# 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.
"""PyTorch Qwen2.5Omni model (Audio, Image, Video)."""

import math
from dataclasses import dataclass
from typing import Any, Dict, List, Optional, Tuple, Union

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

from transformers.models.llama.modeling_llama import rotate_half
from transformers.models.qwen2.configuration_qwen2 import Qwen2Config
from transformers.models.qwen2_5_vl.configuration_qwen2_5_vl import Qwen2_5_VLVisionConfig
from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import (
    Qwen2_5_VisionTransformerPretrainedModel,
    Qwen2_5_VLAttention,
    Qwen2_5_VLMLP,
    Qwen2_5_VLPreTrainedModel,
    Qwen2_5_VLTextModel,
    Qwen2_5_VLVisionBlock,
    Qwen2RMSNorm,
)
from transformers.models.qwen2_audio.configuration_qwen2_audio import Qwen2AudioEncoderConfig
from transformers.models.qwen2_audio.modeling_qwen2_audio import Qwen2AudioEncoderLayer
from transformers.models.qwen2_vl.modeling_qwen2_vl import Qwen2VLRotaryEmbedding

from ...configuration_utils import PretrainedConfig
from ...generation import GenerationMixin
from ...modeling_outputs import BaseModelOutput, ModelOutput
from ...modeling_utils import ALL_ATTENTION_FUNCTIONS
from ...utils import (
    auto_docstring,
    check_torch_load_is_safe,
    is_flash_attn_2_available,
    is_flash_attn_greater_or_equal_2_10,
    logging,
)
from ...utils.hub import cached_file


if is_flash_attn_2_available():
    from flash_attn.flash_attn_interface import flash_attn_varlen_func as flash_attn_varlen_func
    from flash_attn.layers.rotary import apply_rotary_emb
else:
    flash_attn_varlen_func = None
    apply_rotary_emb = None


logger = logging.get_logger(__name__)


class Qwen2_5OmniVisionEncoderConfig(Qwen2_5_VLVisionConfig):
    r"""
    This is the configuration class to store the configuration of a [`Qwen2_5OmniThinkerVision`]. It is used to instantiate a
    Qwen2.5-VL vision encoder according to the specified arguments, defining the model architecture. Instantiating a
    configuration with the defaults will yield a similar configuration to that of the audio encoder of the Qwen2.5-VL
    architecture.

    e.g. [Qwen/Qwen2.5-Omni-7B](https://huggingface.co/Qwen/Qwen2.5-Omni-7B)

    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
    documentation from [`PretrainedConfig`] for more information.

    Args:
        depth (`int`, *optional*, defaults to 32):
            Number of layers (depth) in the model.
        hidden_size (`int`, *optional*, defaults to 3584):
            The size of the hidden layers.
        hidden_act (`str`, *optional*, defaults to `"quick_gelu"`):
            The non-linear activation function used in the model. Supported options include `"quick_gelu"` and others as applicable.
        mlp_ratio (`float`, *optional*, defaults to 4):
            The ratio used to determine the size of the MLP (Multi-Layer Perceptron) hidden layer.
        num_heads (`int`, *optional*, defaults to 16):
            Number of attention heads for each attention layer.
        in_channels (`int`, *optional*, defaults to 3):
            Number of input channels.
        patch_size (`int`, *optional*, defaults to 14):
            The size of the patches extracted from the input.
        spatial_merge_size (`int`, *optional*, defaults to 2):
            The size used for merging spatial dimensions.
        temporal_patch_size (`int`, *optional*, defaults to 2):
            The size used for patches along the temporal dimension.

    Example:

    ```python
    >>> from transformers import Qwen2_5OmniVisionEncoderConfig, Qwen2_5OmniVisionEncoder

    >>> # Initializing a Qwen2_5OmniVisionEncoderConfig
    >>> configuration = Qwen2_5OmniVisionEncoderConfig()

    >>> # Initializing a Qwen2_5OmniVisionEncoder (with random weights)
    >>> model = Qwen2_5OmniVisionEncoder(configuration)

    >>> # Accessing the model configuration
    >>> configuration = model.config
    ```"""

    model_type = "qwen2_5_omni_vision_encoder"

    def __init__(
        self,
        depth=32,
        hidden_size=3584,
        hidden_act="silu",
        intermediate_size=3420,
        num_heads=16,
        in_channels=3,
        patch_size=14,
        spatial_merge_size=2,
        temporal_patch_size=2,
        window_size=112,
        out_hidden_size=3584,
        fullatt_block_indexes=[7, 15, 23, 31],
        initializer_range=0.02,
        **kwargs,
    ):
        super().__init__(
            depth,
            hidden_size,
            hidden_act,
            intermediate_size,
            num_heads,
            in_channels,
            patch_size,
            spatial_merge_size,
            temporal_patch_size,
            window_size,
            out_hidden_size,
            fullatt_block_indexes,
            initializer_range=initializer_range,
            **kwargs,
        )
        del self.tokens_per_second


class Qwen2_5OmniAudioEncoderConfig(Qwen2AudioEncoderConfig):
    r"""
    This is the configuration class to store the configuration of a [`Qwen2_5OmniAudioEncoder`]. It is used to instantiate a
    Qwen2.5-Omni-Thinker audio encoder according to the specified arguments, defining the model architecture. Instantiating a
    configuration with the defaults will yield a similar configuration to that of the audio encoder of the Qwen2-Audio
    architecture.

    e.g. [Qwen/Qwen2.5-Omni-7B](https://huggingface.co/Qwen/Qwen2.5-Omni-7B)

    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
    documentation from [`PretrainedConfig`] for more information.

    Args:
        num_mel_bins (`int`, *optional*, defaults to 128):
            Number of mel features used per input features. Should correspond to the value used in the
            `Qwen2_5OmniProcessor` class.
        encoder_layers (`int`, *optional*, defaults to 32):
            Number of encoder layers.
        encoder_attention_heads (`int`, *optional*, defaults to 20):
            Number of attention heads for each attention layer in the Transformer encoder.
        encoder_ffn_dim (`int`, *optional*, defaults to 5120):
            Dimensionality of the "intermediate" (often named feed-forward) layer in encoder.
        d_model (`int`, *optional*, defaults to 1280):
            Dimensionality of the layers.
        dropout (`float`, *optional*, defaults to 0.0):
            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
        attention_dropout (`float`, *optional*, defaults to 0.0):
            The dropout ratio for the attention probabilities.
        activation_function (`str`, *optional*, defaults to `"gelu"`):
            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
            `"relu"`, `"silu"` and `"gelu_new"` are supported.
        activation_dropout (`float`, *optional*, defaults to 0.0):
            The dropout ratio for activations inside the fully connected layer.
        scale_embedding (`bool`, *optional*, defaults to `False`):
            Scale embeddings by diving by sqrt(d_model).
        initializer_range (`float`, *optional*, defaults to 0.02):
            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
        max_source_positions (`int`, *optional*, defaults to 1500):
            The maximum sequence length of log-mel filter-bank features that this model might ever be used with.
        n_window (`int`, *optional*, defaults to 100):
            The chunk for conv and flash attn in AudioEncoder.
        output_dim (`int`, *optional*, defaults to 3584):
            The output dimension of AudioEncoder.

    Example:

    ```python
    >>> from transformers import Qwen2_5OmniAudioEncoderConfig, Qwen2_5OmniAudioEncoder

    >>> # Initializing a Qwen2_5OmniAudioEncoderConfig
    >>> configuration = Qwen2_5OmniAudioEncoderConfig()

    >>> # Initializing a Qwen2_5OmniAudioEncoder (with random weights)
    >>> model = Qwen2_5OmniAudioEncoder(configuration)

    >>> # Accessing the model configuration
    >>> configuration = model.config
    ```"""

    model_type = "qwen2_5_omni_audio_encoder"

    def __init__(
        self,
        num_mel_bins=128,
        encoder_layers=32,
        encoder_attention_heads=20,
        encoder_ffn_dim=5120,
        d_model=1280,
        dropout=0,
        attention_dropout=0,
        activation_function="gelu",
        activation_dropout=0,
        scale_embedding=False,
        initializer_range=0.02,
        max_source_positions=1500,
        n_window=100,
        output_dim=3584,
        **kwargs,
    ):
        super().__init__(
            num_mel_bins,
            encoder_layers,
            encoder_attention_heads,
            encoder_ffn_dim,
            d_model,
            dropout,
            attention_dropout,
            activation_function,
            activation_dropout,
            scale_embedding,
            initializer_range,
            max_source_positions,
            **kwargs,
        )
        self.n_window = n_window
        self.output_dim = output_dim
        del self.encoder_layerdrop


class Qwen2_5OmniTextConfig(Qwen2Config):
    r"""
    This is the configuration class to store the configuration of a [`Qwen2_5OmniThinkerForConditionalGeneration`]. It is used to instantiate an
    Qwen2.5-Omni-Thinker model according to the specified arguments, defining the model architecture. Instantiating a configuration
    with the defaults will yield a similar configuration to that of the Qwen2.5-Omni-Thinker.

    e.g. [Qwen/Qwen2.5-Omni-7B](https://huggingface.co/Qwen/Qwen2.5-Omni-7B)

    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
    documentation from [`PretrainedConfig`] for more information.

    Args:
        vocab_size (`int`, *optional*, defaults to 152064):
            Vocabulary size of the QwenOmni model. Defines the number of different tokens that can be represented by the
            `inputs_ids` passed when calling [`Qwen2VLModel`]
        hidden_size (`int`, *optional*, defaults to 3584):
            Dimension of the hidden representations.
        intermediate_size (`int`, *optional*, defaults to 18944):
            Dimension of the MLP representations.
        num_hidden_layers (`int`, *optional*, defaults to 28):
            Number of hidden layers in the Transformer encoder.
        num_attention_heads (`int`, *optional*, defaults to 28):
            Number of attention heads for each attention layer in the Transformer encoder.
        num_key_value_heads (`int`, *optional*, defaults to 4):
            This is the number of key_value heads that should be used to implement Grouped Query Attention. If
            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
            `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When
            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
            by meanpooling all the original heads within that group. For more details checkout [this
            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to `32`.
        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
            The non-linear activation function (function or string) in the decoder.
        max_position_embeddings (`int`, *optional*, defaults to 32768):
            The maximum sequence length that this model might ever be used with.
        rms_norm_eps (`float`, *optional*, defaults to 1e-06):
            The epsilon used by the rms normalization layers.
        use_cache (`bool`, *optional*, defaults to `True`):
            Whether or not the model should return the last key/values attentions (not used by all models). Only
            relevant if `config.is_decoder=True`.
        rope_theta (`float`, *optional*, defaults to 1000000.0):
            The base period of the RoPE embeddings.
        use_sliding_window (`bool`, *optional*, defaults to `False`):
            Whether to use sliding window attention.
        sliding_window (`int`, *optional*, defaults to 32768):
            Sliding window attention (SWA) window size. If not specified, will default to `4096`.
        max_window_layers (`int`, *optional*, defaults to 28):
            The number of layers that use SWA (Sliding Window Attention). The bottom layers use SWA while the top use full attention.
        attention_dropout (`float`, *optional*, defaults to 0.0):
            The dropout ratio for the attention probabilities.
        rope_scaling (`Dict`, *optional*):
            Dictionary containing the scaling configuration for the RoPE embeddings. NOTE: if you apply new rope type
            and you expect the model to work on longer `max_position_embeddings`, we recommend you to update this value
            accordingly.
            Expected contents:
                `rope_type` (`str`):
                    The sub-variant of RoPE to use. Can be one of ['default', 'linear', 'dynamic', 'yarn', 'longrope',
                    'llama3'], with 'default' being the original RoPE implementation.
                `factor` (`float`, *optional*):
                    Used with all rope types except 'default'. The scaling factor to apply to the RoPE embeddings. In
                    most scaling types, a `factor` of x will enable the model to handle sequences of length x *
                    original maximum pre-trained length.
                `original_max_position_embeddings` (`int`, *optional*):
                    Used with 'dynamic', 'longrope' and 'llama3'. The original max position embeddings used during
                    pretraining.
                `attention_factor` (`float`, *optional*):
                    Used with 'yarn' and 'longrope'. The scaling factor to be applied on the attention
                    computation. If unspecified, it defaults to value recommended by the implementation, using the
                    `factor` field to infer the suggested value.
                `beta_fast` (`float`, *optional*):
                    Only used with 'yarn'. Parameter to set the boundary for extrapolation (only) in the linear
                    ramp function. If unspecified, it defaults to 32.
                `beta_slow` (`float`, *optional*):
                    Only used with 'yarn'. Parameter to set the boundary for interpolation (only) in the linear
                    ramp function. If unspecified, it defaults to 1.
                `short_factor` (`List[float]`, *optional*):
                    Only used with 'longrope'. The scaling factor to be applied to short contexts (<
                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
                    size divided by the number of attention heads divided by 2
                `long_factor` (`List[float]`, *optional*):
                    Only used with 'longrope'. The scaling factor to be applied to long contexts (<
                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
                    size divided by the number of attention heads divided by 2
                `low_freq_factor` (`float`, *optional*):
                    Only used with 'llama3'. Scaling factor applied to low frequency components of the RoPE
                `high_freq_factor` (`float`, *optional*):
                    Only used with 'llama3'. Scaling factor applied to high frequency components of the RoPE
        initializer_range (`float`, *optional*, defaults to 0.02):
            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.

    Example:

    ```python
    >>> from transformers import Qwen2_5OmniThinkerForConditionalGeneration, Qwen2_5OmniThinkerConfig, Qwen2_5OmniAudioEncoderConfig, Qwen2_5OmniVisionEncoderConfig

    >>> # Initializing a Qwen2_5OmniAudioEncoder config
    >>> audio_config = Qwen2_5OmniAudioEncoderConfig()

    >>> # Initializing a Qwen2_5OmniVisionEncoder config
    >>> vision_config = Qwen2_5OmniVisionEncoderConfig()

    >>> # Initializing a Qwen2.5OmniThinker configuration
    >>> configuration = Qwen2_5OmniThinkerConfig(audio_config, vision_config)

    >>> # Initializing a model from the Qwen-Omni style configuration
    >>> model = Qwen2_5OmniThinkerForConditionalGeneration(configuration)

    >>> # Accessing the model configuration
    >>> configuration = model.config
    ```"""

    model_type = "qwen2_5_omni_text"

    def __init__(
        self,
        vocab_size=152064,
        hidden_size=3584,
        intermediate_size=18944,
        num_hidden_layers=28,
        num_attention_heads=28,
        num_key_value_heads=4,
        rope_theta=1000000.0,
        sliding_window=32768,
        **super_kwargs,
    ):
        super().__init__(**super_kwargs)
        if self.rope_scaling is None:
            self.rope_scaling = {"mrope_section": [16, 24, 24], "rope_type": "default", "type": "default"}


class Qwen2_5OmniThinkerConfig(PretrainedConfig):
    r"""
    This is the configuration class to store the configuration of a [`Qwen2_5OmniThinkerForConditionalGeneration`]. It is used to instantiate an
    Qwen2.5-Omni-Thinker model according to the specified arguments, defining the model architecture. Instantiating a configuration
    with the defaults will yield a similar configuration to that of the Qwen2.5-Omni-Thinker.

    e.g. [Qwen/Qwen2.5-Omni-7B](https://huggingface.co/Qwen/Qwen2.5-Omni-7B)

    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
    documentation from [`PretrainedConfig`] for more information.

    Args:
        audio_config (`dict`,  *optional*):
            The config dictionary of the audio backbone.
        vision_config (`dict`, *optional*):
            The config dictionary of the vision backbone.
        text_config (`dict`, *optional*):
            The config dictionary of the text backbone.
        audio_token_index (`int`, *optional*, defaults to 151646):
            The audio token index to encode the audio prompt.
        image_token_index (`int`, *optional*, defaults to 151655):
            The image token index to encode the image prompt.
        video_token_index (`int`, *optional*, defaults to 151656):
            The video token index to encode the video prompt.
        position_id_per_seconds (`int`, *optional*, defaults to 25):
            The increment of position id per second.
        seconds_per_chunk (`int`, *optional*, defaults to 2):
            The duration in seconds of the chunk of audio and video data.
        audio_start_token_id (`int`, *optional*, defaults to 151647):
            The audio start token index to encode the audio prompt.
        audio_end_token_id (`int`, *optional*, defaults to 151648):
            The audio end token index to encode the audio prompt.
        user_token_id (`int, *optional*, defaults to 872):
            The user token index to encode the user token.
        initializer_range (`float`, *optional*, defaults to 0.02):
            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.

    Example:

    ```python
    >>> from transformers import Qwen2_5OmniThinkerForConditionalGeneration, Qwen2_5OmniThinkerConfig, Qwen2_5OmniAudioEncoderConfig, Qwen2_5OmniVisionEncoderConfig

    >>> # Initializing a Qwen2_5OmniAudioEncoder config
    >>> audio_config = Qwen2_5OmniAudioEncoderConfig()

    >>> # Initializing a Qwen2_5OmniVisionEncoder config
    >>> vision_config = Qwen2_5OmniVisionEncoderConfig()

    >>> # Initializing a Qwen2_5OmniTextConfig config
    >>> text_config = Qwen2_5OmniTextConfig()

    >>> # Initializing a Qwen2.5OmniThinker configuration
    >>> configuration = Qwen2_5OmniThinkerConfig(audio_config, vision_config, text_config)

    >>> # Initializing a model from the Qwen-Omni style configuration
    >>> model = Qwen2_5OmniThinkerForConditionalGeneration(configuration)

    >>> # Accessing the model configuration
    >>> configuration = model.config
    ```"""

    model_type = "qwen2_5_omni_thinker"
    attribute_map = {
        "image_token_id": "image_token_index",
        "video_token_id": "video_token_index",
        "audio_token_id": "audio_token_index",
    }
    sub_configs = {
        "audio_config": Qwen2_5OmniAudioEncoderConfig,
        "vision_config": Qwen2_5OmniVisionEncoderConfig,
        "text_config": Qwen2_5OmniTextConfig,
    }

    def __init__(
        self,
        audio_config=None,
        vision_config=None,
        text_config=None,
        audio_token_index=151646,
        image_token_index=151655,
        video_token_index=151656,
        position_id_per_seconds=25,
        seconds_per_chunk=2,
        audio_start_token_id=151647,
        audio_end_token_id=151648,
        user_token_id=872,
        initializer_range=0.02,
        **kwargs,
    ):
        self.audio_token_index = audio_token_index
        self.image_token_index = image_token_index
        self.video_token_index = video_token_index
        self.user_token_id = user_token_id
        self.position_id_per_seconds = position_id_per_seconds
        self.seconds_per_chunk = seconds_per_chunk
        self.audio_start_token_id = audio_start_token_id
        self.audio_end_token_id = audio_end_token_id
        self.initializer_range = initializer_range

        if isinstance(vision_config, dict):
            vision_config = Qwen2_5OmniVisionEncoderConfig(**vision_config)
        elif vision_config is None:
            vision_config = Qwen2_5OmniVisionEncoderConfig()
        self.vision_config = vision_config

        if isinstance(audio_config, dict):
            audio_config = Qwen2_5OmniAudioEncoderConfig(**audio_config)
        elif audio_config is None:
            audio_config = Qwen2_5OmniAudioEncoderConfig()
        self.audio_config = audio_config

        if isinstance(text_config, dict):
            text_config = Qwen2_5OmniTextConfig(**text_config)
        elif text_config is None:
            text_config = Qwen2_5OmniTextConfig()
        self.text_config = text_config

        super().__init__(**kwargs)


class Qwen2_5OmniTalkerConfig(PretrainedConfig):
    r"""
    This is the configuration class to store the configuration of a [`Qwen2_5OmniTalkerForConditionalGeneration`]. It is used to instantiate an
    Qwen2.5-Omni-Talker model according to the specified arguments, defining the model architecture. Instantiating a configuration
    with the defaults will yield a similar configuration to that of the Qwen2.5-Omni-Thinker.

    e.g. [Qwen/Qwen2.5-Omni-7B](https://huggingface.co/Qwen/Qwen2.5-Omni-7B)

    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
    documentation from [`PretrainedConfig`] for more information.

    Args:
        audio_token_index (`int`, *optional*, defaults to 151646):
            The audio token index to encode the audio prompt.
        image_token_index (`int`, *optional*, defaults to 151655):
            The image token index to encode the image prompt.
        video_token_index (`int`, *optional*, defaults to 151656):
            The video token index to encode the video prompt.
        vocab_size (`int`, *optional*, defaults to 8448):
            Vocabulary size of the QwenOmni model. Defines the number of different tokens that can be represented by the
            `inputs_ids` passed when calling [`Qwen2VLModel`]
        tts_text_start_token_id (`int`, *optional*, defaults to 151860):
            The tts text start token index to encode the start of tts text.
        tts_text_end_token_id (`int`, *optional*, defaults to 151861):
            The tts text end token index to encode the end of tts text.
        tts_text_pad_token_id (`int`, *optional*, defaults to 151859):
            The tts text pad token index to encode the pad of tts text.
        tts_codec_start_token_id (`int`, *optional*, defaults to 8293):
            The tts codec start token index to encode the start of tts codec.
        tts_codec_end_token_id (`int`, *optional*, defaults to 8294):
            The tts codec end token index to encode the end of tts codec.
        tts_codec_pad_token_id (`int`, *optional*, defaults to 8292):
            The tts codec pad token index to encode the pad of tts codec.
        tts_codec_mask_token_id (`int`, *optional*, defaults to 8296):
            The tts codec mask token index to encode the mask of tts codec.
        vision_start_token_id (`int`, *optional*, defaults to 151652):
            The tts vision start token index to encode the start of vision.
        vision_end_token_id (`int`, *optional*, defaults to 151653):
            The tts vision end token index to encode the end of vision.
        embedding_size (`int`, *optional*, defaults to 3584):
            Dimension of the embedding representations.
        hidden_size (`int`, *optional*, defaults to 3584):
            Dimension of the hidden representations.
        intermediate_size (`int`, *optional*, defaults to 18944):
            Dimension of the MLP representations.
        num_hidden_layers (`int`, *optional*, defaults to 28):
            Number of hidden layers in the Transformer encoder.
        num_attention_heads (`int`, *optional*, defaults to 28):
            Number of attention heads for each attention layer in the Transformer encoder.
        num_key_value_heads (`int`, *optional*, defaults to 4):
            This is the number of key_value heads that should be used to implement Grouped Query Attention. If
            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
            `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When
            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
            by meanpooling all the original heads within that group. For more details checkout [this
            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to `32`.
        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
            The non-linear activation function (function or string) in the decoder.
        max_position_embeddings (`int`, *optional*, defaults to 32768):
            The maximum sequence length that this model might ever be used with.
        rms_norm_eps (`float`, *optional*, defaults to 1e-06):
            The epsilon used by the rms normalization layers.
        head_dim (`int`, *optional*, defaults to 128):
            The dimension of each attention head.
        use_cache (`bool`, *optional*, defaults to `True`):
            Whether or not the model should return the last key/values attentions (not used by all models). Only
            relevant if `config.is_decoder=True`.
        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
            Whether the model's input and output word embeddings should be tied.
        rope_theta (`float`, *optional*, defaults to 1000000.0):
            The base period of the RoPE embeddings.
        use_sliding_window (`bool`, *optional*, defaults to `False`):
            Whether to use sliding window attention.
        sliding_window (`int`, *optional*, defaults to 32768):
            Sliding window attention (SWA) window size. If not specified, will default to `4096`.
        max_window_layers (`int`, *optional*, defaults to 28):
            The number of layers that use SWA (Sliding Window Attention). The bottom layers use SWA while the top use full attention.
        attention_dropout (`float`, *optional*, defaults to 0.0):
            The dropout ratio for the attention probabilities.
        rope_scaling (`Dict`, *optional*):
            Dictionary containing the scaling configuration for the RoPE embeddings. NOTE: if you apply new rope type
            and you expect the model to work on longer `max_position_embeddings`, we recommend you to update this value
            accordingly.
            Expected contents:
                `rope_type` (`str`):
                    The sub-variant of RoPE to use. Can be one of ['default', 'linear', 'dynamic', 'yarn', 'longrope',
                    'llama3'], with 'default' being the original RoPE implementation.
                `factor` (`float`, *optional*):
                    Used with all rope types except 'default'. The scaling factor to apply to the RoPE embeddings. In
                    most scaling types, a `factor` of x will enable the model to handle sequences of length x *
                    original maximum pre-trained length.
                `original_max_position_embeddings` (`int`, *optional*):
                    Used with 'dynamic', 'longrope' and 'llama3'. The original max position embeddings used during
                    pretraining.
                `attention_factor` (`float`, *optional*):
                    Used with 'yarn' and 'longrope'. The scaling factor to be applied on the attention
                    computation. If unspecified, it defaults to value recommended by the implementation, using the
                    `factor` field to infer the suggested value.
                `beta_fast` (`float`, *optional*):
                    Only used with 'yarn'. Parameter to set the boundary for extrapolation (only) in the linear
                    ramp function. If unspecified, it defaults to 32.
                `beta_slow` (`float`, *optional*):
                    Only used with 'yarn'. Parameter to set the boundary for interpolation (only) in the linear
                    ramp function. If unspecified, it defaults to 1.
                `short_factor` (`List[float]`, *optional*):
                    Only used with 'longrope'. The scaling factor to be applied to short contexts (<
                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
                    size divided by the number of attention heads divided by 2
                `long_factor` (`List[float]`, *optional*):
                    Only used with 'longrope'. The scaling factor to be applied to long contexts (<
                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
                    size divided by the number of attention heads divided by 2
                `low_freq_factor` (`float`, *optional*):
                    Only used with 'llama3'. Scaling factor applied to low frequency components of the RoPE
                `high_freq_factor` (`float`, *optional*):
                    Only used with 'llama3'. Scaling factor applied to high frequency components of the RoPE
        position_id_per_seconds (`int`, *optional*, defaults to 25):
            The increment of position id per second.
        seconds_per_chunk (`int`, *optional*, defaults to 2):
            The duration in seconds of the chunk of audio and video data.
        audio_start_token_id (`int`, *optional*, defaults to 151647):
            The audio start token index to encode the audio prompt.
        audio_end_token_id (`int`, *optional*, defaults to 151648):
            The audio end token index to encode the audio prompt.
        initializer_range (`float`, *optional*, defaults to 0.02):
            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
        spatial_merge_size (`int`, *optional*, defaults to 2):
            The size used for merging spatial dimensions.

    Example:

    ```python
    >>> from transformers import Qwen2_5OmniTalkerForConditionalGeneration, Qwen2_5OmniThinkerConfig, Qwen2_5OmniAudioEncoderConfig, Qwen2_5OmniVisionEncoderConfig

    >>> # Initializing a Qwen2_5OmniAudioEncoder config
    >>> audio_config = Qwen2_5OmniAudioEncoderConfig()

    >>> # Initializing a Qwen2 config
    >>> text_config = Qwen2Config()

    >>> # Initializing a Qwen2_5Omni configuration
    >>> configuration = Qwen2_5OmniThinkerConfig(audio_config, text_config)

    >>> # Initializing a model from the qwen2-audio style configuration
    >>> model = Qwen2_5OmniTalkerForConditionalGeneration(configuration)

    >>> # Accessing the model configuration
    >>> configuration = model.config
    ```"""

    model_type = "qwen2_5_omni_talker"
    attribute_map = {
        "image_token_id": "image_token_index",
        "video_token_id": "video_token_index",
        "audio_token_id": "audio_token_index",
    }

    def __init__(
        self,
        audio_token_index=151646,
        image_token_index=151655,
        video_token_index=151656,
        vocab_size=8448,
        tts_text_start_token_id=151860,
        tts_text_end_token_id=151861,
        tts_text_pad_token_id=151859,
        tts_codec_start_token_id=8293,
        tts_codec_end_token_id=8294,
        tts_codec_pad_token_id=8292,
        tts_codec_mask_token_id=8296,
        vision_start_token_id=151652,
        vision_end_token_id=151653,
        embedding_size=3584,
        hidden_size=3584,
        intermediate_size=18944,
        num_hidden_layers=28,
        num_attention_heads=28,
        num_key_value_heads=4,
        hidden_act="silu",
        max_position_embeddings=32768,
        rms_norm_eps=1e-06,
        head_dim=128,
        use_cache=True,
        tie_word_embeddings=False,
        rope_theta=1000000.0,
        use_sliding_window=False,
        sliding_window=32768,
        max_window_layers=28,
        attention_dropout=0.0,
        rope_scaling=None,
        position_id_per_seconds=25,
        seconds_per_chunk=2,
        audio_start_token_id=151647,
        audio_end_token_id=151648,
        initializer_range=0.02,
        spatial_merge_size=2,
        **kwargs,
    ):
        self.audio_token_index = audio_token_index
        self.image_token_index = image_token_index
        self.video_token_index = video_token_index

        self.tts_text_start_token_id = tts_text_start_token_id
        self.tts_text_end_token_id = tts_text_end_token_id
        self.tts_text_pad_token_id = tts_text_pad_token_id
        self.tts_codec_start_token_id = tts_codec_start_token_id
        self.tts_codec_end_token_id = tts_codec_end_token_id
        self.tts_codec_pad_token_id = tts_codec_pad_token_id

        self.tts_codec_mask_token_id = tts_codec_mask_token_id

        self.vision_start_token_id = vision_start_token_id
        self.vision_end_token_id = vision_end_token_id

        self.vocab_size = vocab_size
        self.head_dim = head_dim
        self.embedding_size = embedding_size
        self.max_position_embeddings = max_position_embeddings
        self.hidden_size = hidden_size
        self.intermediate_size = intermediate_size
        self.num_hidden_layers = num_hidden_layers
        self.num_attention_heads = num_attention_heads
        self.use_sliding_window = use_sliding_window
        self.sliding_window = sliding_window
        self.max_window_layers = max_window_layers

        # for backward compatibility
        if num_key_value_heads is None:
            num_key_value_heads = num_attention_heads

        self.num_key_value_heads = num_key_value_heads
        self.hidden_act = hidden_act
        self.rms_norm_eps = rms_norm_eps
        self.use_cache = use_cache
        self.rope_theta = rope_theta
        self.attention_dropout = attention_dropout
        self.rope_scaling = rope_scaling
        self.position_id_per_seconds = position_id_per_seconds  # zf
        self.seconds_per_chunk = seconds_per_chunk  # zf
        self.audio_start_token_id = audio_start_token_id  # zf
        self.audio_end_token_id = audio_end_token_id  # zf

        self.initializer_range = initializer_range
        self.spatial_merge_size = spatial_merge_size

        super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)


class Qwen2_5OmniDiTConfig(PretrainedConfig):
    r"""
    This is the configuration class to store the configuration of the Qwen2_5OmniToken2WavDiT used in the Qwen2.5-Omni-Token2Wav model.
    It defines the architecture of the DiT model, which is used for generating mel-spectrograms from tokens.

    Args:
        hidden_size (`int`, *optional*, defaults to 1024):
            The dimension of the model.
        num_hidden_layers (`int`, *optional*, defaults to 22):
            The number of transformer blocks in the DiT model.
        num_attention_heads (`int`, *optional*, defaults to 16):
            The number of attention heads in each transformer block.
        ff_mult (`int`, *optional*, defaults to 2):
            The multiplier for the feedforward layer in each transformer block.
        emb_dim (`int`, *optional*, defaults to 512):
            The dimension of the embedding layer.
        head_dim (`int`, *optional*, defaults to 64):
            The dimension of each attention head.
        repeats (`int`, *optional*, defaults to 2):
            The number of times the codec embeddings are repeated.
        num_embeds (`int`, *optional*, defaults to 8193):
            The number of unique embeddings in the codec.
        mel_dim (`int`, *optional*, defaults to 80):
            The dimension of the mel-spectrogram.
        dropout (`float`, *optional*, defaults to 0.1):
            The dropout rate for the transformer blocks.

        enc_emb_dim (`int`, *optional*, defaults to 192):
            The dimension of the pre-trained speaker embedding.
        enc_dim (`int`, *optional*, defaults to 128):
            The dimension of the encoder output.
        enc_channels (`List[int]`, *optional*, defaults to `[256, 256, 256, 256, 768]`):
            A list of output channels for each TDNN/SERes2Net layer in the encoder.
        enc_kernel_sizes (`List[int]`, *optional*, defaults to `[5, 3, 3, 3, 1]`):
            A list of kernel sizes for each layer in the encoder.
        enc_dilations (`List[int]`, *optional*, defaults to `[1, 2, 3, 4, 1]`):
            A list of dilations for each layer in the encoder.
        enc_attention_channels (`int`, *optional*, defaults to 64):
            The number of attention channels in the SqueezeExcitationBlock.
        enc_res2net_scale (`int`, *optional*, defaults to 2):
            The scale of the Res2Net block in the encoder.
        enc_se_channels (`int`, *optional*, defaults to 64):
            The number of output channels after squeeze in the SqueezeExcitationBlock.
    """

    model_type = "qwen2_5_omni_dit"

    def __init__(
        self,
        hidden_size=1024,
        num_hidden_layers=22,
        num_attention_heads=16,
        ff_mult=2,
        emb_dim=512,
        head_dim=64,
        rope_theta=10000.0,
        max_position_embeddings=32768,
        block_size=24,
        look_ahead_layers=[10],
        look_backward_layers=[0, 20],
        repeats=2,
        num_embeds=8193,
        mel_dim=80,
        dropout=0.1,
        enc_emb_dim=192,
        enc_dim=128,
        enc_channels=[256, 256, 256, 256, 768],
        enc_kernel_sizes=[5, 3, 3, 3, 1],
        enc_dilations=[1, 2, 3, 4, 1],
        enc_attention_channels=64,
        enc_res2net_scale=2,
        enc_se_channels=64,
        **kwargs,
    ):
        self.hidden_size = hidden_size
        self.num_hidden_layers = num_hidden_layers
        self.num_attention_heads = num_attention_heads
        self.ff_mult = ff_mult
        self.emb_dim = emb_dim
        self.head_dim = head_dim
        self.rope_theta = rope_theta
        self.max_position_embeddings = max_position_embeddings
        self.block_size = block_size
        self.look_ahead_layers = look_ahead_layers
        self.look_backward_layers = look_backward_layers
        self.repeats = repeats
        self.num_embeds = num_embeds
        self.mel_dim = mel_dim
        self.dropout = dropout
        self.enc_emb_dim = enc_emb_dim
        self.enc_dim = enc_dim
        self.enc_channels = enc_channels
        self.enc_kernel_sizes = enc_kernel_sizes
        self.enc_dilations = enc_dilations
        self.enc_attention_channels = enc_attention_channels
        self.enc_res2net_scale = enc_res2net_scale
        self.enc_se_channels = enc_se_channels
        super().__init__(**kwargs)


class Qwen2_5OmniBigVGANConfig(PretrainedConfig):
    r"""
    This is the configuration class to store the configuration of the Qwen2_5OmniToken2WavBigVGAN module used in the Qwen2.5-Omni-Token2Wav model.
    It defines the architecture of the BigVGAN model, which is used for converting mel-spectrograms to waveforms.

    Args:
        mel_dim (`int`, *optional*, defaults to 80):
            The dimension of the mel-spectrogram.
        upsample_initial_channel (`int`, *optional*, defaults to 1536):
            The number of channels in the initial upsampling layer.
        resblock_kernel_sizes (`List[int]`, *optional*, defaults to `[3, 7, 11]`):
            A list of kernel sizes for each residual block.
        resblock_dilation_sizes (`List[List[int]]`, *optional*, defaults to `[[1, 3, 5], [1, 3, 5], [1, 3, 5]]`):
            A list of dilation sizes for each residual block.
        upsample_rates (`List[int]`, *optional*, defaults to `[5, 3, 2, 2, 2, 2]`):
            A list of upsampling rates for each upsampling layer.
        upsample_kernel_sizes (`List[int]`, *optional*, defaults to `[11, 7, 4, 4, 4, 4]`):
            A list of kernel sizes for each upsampling layer.
    """

    model_type = "qwen2_5_omni_bigvgan"

    def __init__(
        self,
        mel_dim=80,
        upsample_initial_channel=1536,
        resblock_kernel_sizes=[3, 7, 11],
        resblock_dilation_sizes=[[1, 3, 5], [1, 3, 5], [1, 3, 5]],
        upsample_rates=[5, 3, 2, 2, 2, 2],
        upsample_kernel_sizes=[11, 7, 4, 4, 4, 4],
        **kwargs,
    ):
        self.mel_dim = mel_dim
        self.upsample_initial_channel = upsample_initial_channel
        self.resblock_kernel_sizes = resblock_kernel_sizes
        self.resblock_dilation_sizes = resblock_dilation_sizes
        self.upsample_rates = upsample_rates
        self.upsample_kernel_sizes = upsample_kernel_sizes
        super().__init__(**kwargs)


class Qwen2_5OmniToken2WavConfig(PretrainedConfig):
    r"""
    This is the configuration class to store the configuration of a [`Qwen2_5OmniToken2WavModel`].
    It is used to instantiate the Qwen2.5-Omni-Token2Wav model which combines a Diffusion Transformer (DiT) for mel-spectrogram generation with a BigVGAN model for waveform synthesis. The configuration contains sub-configurations for both components.

    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
    documentation from [`PretrainedConfig`] for more information.

    Args:
        dit_config ([`DiT_Args`], *optional*):
            Configuration class for the Diffusion Transformer (DiT) module responsible for generating mel-spectrograms.
        bigvgan_config ([`BigVGAN_Args`], *optional*):
            Configuration class for the BigVGAN module responsible for converting mel-spectrograms to waveforms.
    Example:

    ```python
    >>> from transformers import Qwen2_5OmniToken2WavModel, DiT_Args, BigVGAN_Args

    >>> # Initialize DiT configuration
    >>> dit_config = DiT_Args(
    ...     dim=1024,
    ...     depth=22,
    ...     heads=16,
    ...     ff_mult=2
    ... )

    >>> # Initialize BigVGAN configuration
    >>> bigvgan_config = BigVGAN_Args(
    ...     mel_dim=80,
    ...     upsample_rates=[5,3,2,2,2,2]
    ... )

    >>> # Initialize main configuration
    >>> config = Qwen2_5OmniToken2WavConfig(dit_config, bigvgan_config)

    >>> # Initialize model with config
    >>> model = Qwen2_5OmniToken2Wav(config)

    >>> # Accessing the model configuration
    >>> configuration = model.config
    ```
    """

    model_type = "qwen2_5_omni_token2wav"
    sub_configs = {
        "dit_config": Qwen2_5OmniDiTConfig,
        "bigvgan_config": Qwen2_5OmniBigVGANConfig,
    }

    def __init__(self, dit_config=None, bigvgan_config=None, **kwargs):
        if dit_config is None:
            dit_config = {}
        if bigvgan_config is None:
            bigvgan_config = {}
        self.dit_config = Qwen2_5OmniDiTConfig(**dit_config)
        self.bigvgan_config = Qwen2_5OmniBigVGANConfig(**bigvgan_config)
        super().__init__(**kwargs)


class Qwen2_5OmniConfig(PretrainedConfig):
    """
    This is the configuration class to store the configuration of a [`Qwen2_5OmniForConditionalGeneration`]. It is used to instantiate a Qwen2.5Omni
    model according to the specified sub-models configurations, defining the model architecture.

    Instantiating a configuration with the defaults will yield a similar configuration to that of the
    [Qwen/Qwen2.5-Omni-7B](https://huggingface.co/Qwen/Qwen2.5-Omni-7B) architecture.

    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
    documentation from [`PretrainedConfig`] for more information.

    Args:
        thinker_config (`dict`, *optional*): Configuration of the underlying thinker sub-model.
        talker_config (`dict`, *optional*): Configuration of the underlying talker sub-model.
        token2wav_config (`dict`, *optional*): Configuration of the underlying codec sub-model.
        enable_audio_output (`bool`, *optional*, defaults to `True`): Whether enable audio output and load talker and token2wav module.

    Example:

    ```python
    >>> from transformers import (
    ...     Qwen2_5OmniThinkerConfig,
    ...     Qwen2_5OmniTalkerConfig,
    ...     Qwen2_5OmniToken2WavConfig,
    ...     Qwen2_5OmniForConditionalGeneration,
    ...     Qwen2_5OmniConfig,
    ... )

    >>> # Initializing sub-modules configurations.
    >>> thinker_config = Qwen2_5OmniThinkerConfig()
    >>> talker_config = Qwen2_5OmniTalkerConfig()
    >>> token2wav_config = Qwen2_5OmniToken2WavConfig()


    >>> # Initializing a module style configuration
    >>> configuration = Qwen2_5OmniConfig.from_sub_model_configs(
    ...     thinker_config, talker_config, token2wav_config
    ... )

    >>> # Initializing a model (with random weights)
    >>> model = Qwen2_5OmniForConditionalGeneration(configuration)

    >>> # Accessing the model configuration
    >>> configuration = model.config
    ```
    """

    model_type = "qwen2_5_omni"
    sub_configs = {
        "thinker_config": Qwen2_5OmniThinkerConfig,
        "talker_config": Qwen2_5OmniTalkerConfig,
        "token2wav_config": Qwen2_5OmniToken2WavConfig,
    }

    def __init__(
        self,
        thinker_config=None,
        talker_config=None,
        token2wav_config=None,
        enable_audio_output: bool = True,
        **kwargs,
    ):
        if thinker_config is None:
            thinker_config = {}
            logger.info("thinker_config is None. Initializing thinker model with default values")

        if talker_config is None:
            talker_config = {}
            logger.info("talker_config is None. Initializing talker model with default values")

        if token2wav_config is None:
            token2wav_config = {}
            logger.info("token2wav_config is None. Initializing token2wav model with default values")

        self.thinker_config = Qwen2_5OmniThinkerConfig(**thinker_config)
        self.talker_config = Qwen2_5OmniTalkerConfig(**talker_config)
        self.token2wav_config = Qwen2_5OmniToken2WavConfig(**token2wav_config)
        self.enable_audio_output = enable_audio_output

        super().__init__(**kwargs)

    def get_text_config(self, decoder=False) -> "PretrainedConfig":
        """
        Returns the config that is meant to be used with text IO. On most models, it is the original config instance
        itself. On specific composite models, it is under a set of valid names.

        Args:
            decoder (`Optional[bool]`, *optional*, defaults to `False`):
                If set to `True`, then only search for decoder config names.
        """
        # Overridden for deeply nested config like Qwen2-Omni. We don't have any omni model
        # except for Qwen yet. This has to be generalized if more deeply nested configs are
        # added. NOTE: currently method used only by vLLM
        return self.thinker_config.get_text_config()


class Qwen2_5OmniPreTrainedModel(Qwen2_5_VLPreTrainedModel):
    config_class = Qwen2_5OmniConfig
    _supports_static_cache = True

    def _init_weights(self, module):
        # important: this ported version of Qwen2.5OmniThinker isn't meant for training from scratch - only
        # inference and fine-tuning - so the proper init weights code has been removed
        std = self.config.initializer_range if hasattr(self.config, "initializer_range") else 0.02

        if isinstance(module, (nn.Linear, nn.Conv1d, nn.Conv3d, nn.ConvTranspose1d)):
            module.weight.data.normal_(mean=0.0, std=std)
            if module.bias is not None:
                module.bias.data.zero_()
        elif isinstance(module, nn.Embedding):
            module.weight.data.normal_(mean=0.0, std=std)
            if module.padding_idx is not None:
                module.weight.data[module.padding_idx].zero_()
        elif isinstance(module, nn.LayerNorm):
            if module.weight is not None:
                module.weight.data.fill_(1.0)
            if module.bias is not None:
                module.bias.data.zero_()
        elif isinstance(module, Qwen2RMSNorm):
            module.weight.data.fill_(1.0)


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 None:
                attention_mask = torch.ones_like(total_input_ids)
            position_ids = torch.ones(
                3,
                input_ids.shape[0],
                input_ids.shape[1],
                dtype=input_ids.dtype,
                device=input_ids.device,
            )
            image_idx, video_idx, audio_idx = 0, 0, 0
            attention_mask = attention_mask.to(total_input_ids.device)
            for i, input_ids in enumerate(total_input_ids):
                input_ids = input_ids[attention_mask[i] == 1]
                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)

                position_ids[..., i, attention_mask[i] == 1] = 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, device=input_ids.device).unsqueeze(1)

            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
class Qwen2_5OmniThinkerCausalLMOutputWithPast(ModelOutput):
    """
    Base class for Qwen2.5OmniThinker causal language model (or autoregressive) outputs.

    Args:
        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 (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
            Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
            `(batch_size, num_heads, sequence_length, embed_size_per_head)`)

            Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
            `past_key_values` input) to speed up sequential decoding.
        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.

            Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
        attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
            sequence_length)`.

            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
            heads.
        rope_deltas (`torch.LongTensor` of shape `(batch_size, )`, *optional*):
            The rope index difference between sequence length and multimodal rope.
    """

    loss: Optional[torch.FloatTensor] = None
    logits: Optional[torch.FloatTensor] = None
    past_key_values: Optional[List[torch.FloatTensor]] = None
    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
    attentions: Optional[Tuple[torch.FloatTensor]] = None
    rope_deltas: Optional[torch.LongTensor] = None


class 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.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.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,
    ) -> 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)
        key_states = key_states.transpose(0, 1)
        value_states = value_states.transpose(0, 1)
        attn_weights = torch.matmul(query_states, key_states.transpose(1, 2)) / math.sqrt(self.head_dim)

        attention_mask = torch.full(
            [1, seq_length, key_states.shape[1]],
            torch.finfo(query_states.dtype).min,
            device=query_states.device,
            dtype=query_states.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

        attn_weights = attn_weights + attention_mask

        attn_weights = nn.functional.softmax(attn_weights, dim=-1).to(query_states.dtype)

        attn_output = torch.matmul(attn_weights, value_states).transpose(0, 1).reshape(seq_length, self.embed_dim)

        # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
        # partitioned across GPUs when using tensor-parallelism.

        attn_output = self.out_proj(attn_output)

        return attn_output


class Qwen2_5OmniAudioFlashAttention2(Qwen2_5OmniAudioAttention):
    """
    Qwen2.5OmniThinker flash attention module. This module inherits from `Qwen2_5OmniAudioAttention` as the weights of the module stays
    untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
    flash attention and deal with padding tokens in case the input contains any of them.
    """

    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)

        # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
        # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignment, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
        # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
        self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()

    def forward(
        self,
        hidden_states: torch.Tensor,
        cu_seqlens: Optional[torch.Tensor] = None,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        seq_length, all_dim = hidden_states.size()
        query_states = self.q_proj(hidden_states)
        query_states = query_states.reshape(seq_length, self.num_heads, -1)

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

        max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max().item()
        attn_output = flash_attn_varlen_func(
            query_states, key_states, value_states, cu_seqlens, cu_seqlens, max_seqlen, max_seqlen, dropout_p=0.0
        )
        attn_output = attn_output.reshape(seq_length, all_dim)
        attn_output = self.out_proj(attn_output)

        return attn_output


class Qwen2_5OmniAudioSdpaAttention(Qwen2_5OmniAudioAttention):
    def forward(
        self,
        hidden_states: torch.Tensor,
        cu_seqlens: Optional[torch.Tensor] = None,
    ) -> 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)

        attention_mask = torch.zeros(
            [1, seq_length, key_states.shape[0]], device=query_states.device, dtype=torch.bool
        )
        for i in range(1, len(cu_seqlens)):
            attention_mask[..., cu_seqlens[i - 1] : cu_seqlens[i], cu_seqlens[i - 1] : cu_seqlens[i]] = True

        # We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment
        # in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling.
        # The tgt_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case tgt_len == 1.
        query_states = query_states.transpose(0, 1)
        key_states = key_states.transpose(0, 1)
        value_states = value_states.transpose(0, 1)

        # NOTE: SDPA with memory-efficient backend is currently (torch==2.1.2) bugged when using non-contiguous inputs and a custom attn_mask,
        # but we are fine here as `_shape` do call `.contiguous()`. Reference: https://github.com/pytorch/pytorch/issues/112577
        attn_output = torch.nn.functional.scaled_dot_product_attention(
            query_states,
            key_states,
            value_states,
            attn_mask=attention_mask,
            dropout_p=self.dropout if self.training else 0.0,
        )
        attn_output = attn_output.transpose(0, 1)
        # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
        # partitioned across GPUs when using tensor-parallelism.
        attn_output = attn_output.reshape(seq_length, self.embed_dim)
        attn_output = self.out_proj(attn_output)
        return attn_output


QWEN2_5_OMNI_AUDIO_ATTENTION_CLASSES = {
    "eager": Qwen2_5OmniAudioAttention,
    "flash_attention_2": Qwen2_5OmniAudioFlashAttention2,
    "sdpa": Qwen2_5OmniAudioSdpaAttention,
}


class Qwen2_5OmniAudioEncoderLayer(Qwen2AudioEncoderLayer):
    def __init__(self, config: Qwen2_5OmniAudioEncoderConfig):
        super().__init__(config)
        self.self_attn = QWEN2_5_OMNI_AUDIO_ATTENTION_CLASSES[config._attn_implementation](config)

    def forward(
        self,
        hidden_states: torch.Tensor,
        cu_seqlens: torch.Tensor,
    ) -> torch.Tensor:
        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,
        )
        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_class = 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

    @auto_docstring
    def forward(
        self,
        input_features,
        feature_lens=None,
        aftercnn_lens=None,
    ):
        r"""
        input_features (`torch.LongTensor` of shape `(batch_size, feature_size, sequence_length)`):
            Float values of mel features extracted from the raw speech waveform. Raw speech waveform can be
            obtained by loading a `.flac` or `.wav` audio file into an array of type `List[float]` or a
            `numpy.ndarray`, *e.g.* via the soundfile library (`pip install soundfile`). To prepare the array into
            `input_features`, the [`AutoFeatureExtractor`] should be used for extracting the mel features, padding
            and conversion into a tensor of type `torch.FloatTensor`. See [`~WhisperFeatureExtractor.__call__`]
        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)

        for idx, encoder_layer in enumerate(self.layers):
            if self.gradient_checkpointing and self.training:
                layer_outputs = self._gradient_checkpointing_func(
                    encoder_layer.__call__,
                    hidden_states,
                    cu_seqlens,
                )
            else:
                layer_outputs = encoder_layer(
                    hidden_states,
                    cu_seqlens,
                )

            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 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, dim: int, num_heads: int = 16) -> None:
        super().__init__()
        self.num_heads = num_heads
        self.head_dim = dim // num_heads
        self.q = nn.Linear(dim, dim, bias=True)
        self.k = nn.Linear(dim, dim, bias=True)
        self.v = nn.Linear(dim, dim, bias=True)
        self.proj = nn.Linear(dim, dim)

    def forward(
        self, hidden_states: torch.Tensor, cu_seqlens: torch.Tensor, rotary_pos_emb: torch.Tensor = None
    ) -> torch.Tensor:
        seq_length = hidden_states.shape[0]
        q = self.q(hidden_states).reshape(seq_length, self.num_heads, -1)
        k = self.k(hidden_states).reshape(seq_length, self.num_heads, -1)
        v = self.v(hidden_states).reshape(seq_length, self.num_heads, -1)
        q = apply_rotary_pos_emb_vision(q.unsqueeze(0), rotary_pos_emb).squeeze(0)
        k = apply_rotary_pos_emb_vision(k.unsqueeze(0), rotary_pos_emb).squeeze(0)

        attention_mask = torch.full(
            [1, seq_length, seq_length], torch.finfo(q.dtype).min, device=q.device, dtype=q.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

        q = q.transpose(0, 1)
        k = k.transpose(0, 1)
        v = v.transpose(0, 1)
        attn_weights = torch.matmul(q, k.transpose(1, 2)) / math.sqrt(self.head_dim)
        attn_weights = attn_weights + attention_mask
        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(q.dtype)
        attn_output = torch.matmul(attn_weights, v)
        attn_output = attn_output.transpose(0, 1)
        attn_output = attn_output.reshape(seq_length, -1)
        attn_output = self.proj(attn_output)
        return attn_output


class Qwen2_5OmniVisionFlashAttention2(nn.Module):
    def __init__(self, dim: int, num_heads: int = 16) -> None:
        super().__init__()
        self.num_heads = num_heads
        self.q = nn.Linear(dim, dim, bias=True)
        self.k = nn.Linear(dim, dim, bias=True)
        self.v = nn.Linear(dim, dim, bias=True)
        self.proj = nn.Linear(dim, dim)

    def _apply_rotary_pos_emb_flashatt(self, tensor: torch.Tensor, freqs: torch.Tensor) -> torch.Tensor:
        tensor_ = tensor.float()
        cos = freqs.cos()  # .type_as(tensor_)
        sin = freqs.sin()  # .type_as(tensor_)
        output = apply_rotary_emb(tensor_, cos, sin).type_as(tensor)
        return output

    def forward(
        self, hidden_states: torch.Tensor, cu_seqlens: torch.Tensor, rotary_pos_emb: torch.Tensor = None
    ) -> torch.Tensor:
        seq_length = hidden_states.shape[0]
        q = self.q(hidden_states).reshape(seq_length, self.num_heads, -1)
        k = self.k(hidden_states).reshape(seq_length, self.num_heads, -1)
        v = self.v(hidden_states).reshape(seq_length, self.num_heads, -1)
        q = self._apply_rotary_pos_emb_flashatt(q.unsqueeze(0), rotary_pos_emb).squeeze(0)
        k = self._apply_rotary_pos_emb_flashatt(k.unsqueeze(0), rotary_pos_emb).squeeze(0)

        max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max().item()
        attn_output = flash_attn_varlen_func(q, k, v, cu_seqlens, cu_seqlens, max_seqlen, max_seqlen).reshape(
            seq_length, -1
        )
        attn_output = self.proj(attn_output)
        return attn_output


class Qwen2_5OmniVisionSdpaAttention(nn.Module):
    def __init__(self, dim: int, num_heads: int = 16) -> None:
        super().__init__()
        self.num_heads = num_heads
        self.q = nn.Linear(dim, dim, bias=True)
        self.k = nn.Linear(dim, dim, bias=True)
        self.v = nn.Linear(dim, dim, bias=True)
        self.proj = nn.Linear(dim, dim)

    def forward(
        self, hidden_states: torch.Tensor, cu_seqlens: torch.Tensor, rotary_pos_emb: torch.Tensor = None
    ) -> torch.Tensor:
        seq_length = hidden_states.shape[0]
        q = self.q(hidden_states).reshape(seq_length, self.num_heads, -1)
        k = self.k(hidden_states).reshape(seq_length, self.num_heads, -1)
        v = self.v(hidden_states).reshape(seq_length, self.num_heads, -1)
        q = apply_rotary_pos_emb_vision(q.unsqueeze(0), rotary_pos_emb).squeeze(0)
        k = apply_rotary_pos_emb_vision(k.unsqueeze(0), rotary_pos_emb).squeeze(0)

        attention_mask = torch.zeros([1, seq_length, seq_length], device=q.device, dtype=torch.bool)
        for i in range(1, len(cu_seqlens)):
            attention_mask[..., cu_seqlens[i - 1] : cu_seqlens[i], cu_seqlens[i - 1] : cu_seqlens[i]] = True
        q = q.transpose(0, 1)
        k = k.transpose(0, 1)
        v = v.transpose(0, 1)
        attn_output = F.scaled_dot_product_attention(q, k, v, attention_mask, dropout_p=0.0)
        attn_output = attn_output.transpose(0, 1)
        attn_output = attn_output.reshape(seq_length, -1)
        attn_output = self.proj(attn_output)
        return attn_output


QWEN2_5_OMNI_VISION_ATTENTION_CLASSES = {
    "eager": Qwen2_5OmniVisionAttention,
    "flash_attention_2": Qwen2_5OmniVisionFlashAttention2,
    "sdpa": Qwen2_5OmniVisionSdpaAttention,
}


class Qwen2_5OmniVisionBlock(Qwen2_5_VLVisionBlock):
    def __init__(self, config: Qwen2_5OmniVisionEncoderConfig) -> None:
        super().__init__(config, config._attn_implementation)
        self.attn = QWEN2_5_OMNI_VISION_ATTENTION_CLASSES[config._attn_implementation](
            config.hidden_size, num_heads=config.num_heads
        )

    def forward(self, hidden_states, cu_seqlens, rotary_pos_emb) -> torch.Tensor:
        hidden_states = hidden_states + self.attn(
            self.norm1(hidden_states), cu_seqlens=cu_seqlens, rotary_pos_emb=rotary_pos_emb
        )
        hidden_states = hidden_states + self.mlp(self.norm2(hidden_states))
        return hidden_states


class Qwen2_5OmniVisionEncoder(Qwen2_5_VisionTransformerPretrainedModel):
    config_class = Qwen2_5OmniVisionEncoderConfig
    _no_split_modules = ["Qwen2_5OmniVisionBlock"]

    def __init__(self, config: Qwen2_5OmniVisionEncoderConfig, *inputs, **kwargs) -> None:
        super().__init__(config, *inputs, **kwargs)
        self.blocks = nn.ModuleList([Qwen2_5OmniVisionBlock(config) for _ in range(config.depth)])

    def forward(self, hidden_states: torch.Tensor, grid_thw: torch.Tensor) -> 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
            if self.gradient_checkpointing and self.training:
                hidden_states = self._gradient_checkpointing_func(
                    blk.__call__, hidden_states, cu_seqlens_now, rotary_pos_emb
                )
            else:
                hidden_states = blk(
                    hidden_states,
                    cu_seqlens=cu_seqlens_now,
                    rotary_pos_emb=rotary_pos_emb,
                )
        hidden_states = self.merger(hidden_states)
        reverse_indices = torch.argsort(window_index)
        hidden_states = hidden_states[reverse_indices, :]

        return hidden_states


class Qwen2_5OmniRotaryEmbedding(Qwen2VLRotaryEmbedding):
    def __init__(self, config: Qwen2_5OmniThinkerConfig, device=None):
        super().__init__(config, device)


# It's same as `Qwen2_5_VLAttention`, but talker model's hidden_size isn't divisible by num_heads.
# Removes the value error as a workaround.
class Qwen2_5OmniAttention(Qwen2_5_VLAttention, nn.Module):
    def __init__(self, config: Qwen2_5OmniConfig, layer_idx: Optional[int] = None):
        nn.Module.__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.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.rotary_emb = Qwen2_5OmniRotaryEmbedding(config=config)


class Qwen2MLP(Qwen2_5_VLMLP):
    pass


class Qwen2_5OmniThinkerTextModel(Qwen2_5_VLTextModel):
    config_class = Qwen2_5OmniTextConfig
    _no_split_modules = ["Qwen2_5OmniDecoderLayer"]

    def __init__(self, config: Qwen2_5OmniTextConfig):
        super().__init__(config)


@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_class = Qwen2_5OmniThinkerConfig
    base_model_prefix = "thinker"
    _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, attn_implementation=config._attn_implementation
        )

        self.visual = Qwen2_5OmniVisionEncoder._from_config(
            config.vision_config, attn_implementation=config._attn_implementation
        )

        self.vocab_size = config.text_config.vocab_size
        self.model = Qwen2_5OmniThinkerTextModel._from_config(
            config.text_config, attn_implementation=config._attn_implementation
        )
        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

    @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[List[torch.FloatTensor]] = 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,
    ) -> Union[Tuple, Qwen2_5OmniThinkerCausalLMOutputWithPast]:
        r"""
        input_features (`torch.FloatTensor` of shape `(batch_size, feature_size, feature_sequence_length)`):
            Float values mel features extracted from the raw speech waveform. Raw speech waveform can be obtained by
            loading a `.flac` or `.wav` audio file into an array of type `List[float]` or a `numpy.ndarray`, *e.g.* via
            the soundfile library (`pip install soundfile`). To prepare the array into `input_features`, the
            [`AutoFeatureExtractor`] should be used for extracting the mel features, padding and conversion into a
            tensor of type `torch.FloatTensor`. See [`~WhisperFeatureExtractor.__call__`]
        pixel_values_videos (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size), *optional*):
            The tensors corresponding to the input videos. Pixel values can be obtained using
            [`AutoImageProcessor`]. See [`SiglipImageProcessor.__call__`] for details ([]`NewTaskModelProcessor`] uses
            [`SiglipImageProcessor`] for processing videos).
        image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
            The temporal, height and width of feature shape of each image in LLM.
        video_grid_thw (`torch.LongTensor` of shape `(num_videos, 3)`, *optional*):
            The temporal, height and width of feature shape of each video in LLM.
        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_ids is not None and input_ids.shape[1] != 1:  # Prefill stage
            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_mask = (
                    (input_ids == self.config.audio_token_id)
                    .unsqueeze(-1)
                    .expand_as(inputs_embeds)
                    .to(inputs_embeds.device)
                )
                audio_features = audio_features.to(inputs_embeds.device, inputs_embeds.dtype)
                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_mask = (
                    (input_ids == self.config.image_token_id)
                    .unsqueeze(-1)
                    .expand_as(inputs_embeds)
                    .to(inputs_embeds.device)
                )
                image_embeds = image_embeds.to(inputs_embeds.device, inputs_embeds.dtype)
                inputs_embeds = inputs_embeds.masked_scatter(image_mask, image_embeds)

            if pixel_values_videos is not None:
                video_embeds = self.get_video_features(pixel_values_videos, video_grid_thw)
                video_mask = (
                    (input_ids == self.config.video_token_id)
                    .unsqueeze(-1)
                    .expand_as(inputs_embeds)
                    .to(inputs_embeds.device)
                )
                video_embeds = video_embeds.to(inputs_embeds.device, inputs_embeds.dtype)
                inputs_embeds = inputs_embeds.masked_scatter(video_mask, video_embeds)

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

        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,
        )

        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

        return model_inputs


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


@dataclass
class Qwen2_5OmniTalkerCausalLMOutputWithPast(ModelOutput):
    """
    Base class for Qwen2.5OmniTalker causal language model (or autoregressive) outputs.

    Args:
        loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
            Language modeling loss (for next-token prediction).
        logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
        past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
            Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
            `(batch_size, num_heads, sequence_length, embed_size_per_head)`)

            Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
            `past_key_values` input) to speed up sequential decoding.
        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.

            Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
        attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
            sequence_length)`.

            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
            heads.
        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: torch.FloatTensor = None
    past_key_values: Optional[List[torch.FloatTensor]] = None
    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
    attentions: Optional[Tuple[torch.FloatTensor]] = None
    rope_deltas: Optional[torch.LongTensor] = None
    thinker_reply_part: torch.FloatTensor = None


class Qwen2_5OmniTalkerModel(Qwen2_5_VLTextModel):
    config_class = Qwen2_5OmniTalkerConfig
    _no_split_modules = ["Qwen2_5OmniTalkerDecoderLayer"]

    def __init__(self, config: Qwen2_5OmniTalkerConfig):
        super().__init__(config)
        self.embed_tokens = nn.Embedding(config.vocab_size, config.embedding_size, self.padding_idx)


class Qwen2_5OmniTalkerForConditionalGeneration(Qwen2_5OmniPreTrainedModelForConditionalGeneration, GenerationMixin):
    config_class = 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: torch.LongTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[List[torch.FloatTensor]] = 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"""
        rope_deltas (`torch.LongTensor` of shape `(batch_size, )`, *optional*):
            The rope index difference between sequence length and multimodal rope.
        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.
        audio_feature_lengths (`torch.LongTensor` of shape `(num_audios)`, *optional*):
            The length of feature shape of each audio in LLM.
        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.
        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.
        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 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):
    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 isinstance(device_type, str) and 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)


# 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 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://arxiv.org/abs/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


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._attn_implementation = config._attn_implementation
        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._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):  # noqa: F821
        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://arxiv.org/abs/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 ValueError("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_class = 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_class = 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_class = 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_class = 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("Speaker {} loaded".format(list(self.speaker_map.keys())))

    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(self.talker.device)
        if thinker_kwargs.get("input_features", None) 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).to(embeds_to_talker.device)
            audio_mask_tensor = torch.zeros(
                [audio_ids_mask.sum(), embeds_to_talker.shape[-1]],
                dtype=embeds_to_talker.dtype,
                device=self.talker.device,
            )
            embeds_to_talker.masked_scatter_(audio_mask, audio_mask_tensor)
        if thinker_kwargs.get("pixel_values", None) 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).to(embeds_to_talker.device)
            image_mask_tensor = torch.zeros(
                [image_ids_mask.sum(), embeds_to_talker.shape[-1]],
                dtype=embeds_to_talker.dtype,
                device=self.talker.device,
            )
            embeds_to_talker.masked_scatter_(image_mask, image_mask_tensor)
        if thinker_kwargs.get("pixel_values_videos", None) 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).to(embeds_to_talker.device)
            video_mask_tensor = torch.zeros(
                [video_ids_mask.sum(), embeds_to_talker.shape[-1]],
                dtype=embeds_to_talker.dtype,
                device=self.talker.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(self.talker.device)
        thinker_token_embeds = [
            token_hidden_states[0].to(self.talker.device) for token_hidden_states in processed_thinker_hidden
        ]
        thinker_hidden_states = [
            token_hidden_states[-1].to(self.talker.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.to(self.talker.device),
                torch.tensor([[talker_text_bos_token]], dtype=torch.long, device=self.talker.device),
                thinker_generate_ids[:, :1],
            ],
            dim=-1,
        )

        talker_input_ids = torch.cat(
            [
                torch.full_like(input_ids, fill_value=self.talker.codec_mask_token, device=self.talker.device),
                torch.tensor([[self.talker.codec_pad_token]], dtype=torch.long, device=self.talker.device),
                torch.tensor([[self.talker.codec_bos_token]], dtype=torch.long, device=self.talker.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=self.thinker.device)
        talker_text_bos_embed = thinker_embed_tokens(talker_text_bos_token).to(self.talker.device)
        talker_inputs_embeds = torch.cat(
            [
                talker_inputs_embeds,
                talker_text_bos_embed,
                thinker_reply_part[:, :1, :],
            ],
            dim=1,
        )

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

        pad_embedding = thinker_embed_tokens(
            torch.tensor([[self.talker.text_pad_token]], dtype=torch.long, device=self.thinker.device)
        ).to(self.talker.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(self.talker.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(self.talker.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(self.token2wav.device),
            conditioning=speaker_params["cond"].to(self.token2wav.device).float(),
            reference_mel=speaker_params["ref_mel"].to(self.token2wav.device).float(),
            **token2wav_kwargs,
        )

        return thinker_result.sequences, wav.float()


__all__ = [
    "Qwen2_5OmniConfig",
    "Qwen2_5OmniThinkerConfig",
    "Qwen2_5OmniTalkerConfig",
    "Qwen2_5OmniToken2WavConfig",
    "Qwen2_5OmniForConditionalGeneration",
    "Qwen2_5OmniThinkerTextModel",
    "Qwen2_5OmniThinkerForConditionalGeneration",
    "Qwen2_5OmniTalkerModel",
    "Qwen2_5OmniTalkerForConditionalGeneration",
    "Qwen2_5OmniToken2WavDiTModel",
    "Qwen2_5OmniToken2WavBigVGANModel",
    "Qwen2_5OmniToken2WavModel",
    "Qwen2_5OmniPreTrainedModel",
    "Qwen2_5OmniPreTrainedModelForConditionalGeneration",
]