom_intro="""
    The full Qwen2.5Omni Token2WavBigVGAN model. Which take mel spectrogram as input and predict waveform.
    """
)
class Qwen2_5OmniToken2WavBigVGANModel(Qwen2_5OmniPreTrainedModel):
    config: Qwen2_5OmniBigVGANConfig

    def __init__(self, config: Qwen2_5OmniBigVGANConfig):
        super().__init__(config)
        self.num_residual_blocks = len(config.resblock_kernel_sizes)
        self.num_upsample_layers = len(config.upsample_rates)

        self.conv_pre = nn.Conv1d(config.mel_dim, config.upsample_initial_channel, 7, 1, padding=3)

        # Removing extra ModuleList breaks official state dict
        ups = [
            nn.ModuleList(
                [
                    nn.ConvTranspose1d(
                        config.upsample_initial_channel // (2**layer_idx),
                        config.upsample_initial_channel // (2 ** (layer_idx + 1)),
                        kernel_size,
                        stride,
                        padding=(kernel_size - stride) // 2,
                    )
                ]
            )
            for layer_idx, (stride, kernel_size) in enumerate(zip(config.upsample_rates, config.upsample_kernel_sizes))
        ]
        self.ups = nn.ModuleList(ups)

        self.resblocks = nn.ModuleList(
            [
                AMPBlock(config.upsample_initial_channel // (2 ** (layer_idx + 1)), kernel_size, dilation)
                for layer_idx in range(self.num_upsample_layers)
                for kernel_size, dilation in zip(config.resblock_kernel_sizes, config.resblock_dilation_sizes)
            ]
        )

        self.activation_post = TorchActivation1d(
            activation=SnakeBeta(config.upsample_initial_channel // (2**self.num_upsample_layers))
        )
        self.conv_post = nn.Conv1d(
            config.upsample_initial_channel // (2**self.num_upsample_layers), 1, 7, 1, padding=3, bias=False
        )

    def normalize_spectrogram(self, spectrogram, max_value, min_db):
        return torch.clamp((2 * max_value) * ((spectrogram - min_db) / (-min_db)) - max_value, -max_value, max_value)

    def amplitude_to_db(self, amplitude, min_db_level):
        min_level = torch.exp(
            torch.tensor(min_db_level / 20.0 * np.log(10), device=amplitude.device, dtype=amplitude.dtype)
        )
        return 20 * torch.log10(torch.clamp(amplitude, min=min_level))

    def process_mel_spectrogram(self, mel_spectrogram):
        amplitude_spectrum = torch.exp(mel_spectrogram)
        decibel_spectrum = self.amplitude_to_db(amplitude_spectrum, -115) - 20
        return self.normalize_spectrogram(decibel_spectrum, 1, -115)

    def forward(self, mel_spectrogram):
        processed_spectrogram = self.process_mel_spectrogram(mel_spectrogram)
        hidden_representation = self.conv_pre(processed_spectrogram)

        for layer_index in range(self.num_upsample_layers):
            hidden_representation = self.ups[layer_index][0](hidden_representation)
            residual_output = sum(
                self.resblocks[layer_index * self.num_residual_blocks + block_index](hidden_representation)
                for block_index in range(self.num_residual_blocks)
            )
            residual_output = residual_output / self.num_residual_blocks
            hidden_representation = residual_output

        hidden_representation = self.activation_post(hidden_representation)
        output_waveform = self.conv_post(hidden_representation)
        return torch.clamp(output_waveform, min=-1.0, max=1.0).squeeze().cpu()


class RungeKutta4ODESolver:
    def __init__(self, function, initial_value):
        self.function = function
        self.initial_value = initial_value

        self._one_third = 1 / 3
        self._two_thirds = 2 / 3

    def _rk4_step(self, function, time_start, time_step, time_end, value_start, function_value_start=None):
        k1 = function_value_start if function_value_start is not None else function(time_start, value_start)
        k2 = function(time_start + time_step * self._one_third, value_start + time_step * k1 * self._one_third)
        k3 = function(time_start + time_step * self._two_thirds, value_start + time_step * (k2 - k1 * self._one_third))
        k4 = function(time_end, value_start + time_step * (k1 - k2 + k3))
        return (k1 + 3 * (k2 + k3) + k4) * time_step / 8

    def _compute_step(self, function, time_start, time_step, time_end, value_start):
        function_value_start = function(time_start, value_start)
        return self._rk4_step(
            function, time_start, time_step, time_end, value_start, function_value_start=function_value_start
        ), function_value_start

    def _linear_interpolation(self, time_start, time_end, value_start, value_end, time_point):
        if time_point == time_start:
            return value_start
        if time_point == time_end:
            return value_end
        weight = (time_point - time_start) / (time_end - time_start)
        return value_start + weight * (value_end - value_start)

    def integrate(self, time_points):
        solution = torch.empty(
            len(time_points),
            *self.initial_value.shape,
            dtype=self.initial_value.dtype,
            device=self.initial_value.device,
        )
        solution[0] = self.initial_value

        current_index = 1
        current_value = self.initial_value
        for time_start, time_end in zip(time_points[:-1], time_points[1:]):
            time_step = time_end - time_start
            delta_value, _ = self._compute_step(self.function, time_start, time_step, time_end, current_value)
            next_value = current_value + delta_value

            while current_index < len(time_points) and time_end >= time_points[current_index]:
                solution[current_index] = self._linear_interpolation(
                    time_start, time_end, current_value, next_value, time_points[current_index]
                )
                current_index += 1

            current_value = next_value

        return solution


@auto_docstring(
    custom_intro="""
    The full Qwen2.5Omni Token2WavDiT model. Which take speech tokens as input and predict mel spectrogram.
    """
)
class Qwen2_5OmniToken2WavDiTModel(Qwen2_5OmniPreTrainedModel):
    config: Qwen2_5OmniDiTConfig
    _no_split_modules = ["DiTDecoderLayer"]

    def __init__(self, config: Qwen2_5OmniDiTConfig):
        super().__init__(config)
        self.mel_dim = config.mel_dim
        self.repeats = config.repeats
        self.time_embed = DiTTimestepEmbedding(config.hidden_size)

        self.text_embed = DiTCodecEmbedding(config.num_embeds, config.emb_dim, config.repeats)
        self.input_embed = DiTInputEmbedding(config)

        self.rotary_embed = Qwen2_5OmniDiTRotaryEmbedding(config.head_dim)

        self.hidden_size = config.hidden_size
        self.layers = config.num_hidden_layers
        self.block_size = config.block_size
        self.num_attention_heads = config.num_attention_heads

        self.transformer_blocks = nn.ModuleList()
        for i in range(config.num_hidden_layers):
            self.transformer_blocks.append(
                DiTDecoderLayer(
                    config,
                    look_ahead_block=1 if i in config.look_ahead_layers else 0,
                    look_backward_block=1 if i in config.look_backward_layers else 0,
                )
            )

        self.norm_out = Qwen2_5_OmniAdaLayerNormZero_Final(config.hidden_size)  # final modulation
        self.proj_out = nn.Linear(config.hidden_size, config.mel_dim)

    def _create_block_diff(self, hidden_states):
        batch, seq_len = hidden_states.shape[0], hidden_states.shape[1]
        block_indices = torch.arange(seq_len, device=hidden_states.device) // self.block_size  # [seq_length]

        block_i = block_indices.unsqueeze(1)  # [seq_length, 1]
        block_j = block_indices.unsqueeze(0)  # [1, seq_length]
        block_diff = block_j - block_i  # (n, n)

        return block_diff.expand(batch, self.num_attention_heads, seq_len, seq_len)

    def forward(
        self,
        hidden_states,
        condition_vector,
        speaker_embedding,
        quantized_code,
        time_step,
        drop_audio_conditioning=False,
        drop_code=False,
        apply_cfg=True,
    ):
        batch_size = hidden_states.shape[0]
        if time_step.ndim == 0:
            time_step = time_step.repeat(batch_size)

        # Compute embeddings
        time_embedding = self.time_embed(time_step)
        text_embedding = self.text_embed(quantized_code, drop_code=False if apply_cfg else drop_code)
        text_embedding_unconditioned = self.text_embed(quantized_code, drop_code=True) if apply_cfg else None

        hidden_states = self.input_embed(
            hidden_states,
            speaker_embedding,
            condition_vector,
            text_embedding,
            drop_audio_cond=drop_audio_conditioning,
            code_embed_uncond=text_embedding_unconditioned,
            apply_cfg=apply_cfg,
        )

        # Compute positional encodings
        position_embeddings = self.rotary_embed(hidden_states)
        blockwise_difference = self._create_block_diff(hidden_states)

        # Transformer blocks
        for transformer_block in self.transformer_blocks:
            hidden_states = transformer_block(
                hidden_states,
                time_embedding,
                position_embeddings=position_embeddings,
                block_diff=blockwise_difference,
            )

        hidden_states = self.norm_out(hidden_states, time_embedding)
        output = self.proj_out(hidden_states)

        return output

    @torch.no_grad()
    def sample(
        self,
        conditioning_vector,
        reference_mel_spectrogram,
        quantized_code,
        num_steps=10,
        guidance_scale=0.5,
        sway_coefficient=-1.0,
    ):
        noise_initialization = torch.randn([1, 30000, self.mel_dim], dtype=reference_mel_spectrogram.dtype)
        maximum_duration = quantized_code.shape[1] * self.repeats
        initial_state = noise_initialization[:, :maximum_duration].to(quantized_code.device)
        batch_size = reference_mel_spectrogram.shape[0]
        conditioning_vector = conditioning_vector.unsqueeze(1).repeat(1, maximum_duration, 1)

        if batch_size != 1:
            raise ValueError("Only batch size = 1 is currently supported")

        def ode_function(time_step, hidden_states):
            if guidance_scale < 1e-5:
                prediction = self(
                    hidden_states=hidden_states,
                    speaker_embedding=conditioning_vector,
                    condition_vector=reference_mel_spectrogram,
                    quantized_code=quantized_code,
                    time_step=time_step,
                    drop_audio_conditioning=False,
                    drop_code=False,
                )
                return prediction

            model_output = self(
                hidden_states=hidden_states,
                quantized_code=quantized_code,
                speaker_embedding=conditioning_vector,
                condition_vector=reference_mel_spectrogram,
                time_step=time_step,
                apply_cfg=True,
            )
            guided_prediction, null_prediction = torch.chunk(model_output, 2, dim=0)
            return guided_prediction + (guided_prediction - null_prediction) * guidance_scale

        initial_time = 0
        time_embedding = torch.linspace(
            initial_time, 1, num_steps, device=quantized_code.device, dtype=conditioning_vector.dtype
        )

        if sway_coefficient is not None:
            time_embedding += sway_coefficient * (torch.cos(torch.pi / 2 * time_embedding) - 1 + time_embedding)

        ode_solver = RungeKutta4ODESolver(function=ode_function, initial_value=initial_state)
        solution_trajectory = ode_solver.integrate(time_embedding)

        generated_waveform = solution_trajectory[-1]
        generated_mel_spectrogram = generated_waveform.permute(0, 2, 1)
        return generated_mel_spectrogram


@auto_docstring(
    custom_intro="""
    The full Qwen2.5Omni Token2Wav model. Consists a DiT model take speech tokens as input and predict mel spectrogram and a BigVGAN vocoder take mel spectrogram as input and predict waveform.
    """
)
class Qwen2_5OmniToken2WavModel(Qwen2_5OmniPreTrainedModel):
    config: Qwen2_5OmniToken2WavConfig
    base_model_prefix = "model"
    _no_split_modules = ["Qwen2_5OmniToken2WavDiTModel", "Qwen2_5OmniToken2WavBigVGANModel"]

    def __init__(self, config: Qwen2_5OmniToken2WavConfig):
        super().__init__(config)
        attn_impl = config._attn_implementation
        if config._attn_implementation == "flash_attention_2":
            logger.warning_once(
                "Qwen2_5OmniToken2WavModel must inference with fp32, but flash_attention_2 only supports fp16 and bf16, "
                "attention implementation of Qwen2_5OmniToken2WavModel will fallback to sdpa."
            )
            attn_impl = "sdpa"
        elif config._attn_implementation == "eager":
            logger.warning_once(
                "Qwen2_5OmniToken2WavModel does not support eager attention implementation, fall back to sdpa"
            )
            attn_impl = "sdpa"
        self.code2wav_dit_model = Qwen2_5OmniToken2WavDiTModel._from_config(
            config.dit_config, attn_implementation=attn_impl
        )
        self.code2wav_bigvgan_model = Qwen2_5OmniToken2WavBigVGANModel._from_config(
            config.bigvgan_config, attn_implementation=attn_impl
        )

    def forward(
        self,
        code,
        conditioning,
        reference_mel,
        num_steps=10,
        guidance_scale=0.5,
        sway_coefficient=-1.0,
        **kwargs,
    ):
        """Generates a waveform from input code and conditioning parameters."""

        mel_spectrogram = self.code2wav_dit_model.sample(
            conditioning,
            reference_mel,
            code,
            num_steps=num_steps,
            guidance_scale=guidance_scale,
            sway_coefficient=sway_coefficient,
        )

        waveform = self.code2wav_bigvgan_model(mel_spectrogram)

        return waveform


############################
#    Start Qwen2.5Omni     #
############################


@auto_docstring(
    custom_intro="""
    The full Qwen2.5Omni model, a multimodal model composed of 3 sub-models:
    - [`Qwen2_5OmniThinkerForConditionalGeneration`]:
    a causal auto-regressive transformer takes text, audio, image, video as input and predict text tokens.
    - [`Qwen2_5OmniTalkerForConditionalGeneration`]:
    a causal auto-regressive transformer takes thinker hidden states and response as input and predict speech tokens.
    - [`Qwen2_5OmniToken2WavModel`]:
    a DiT model take speech tokens as input and predict mel spectrogram and a BigVGAN vocoder take mel spectrogram as input and predict waveform.
    """
)
class Qwen2_5OmniForConditionalGeneration(Qwen2_5OmniPreTrainedModel, GenerationMixin):
    config: Qwen2_5OmniConfig
    _no_split_modules = [
        "Qwen2_5OmniTalkerForConditionalGeneration",
        "Qwen2_5OmniToken2WavModel",
    ]

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

        self.thinker = Qwen2_5OmniThinkerForConditionalGeneration(config.thinker_config)

        self.has_talker = config.enable_audio_output
        self.speaker_map = {}
        if config.enable_audio_output:
            self.enable_talker()
        self.post_init()

    def enable_talker(self):
        self.talker = Qwen2_5OmniTalkerForConditionalGeneration(self.config.talker_config)
        self.token2wav = Qwen2_5OmniToken2WavModel(self.config.token2wav_config)
        self.token2wav.float()
        self.has_talker = True

    def load_speakers(self, path):
        check_torch_load_is_safe()
        for key, value in torch.load(path, weights_only=True).items():
            self.speaker_map[key] = value
        logger.info(f"Speaker {list(self.speaker_map.keys())} loaded")

    def disable_talker(self):
        if hasattr(self, "talker"):
            del self.talker
        if hasattr(self, "token2wav"):
            del self.token2wav
        self.has_talker = False

    @classmethod
    def from_pretrained(
        cls,
        pretrained_model_name_or_path,
        *model_args,
        config=None,
        cache_dir=None,
        ignore_mismatched_sizes=False,
        force_download=False,
        local_files_only=False,
        token=None,
        revision="main",
        use_safetensors=None,
        weights_only=True,
        **kwargs,
    ):
        model = super().from_pretrained(
            pretrained_model_name_or_path,
            *model_args,
            config=config,
            cache_dir=cache_dir,
            ignore_mismatched_sizes=ignore_mismatched_sizes,
            force_download=force_download,
            local_files_only=local_files_only,
            token=token,
            revision=revision,
            use_safetensors=use_safetensors,
            weights_only=weights_only,
            **kwargs,
        )
        spk_path = cached_file(
            pretrained_model_name_or_path,
            "spk_dict.pt",
            subfolder=kwargs.pop("subfolder", None),
            cache_dir=kwargs.pop("cache_dir", None),
            force_download=kwargs.pop("force_download", False),
            proxies=kwargs.pop("proxies", None),
            resume_download=kwargs.pop("resume_download", None),
            local_files_only=kwargs.pop("local_files_only", False),
            token=kwargs.pop("use_auth_token", None),
            revision=kwargs.pop("revision", None),
        )
        if spk_path is None:
            raise ValueError(f"""{pretrained_model_name_or_path}/{spk_path} not exists""")
        model.load_speakers(spk_path)

        return model

    @torch.no_grad()
    # TODO: raushan, defaults should be saved in generation config
    def generate(
        self,
        input_ids: Optional[torch.Tensor] = None,
        speaker: str = "Chelsie",
        use_audio_in_video: bool = False,
        return_audio: Optional[bool] = None,
        thinker_max_new_tokens: int = 1024,
        talker_max_new_tokens: int = 4096,
        talker_do_sample: bool = True,
        talker_top_k: int = 40,
        talker_top_p: float = 0.8,
        talker_temperature: float = 0.9,
        talker_eos_token_id: list[int] = [8292, 8294],
        talker_repetition_penalty: float = 1.05,
        **kwargs,
    ):
        r"""
        Generate text response and audio from input.

        Args:
            input_ids (`Optional[torch.Tensor]`, *optional*):
                Input ids, should obtain from processor.
            speaker (`str` , defaults to "Chelsie"):
                Which speaker should be used in audio response.
            use_audio_in_video (`bool`, defaults to False):
                Whether or not use audio track in video, should same as the parameter in `process_audio_info`.
            return_audio (`Optional[bool]`, *optional*):
                Whether or not return response in audio format. When `return_audio=None`, this parameter is same as `config.enable_audio_output`.
            kwargs (*optional*):
                - Without a prefix, they will be entered as `**kwargs` for the `generate` method of each sub-model.
                - With a *thinker_*, *talker_*, *token2wav_* prefix, they will be input for the `generate` method of the
                thinker, talker and token2wav respectively. It has the priority over the keywords without a prefix.
        Returns:
            When `return_audio=False`:
                - **Text** (`torch.Tensor`): Generated text token sequence.
            When `return_audio=True`:
                - **Text** (`torch.Tensor`): Generated text token sequence.
                - **Audio waveform** (`torch.Tensor`): Generated audio waveform.
        """
        if speaker not in self.speaker_map:
            raise ValueError(f"{speaker} is not available, available speakers: {self.speaker_map.keys()}")
        if return_audio and not self.has_talker:
            raise ValueError(
                "Cannot use talker when talker module not initialized. Use `enable_talker` method or set enable_talker in config to enable talker."
            )
        if return_audio is None:
            return_audio = self.has_talker
        if input_ids.shape[0] != 1 and return_audio:
            raise NotImplementedError("Qwen2.5-Omni currently does not support batched inference with audio output")

        shared_kwargs = {"use_audio_in_video": use_audio_in_video}
        thinker_kwargs = {
            "max_new_tokens": thinker_max_new_tokens,
        }
        talker_kwargs = {
            "max_new_tokens": talker_max_new_tokens,
            "do_sample": talker_do_sample,
            "top_k": talker_top_k,
            "top_p": talker_top_p,
            "temperature": talker_temperature,
            "eos_token_id": talker_eos_token_id,
            "repetition_penalty": talker_repetition_penalty,
        }
        token2wav_kwargs = {}

        for key, value in kwargs.items():
            if key.startswith("thinker_"):
                thinker_kwargs[key[len("thinker_") :]] = value
            elif key.startswith("talker_"):
                talker_kwargs[key[len("talker_") :]] = value
            elif key.startswith("token2wav_"):
                token2wav_kwargs[key[len("token2wav_") :]] = value
            # Process special input values
            elif key == "feature_attention_mask":
                thinker_kwargs[key] = value
                talker_kwargs["audio_feature_lengths"] = torch.sum(value, dim=1)
            elif key == "input_features" or key == "attention_mask":
                thinker_kwargs[key] = value
            # Put other key to shared kwargs
            else:
                shared_kwargs[key] = value

        # Merge kwargs
        for key, value in shared_kwargs.items():
            if key not in thinker_kwargs:
                thinker_kwargs[key] = value
            if key not in talker_kwargs:
                talker_kwargs[key] = value
            if key not in token2wav_kwargs:
                token2wav_kwargs[key] = value
        speaker_params = self.speaker_map[speaker]

        # 1. Generate from thinker module
        generate_audio = return_audio and self.has_talker
        if generate_audio:
            thinker_kwargs["output_hidden_states"] = True
            thinker_kwargs["return_dict_in_generate"] = True

        thinker_result = self.thinker.generate(input_ids=input_ids, **thinker_kwargs)

        if not generate_audio:
            return thinker_result

        # 2. Generate speech tokens from talker module
        embeds_to_talker = thinker_result.hidden_states[0][0].clone().to(input_ids.device)
        if thinker_kwargs.get("input_features") is not None:
            audio_ids_mask = input_ids == self.config.thinker_config.audio_token_index
            audio_mask = audio_ids_mask.unsqueeze(-1).expand_as(embeds_to_talker)
            audio_mask_tensor = torch.zeros(
                [audio_ids_mask.sum(), embeds_to_talker.shape[-1]],
                dtype=embeds_to_talker.dtype,
                device=input_ids.device,
            )
            embeds_to_talker.masked_scatter_(audio_mask, audio_mask_tensor)
        if thinker_kwargs.get("pixel_values") is not None:
            image_ids_mask = input_ids == self.config.thinker_config.image_token_index
            image_mask = image_ids_mask.unsqueeze(-1).expand_as(embeds_to_talker)
            image_mask_tensor = torch.zeros(
                [image_ids_mask.sum(), embeds_to_talker.shape[-1]],
                dtype=embeds_to_talker.dtype,
                device=input_ids.device,
            )
            embeds_to_talker.masked_scatter_(image_mask, image_mask_tensor)
        if thinker_kwargs.get("pixel_values_videos") is not None:
            video_ids_mask = input_ids == self.config.thinker_config.video_token_index
            video_mask = video_ids_mask.unsqueeze(-1).expand_as(embeds_to_talker)
            video_mask_tensor = torch.zeros(
                [video_ids_mask.sum(), embeds_to_talker.shape[-1]],
                dtype=embeds_to_talker.dtype,
                device=input_ids.device,
            )
            embeds_to_talker.masked_scatter_(video_mask, video_mask_tensor)

        processed_thinker_hidden = (
            (embeds_to_talker,) + thinker_result.hidden_states[0][1:],
        ) + thinker_result.hidden_states[1:]
        thinker_generate_ids = thinker_result.sequences[:, input_ids.size(1) :].to(input_ids.device)
        thinker_token_embeds = [
            token_hidden_states[0].to(input_ids.device) for token_hidden_states in processed_thinker_hidden
        ]
        thinker_hidden_states = [
            token_hidden_states[-1].to(input_ids.device) for token_hidden_states in processed_thinker_hidden
        ]

        talker_text_bos_token = speaker_params["bos_token"]
        talker_input_text_ids = torch.cat(
            [
                input_ids,
                torch.tensor([[talker_text_bos_token]], dtype=torch.long, device=input_ids.device),
                thinker_generate_ids[:, :1],
            ],
            dim=-1,
        )

        talker_input_ids = torch.cat(
            [
                torch.full_like(input_ids, fill_value=self.talker.codec_mask_token),
                torch.tensor([[self.talker.codec_pad_token]], dtype=torch.long, device=input_ids.device),
                torch.tensor([[self.talker.codec_bos_token]], dtype=torch.long, device=input_ids.device),
            ],
            dim=1,
        )

        thinker_embed_tokens = self.thinker.get_input_embeddings()
        thinker_reply_part = torch.cat(thinker_hidden_states[1:], dim=1) + torch.cat(thinker_token_embeds[1:], dim=1)
        talker_inputs_embeds = thinker_hidden_states[0] + thinker_token_embeds[0]
        talker_text_bos_token = torch.tensor([[talker_text_bos_token]], dtype=torch.long, device=input_ids.device)
        talker_text_bos_embed = thinker_embed_tokens(talker_text_bos_token).to(input_ids.device)
        talker_inputs_embeds = torch.cat(
            [
                talker_inputs_embeds,
                talker_text_bos_embed,
                thinker_reply_part[:, :1, :],
            ],
            dim=1,
        )

        eos_token = torch.tensor([[self.talker.text_eos_token]], dtype=torch.long, device=input_ids.device)
        eos_embedding = thinker_embed_tokens(eos_token).to(input_ids.device)

        pad_token = torch.tensor([[self.talker.text_pad_token]], dtype=torch.long, device=input_ids.device)
        pad_embedding = thinker_embed_tokens(pad_token).to(input_ids.device)

        thinker_reply_part = torch.cat(
            [
                thinker_reply_part[:, 1:, :],
                eos_embedding,
                pad_embedding,
            ],
            dim=1,
        )

        talker_attention_mask = None
        if "attention_mask" in kwargs:
            talker_attention_mask = torch.cat(
                [kwargs["attention_mask"], kwargs["attention_mask"].new_ones((1, 2))], dim=1
            ).to(input_ids.device)

        talker_result = self.talker.generate(
            input_ids=talker_input_ids,
            input_text_ids=talker_input_text_ids,
            thinker_reply_part=thinker_reply_part,
            inputs_embeds=talker_inputs_embeds,
            attention_mask=talker_attention_mask,
            suppress_tokens=[self.talker.codec_bos_token],
            **{k: (v.to(input_ids.device) if torch.is_tensor(v) else v) for k, v in talker_kwargs.items()},
        )
        talker_generate_codes = talker_result[:, talker_input_ids.shape[1] : -1]

        # 3. Generate wavs from code
        if self.token2wav.dtype != torch.float:
            self.token2wav.float()

        wav = self.token2wav(
            talker_generate_codes.to(input_ids.device),
            conditioning=speaker_params["cond"].to(input_ids.device).float(),
            reference_mel=speaker_params["ref_mel"].to(input_ids.device).float(),
            **token2wav_kwargs,
        )

        return thinker_result.sequences, wav.float()


__all__ = [
    "Qwen2_5OmniForConditionalGeneration",
    "Qwen2_5OmniThinkerTextModel",
    "Qwen2_5OmniThinkerForConditionalGeneration",
    "Qwen2_5OmniTalkerModel",
    "Qwen2_5OmniTalkerForConditionalGeneration",
    "Qwen2_5OmniToken2WavDiTModel",
    "Qwen2_5OmniToken2WavBigVGANModel",
    "Qwen2_5OmniToken2WavModel",
    "Qwen2_5OmniPreTrainedModel",
    "Qwen2_5OmniPreTrainedModelForConditionalGeneration",
]