modeling_hithinkomni.py

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

import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.utils.checkpoint
from torch.nn import CrossEntropyLoss, BCEWithLogitsLoss

from transformers.activations import ACT2FN
from transformers.cache_utils import Cache, EncoderDecoderCache, SlidingWindowCache, StaticCache
from transformers.generation import GenerationMixin
from transformers.modeling_attn_mask_utils import (
    AttentionMaskConverter,
)
from transformers.modeling_outputs import (
    BaseModelOutputWithPast,
    BaseModelOutput,
    ModelOutput,
)
from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS
from transformers.modeling_utils import PreTrainedModel
from transformers.cache_utils import DynamicCache
from transformers.utils import (
    add_start_docstrings,
    add_start_docstrings_to_model_forward,
    is_flash_attn_2_available,
    is_flash_attn_greater_or_equal_2_10,
    logging,
    replace_return_docstrings,
)

# from seallm 
from transformers.generation.logits_process import (
    RepetitionPenaltyLogitsProcessor,
    TopKLogitsWarper,
    TopPLogitsWarper,
    TemperatureLogitsWarper,
    ExponentialDecayLengthPenalty
)

# import config class
from .configuration_hithinkomni import HithinkOmniConfig, HithinkOmniVisionConfig, HithinkAudioEncoderConfig

if is_flash_attn_2_available():
    from flash_attn import flash_attn_varlen_func

    from transformers.modeling_flash_attention_utils import _flash_attention_forward
else:
    flash_attn_varlen_func = None

try:
    from flash_attn.layers.rotary import apply_rotary_emb_func
except ImportError:
    apply_rotary_emb_func = None

try:
    from flash_attn.ops.rms_norm import dropout_add_rms_norm
except ImportError:
    dropout_add_rms_norm = None

try:
    from flash_attn.ops.activations import swiglu
except ImportError:
    swiglu = None

if torch.cuda.is_available():
    try:
        from flash_attn.losses.cross_entropy import CrossEntropyLoss
    except ImportError:
        pass

logger = logging.get_logger(__name__)

_CONFIG_FOR_DOC = "HithinkOmniConfig"


@dataclass
class HithinkOmniCausalLMOutputWithPast(ModelOutput):
    """
    Base class for HithinkOmni 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
    audio_past_key_values: Optional[Cache] = None
    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
    attentions: Optional[Tuple[torch.FloatTensor]] = None
    rope_deltas: Optional[torch.LongTensor] = None

# audio part start
class CausalConv1d(nn.Conv1d):
    def __init__(
        self,
        in_channels,
        out_channels,
        kernel_size,
        stride=1,
        padding=0,
        dilation=1,
        groups=1,
        bias=True,
        **kwargs
    ):
        super(CausalConv1d, self).__init__(
            in_channels,
            out_channels,
            kernel_size,
            stride=stride,
            padding=0,
            dilation=dilation,
            groups=groups,
            bias=bias,
            **kwargs
        )

        self.left_padding = dilation * (kernel_size - 1)

    def forward(self, input: torch.Tensor) -> torch.Tensor:
        x = torch.nn.functional.pad(input.unsqueeze(2), (self.left_padding, 0, 0, 0)).squeeze(2)
        return super().forward(x)


# Copied from transformers.models.whisper.modeling_whisper.WhisperAttention with Whisper->HithinkAudio
class HithinkAudioAttention(nn.Module):
    """Multi-headed attention from 'Attention Is All You Need' paper"""

    def __init__(
        self,
        embed_dim: int,
        num_heads: int,
        dropout: float = 0.0,
        is_decoder: bool = False,
        bias: bool = True,
        is_causal: bool = False,
        layer_idx: Optional[int] = None,
        config: Optional[HithinkOmniConfig] = None,
    ):
        super().__init__()
        self.embed_dim = embed_dim
        self.num_heads = num_heads
        self.dropout = dropout
        self.head_dim = embed_dim // num_heads
        self.config = config

        if (self.head_dim * 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`: {num_heads})."
            )
        self.scaling = self.head_dim**-0.5
        self.is_decoder = is_decoder
        self.is_causal = is_causal

        if layer_idx is None and is_decoder:
            logger.warning_once(
                f"Instantiating a decoder {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.layer_idx = layer_idx

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

    # Copied from transformers.models.bart.modeling_bart.BartAttention._shape with BART->whisper
    def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
        return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        layer_head_mask: Optional[torch.Tensor] = None,
        past_key_value: Optional[Cache] = None,
        output_attentions: bool = False,
        cache_position: Optional[torch.LongTensor] = None,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        """Input shape: Batch x Time x Channel"""

        bsz, tgt_len, _ = hidden_states.size()

        # get query proj
        query_states = self._shape(self.q_proj(hidden_states) * self.scaling, tgt_len, bsz)
        key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
        value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
        if past_key_value is not None:
            # save all key/value_states to cache to be re-used for fast auto-regressive generation
            key_states, value_states = past_key_value.update(
                key_states, value_states, self.layer_idx, {"cache_position": cache_position}
            )

        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3))

        if attention_mask is not None:  # no matter the length, we just slice it
            causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
            attn_weights = attn_weights + causal_mask

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

        if layer_head_mask is not None:
            if layer_head_mask.size() != (self.num_heads,):
                raise ValueError(
                    f"Head mask for a single layer should be of size {(self.num_heads,)}, but is"
                    f" {layer_head_mask.size()}"
                )
            attn_weights = layer_head_mask.view(1, -1, 1, 1) * attn_weights

        attn_probs = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)
        attn_output = torch.matmul(attn_probs, value_states)

        if attn_output.size() != (bsz, self.num_heads, tgt_len, self.head_dim):
            raise ValueError(
                f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
                f" {attn_output.size()}"
            )

        attn_output = attn_output.transpose(1, 2)
        # 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(bsz, tgt_len, self.embed_dim)

        attn_output = self.out_proj(attn_output)

        return attn_output, attn_weights, past_key_value


# Copied from transformers.models.whisper.modeling_whisper.WhisperFlashAttention2 with Whisper->HithinkAudio
class HithinkAudioFlashAttention2(HithinkAudioAttention):
    """
    HithinkAudio flash attention module. This module inherits from `HithinkAudioAttention` 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.
    """

    # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2.__init__
    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 alignement, 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,
        key_value_states: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        layer_head_mask: Optional[torch.Tensor] = None,
        past_key_value: Optional[Cache] = None,
        output_attentions: bool = False,
        cache_position: Optional[torch.LongTensor] = None,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        if isinstance(past_key_value, StaticCache):
            raise ValueError(
                "The `static` cache implementation is not compatible with `attn_implementation='flash_attention_2'`. "
                "Use `attn_implementation='sdpa'` in the meantime, and open an issue at https://github.com/huggingface/transformers"
            )
        # HithinkAudioFlashAttention2 attention does not support output_attentions
        if output_attentions:
            raise ValueError("HithinkAudioFlashAttention2 attention does not support output_attentions")

        bsz, tgt_len, _ = hidden_states.size()

        # get query proj
        query_states = torch.reshape(self.q_proj(hidden_states), (bsz, tgt_len, self.num_heads, self.head_dim))
        key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
        value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
        if past_key_value is not None:
            # save all key/value_states to cache to be re-used for fast auto-regressive generation
            key_states, value_states = past_key_value.update(
                key_states, value_states, self.layer_idx, {"cache_position": cache_position}
            )

        # TODO: These transpose are quite inefficient but Flash Attention requires the layout [batch_size, sequence_length, num_heads, head_dim]
        #  We would need to refactor the KV cache to be able to avoid many of these transpose/reshape/view.
        key_states = key_states.transpose(1, 2)
        value_states = value_states.transpose(1, 2)

        causal_mask = attention_mask
        if attention_mask is not None:  # no matter the length, we just slice it
            causal_mask = attention_mask[:, : key_states.shape[1]]

        # In PEFT, usually we cast the layer norms in float32 for training stability reasons
        # therefore the input hidden states gets silently casted in float32. Hence, we need
        # cast them back in the correct dtype just to be sure everything works as expected.
        # This might slowdown training & inference so it is recommended to not cast the LayerNorms
        # in fp32. (LlamaRMSNorm handles it correctly)

        input_dtype = query_states.dtype
        if input_dtype == torch.float32:
            if torch.is_autocast_enabled():
                target_dtype = torch.get_autocast_gpu_dtype()
            # Handle the case where the model is quantized
            elif hasattr(self.config, "_pre_quantization_dtype"):
                target_dtype = self.config._pre_quantization_dtype
            else:
                target_dtype = self.q_proj.weight.dtype

            logger.warning_once(
                f"The input hidden states seems to be silently casted in float32, this might be related to"
                f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
                f" {target_dtype}."
            )

            query_states = query_states.to(target_dtype)
            key_states = key_states.to(target_dtype)
            value_states = value_states.to(target_dtype)

        attn_output = _flash_attention_forward(
            query_states,
            key_states,
            value_states,
            causal_mask,
            tgt_len,
            dropout=self.dropout if self.training else 0.0,
            is_causal=self.is_causal,
            use_top_left_mask=self._flash_attn_uses_top_left_mask,
        )

        attn_output = attn_output.reshape(bsz, tgt_len, -1)
        attn_output = self.out_proj(attn_output)

        if not output_attentions:
            attn_weights = None

        return attn_output, attn_weights, past_key_value


# Copied from transformers.models.whisper.modeling_whisper.WhisperSdpaAttention with Whisper->HithinkAudio
class HithinkAudioSdpaAttention(HithinkAudioAttention):
    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        layer_head_mask: Optional[torch.Tensor] = None,
        past_key_value: Optional[Cache] = None,
        output_attentions: bool = False,
        cache_position: Optional[torch.LongTensor] = None,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        """Input shape: Batch x Time x Channel"""
        if output_attentions or layer_head_mask is not None:
            # TODO: Improve this warning with e.g. `model.config._attn_implementation = "manual"` once this is implemented.
            logger.warning_once(
                "HithinkAudioModel is using HithinkAudioSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True` or `layer_head_mask` not None. Falling back to the manual attention"
                ' implementation, but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
            )
            return super().forward(
                hidden_states,
                past_key_value=past_key_value,
                attention_mask=attention_mask,
                layer_head_mask=layer_head_mask,
                output_attentions=output_attentions,
                cache_position=cache_position,
            )
        bsz, tgt_len, _ = hidden_states.size()

        # get query proj
        query_states = self._shape(self.q_proj(hidden_states), tgt_len, bsz)
        key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
        value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
        if past_key_value is not None:
            # save all key/value_states to cache to be re-used for fast auto-regressive generation
            key_states, value_states = past_key_value.update(
                key_states, value_states, self.layer_idx, {"cache_position": cache_position}
            )

        causal_mask = attention_mask
        if attention_mask is not None:  # no matter the length, we just slice it
            causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]

        # 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.
        is_causal = True if self.is_causal and causal_mask is None and tgt_len > 1 else False

        # 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=causal_mask,
            dropout_p=self.dropout if self.training else 0.0,
            is_causal=is_causal,
        )

        if attn_output.size() != (bsz, self.num_heads, tgt_len, self.head_dim):
            raise ValueError(
                f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
                f" {attn_output.size()}"
            )

        attn_output = attn_output.transpose(1, 2)

        # 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(bsz, tgt_len, self.embed_dim)

        attn_output = self.out_proj(attn_output)

        return attn_output, None, past_key_value


HITHINKAUDIO_ATTENTION_CLASSES = {
    "eager": HithinkAudioAttention,
    "flash_attention_2": HithinkAudioFlashAttention2,
    "sdpa": HithinkAudioSdpaAttention,
}


# Copied from transformers.models.whisper.modeling_whisper.WhisperEncoderLayer with Whisper->HithinkAudio, WHISPER->HITHINKAUDIO
class HithinkAudioEncoderLayer(nn.Module):
    def __init__(self, config: HithinkAudioEncoderConfig, layer_idx: int):
        super().__init__()
        self.embed_dim = config.d_model

        self.self_attn = HITHINKAUDIO_ATTENTION_CLASSES[config._attn_implementation](
            embed_dim=self.embed_dim,
            num_heads=config.encoder_attention_heads,
            dropout=config.attention_dropout,
            config=config,
            layer_idx=layer_idx,
            is_causal=True
        )
        self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
        self.dropout = config.dropout
        self.activation_fn = ACT2FN[config.activation_function]
        self.activation_dropout = config.activation_dropout
        self.fc1 = nn.Linear(self.embed_dim, config.encoder_ffn_dim)
        self.fc2 = nn.Linear(config.encoder_ffn_dim, self.embed_dim)
        self.final_layer_norm = nn.LayerNorm(self.embed_dim)

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.Tensor,
        layer_head_mask: torch.Tensor,
        past_key_value: Optional[Cache],
        output_attentions: bool = False,
        use_cache: bool = False,
    ) -> torch.Tensor:
        """
        Args:
            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
            attention_mask (`torch.FloatTensor`): attention mask of size
                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
            layer_head_mask (`torch.FloatTensor`): mask for attention heads in a given layer of size
                `(encoder_attention_heads,)`.
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
        """
        residual = hidden_states
        hidden_states = self.self_attn_layer_norm(hidden_states)
        hidden_states, attn_weights, present_key_value = self.self_attn(
            hidden_states=hidden_states,
            attention_mask=attention_mask,
            layer_head_mask=layer_head_mask,
            past_key_value=past_key_value,
            output_attentions=output_attentions,
        )
        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
        hidden_states = residual + hidden_states

        residual = hidden_states
        hidden_states = self.final_layer_norm(hidden_states)
        hidden_states = self.activation_fn(self.fc1(hidden_states))
        hidden_states = nn.functional.dropout(hidden_states, p=self.activation_dropout, training=self.training)
        hidden_states = self.fc2(hidden_states)
        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
        hidden_states = residual + hidden_states

        if hidden_states.dtype == torch.float16 and (
            torch.isinf(hidden_states).any() or torch.isnan(hidden_states).any()
        ):
            clamp_value = torch.finfo(hidden_states.dtype).max - 1000
            hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)

        outputs = (hidden_states,)

        if output_attentions:
            outputs += (attn_weights,)

        if use_cache:
            outputs += (present_key_value,)

        return outputs


HITHINKAUDIO_START_DOCSTRING = r"""
    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
    etc.)

    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
    and behavior.

    Parameters:
        config ([`HithinkOmniConfig`]):
            Model configuration class with all the parameters of the model. Initializing with a config file does not
            load the weights associated with the model, only the configuration. Check out the
            [`~PreTrainedModel.from_pretrained`] method to load the model weights.
"""


@add_start_docstrings(
    "The bare HithinkAudio Model outputting raw hidden-states without any specific head on top.",
    HITHINKAUDIO_START_DOCSTRING,
)
class HithinkAudioPreTrainedModel(PreTrainedModel):
    config_class = HithinkOmniConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["HithinkAudioAttention"]
    _skip_keys_device_placement = "past_key_values"
    _supports_flash_attn_2 = True

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

        if isinstance(module, (nn.Linear, nn.Conv1d)):
            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_()

    @property
    def _supports_sdpa(self):
        """
        Retrieve language_model's attribute to check whether the model supports
        SDPA or not.
        """
        return self.language_model._supports_sdpa


HITHINKAUDIOENCODER_START_DOCSTRING = r"""
    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
    etc.)

    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
    and behavior.

    Parameters:
        config ([`HithinkAudioEncoderConfig`]):
            Model configuration class with all the parameters of the model. Initializing with a config file does not
            load the weights associated with the model, only the configuration. Check out the
            [`~PreTrainedModel.from_pretrained`] method to load the model weights.
"""


@add_start_docstrings(
    """The audio model from HithinkAudio without any head or projection on top.""",
    HITHINKAUDIOENCODER_START_DOCSTRING,
)
# Copied from transformers.models.whisper.modeling_whisper.WhisperEncoder with Whisper->HithinkAudio
class HithinkAudioEncoder(HithinkAudioPreTrainedModel):
    """
    Transformer encoder consisting of *config.encoder_layers* self attention layers. Each layer is a
    [`HithinkAudioEncoderLayer`].

    Args:
        config: HithinkAudioEncoderConfig
    """

    # Ignore copy
    config_class = HithinkAudioEncoderConfig
    main_input_name = "input_features"
    _no_split_modules = ["HithinkAudioEncoderLayer"]

    def __init__(self, config: HithinkAudioEncoderConfig):
        super().__init__(config)
        self.dropout = config.dropout
        self.layerdrop = config.encoder_layerdrop

        embed_dim = config.d_model
        self.num_mel_bins = config.num_mel_bins
        self.padding_idx = config.pad_token_id
        self.max_source_positions = config.max_source_positions
        self.embed_scale = math.sqrt(embed_dim) if config.scale_embedding else 1.0

        self.conv1 = CausalConv1d(self.num_mel_bins, embed_dim, kernel_size=3, padding=1)
        self.conv2 = CausalConv1d(embed_dim, embed_dim, kernel_size=3, stride=2, padding=1)

        self.embed_positions = nn.Embedding(self.max_source_positions, embed_dim)
        self.embed_positions.requires_grad_(False)

        self.layers = nn.ModuleList([
            HithinkAudioEncoderLayer(config, layer_idx) for layer_idx in range(config.encoder_layers)
        ])
        self.layer_norm = nn.LayerNorm(config.d_model)
        # Ignore copy
        self.avg_pooler = nn.AvgPool1d(2, stride=2)

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

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

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

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

    def forward(
        self,
        input_features,
        attention_mask=None,
        head_mask=None,
        past_key_values=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
        use_cache=False,
    ):
        r"""
        Args:
            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__`]
            attention_mask (`torch.Tensor`)`, *optional*):
                HithinkAudio does not support masking of the `input_features`, this argument is preserved for compatibility,
                but it is not used. By default the silence in the input log mel spectrogram are ignored.
            head_mask (`torch.Tensor` of shape `(encoder_layers, encoder_attention_heads)`, *optional*):
                Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:

                - 1 indicates the head is **not masked**,
                - 0 indicates the head is **masked**.
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
            output_hidden_states (`bool`, *optional*):
                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
                for more detail.
            return_dict (`bool`, *optional*):
                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
        """
        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

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

        past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0

        # Ignore copy
        input_features = input_features.to(dtype=self.conv1.weight.dtype, device=self.conv1.weight.device)

        inputs_embeds = nn.functional.gelu(self.conv1(input_features))
        inputs_embeds = nn.functional.gelu(self.conv2(inputs_embeds))

        inputs_embeds = inputs_embeds.permute(0, 2, 1)
        if past_seen_tokens > 0:  # 流式输入的后续chunk，需要去除与之前chunk重合的部分（这部分保留在输入中只是为了卷积计算正确性）
            inputs_embeds = inputs_embeds[:, 2:]
        embed_pos = self.embed_positions.weight[past_seen_tokens: past_seen_tokens + inputs_embeds.shape[1]]

        hidden_states = inputs_embeds + embed_pos
        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)

        encoder_states = () if output_hidden_states else None
        all_attentions = () if output_attentions else None
        next_cache = None

        attention_mask = self._prepare_attention_mask(input_features, attention_mask, past_seen_tokens)

        # check if head_mask has a correct number of layers specified if desired
        if head_mask is not None:
            assert head_mask.size()[0] == (
                len(self.layers)
            ), f"The head_mask should be specified for {len(self.layers)} layers, but it is for {head_mask.size()[0]}."

        for idx, encoder_layer in enumerate(self.layers):
            if output_hidden_states:
                encoder_states = encoder_states + (hidden_states,)
            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
            to_drop = False
            if self.training:
                dropout_probability = torch.rand([])
                if dropout_probability < self.layerdrop:  # skip the layer
                    to_drop = True

            # Ignore copy
            if to_drop:
                layer_outputs = (None, None)
            else:
                if self.gradient_checkpointing and self.training:
                    layer_outputs = self._gradient_checkpointing_func(
                        encoder_layer.__call__,
                        hidden_states,
                        attention_mask,
                        (head_mask[idx] if head_mask is not None else None),
                        past_key_values,
                        output_attentions,
                        use_cache,
                    )
                else:
                    layer_outputs = encoder_layer(
                        hidden_states,
                        attention_mask,
                        layer_head_mask=(head_mask[idx] if head_mask is not None else None),
                        past_key_value=past_key_values,
                        output_attentions=output_attentions,
                        use_cache=use_cache,
                    )

                hidden_states = layer_outputs[0]

            if use_cache:
                next_cache = layer_outputs[2 if output_attentions else 1]

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

        # Ignore copy
        hidden_states = hidden_states.permute(0, 2, 1)
        hidden_states = self.avg_pooler(hidden_states)
        hidden_states = hidden_states.permute(0, 2, 1)

        hidden_states = self.layer_norm(hidden_states)
        if output_hidden_states:
            encoder_states = encoder_states + (hidden_states,)

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

    # 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 _prepare_attention_mask(self, input_features, feature_attention_mask, past_seen_tokens):
        feat_lengths, output_lengths = self._get_feat_extract_output_lengths(
            feature_attention_mask.sum(-1)
        )
        batch_size, _, max_mel_seq_len = input_features.shape
        max_seq_len = (max_mel_seq_len - 1) // 2 + 1
        # Create a sequence tensor of shape (batch_size, max_seq_len)
        seq_range = (
            torch.arange(0, max_seq_len, dtype=feat_lengths.dtype, device=feat_lengths.device)
            .unsqueeze(0)
            .expand(batch_size, max_seq_len)
        )
        lengths_expand = feat_lengths.unsqueeze(1).expand(batch_size, max_seq_len)
        # Create mask
        padding_mask = seq_range >= lengths_expand
        if self.config._attn_implementation == "flash_attention_2":
            attention_mask = ~padding_mask
            if past_seen_tokens > 0:
                attention_mask = attention_mask[:, 2:]  # 去除当前chunk与之前重合的部分
                past_mask = torch.ones(
                    (batch_size, past_seen_tokens,),
                    dtype=attention_mask.dtype,
                    device=attention_mask.device
                )
                attention_mask = torch.cat([past_mask, attention_mask], dim=1)
        else:
            position_ids = torch.arange(max_seq_len, device=padding_mask.device)
            causal_mask = position_ids > position_ids.reshape(-1, 1)
            causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
            attention_mask_ = padding_mask.view(batch_size, 1, 1, max_seq_len).expand(
                batch_size, 1, max_seq_len, max_seq_len
            )
            attention_mask_ = attention_mask_.clone() | causal_mask
            attention_mask = attention_mask_.to(
                dtype=input_features.dtype, device=input_features.device
            )
            attention_mask[attention_mask_] = float("-inf")
            if past_seen_tokens > 0:
                attention_mask = attention_mask[:, :, 2:, 2:]  # 去除当前chunk与之前重合的部分
                past_mask = torch.zeros(
                    attention_mask.shape[:3] + (past_seen_tokens,),
                    dtype=attention_mask.dtype,
                    device=attention_mask.device
                )
                attention_mask = torch.cat([past_mask, attention_mask], dim=-1)
        return attention_mask


class HithinkAudioMultiModalProjector(nn.Module):
    def __init__(self, config: HithinkOmniConfig):
        super().__init__()
        self.linear = nn.Linear(config.audio_config.d_model, config.hidden_size, bias=True)

    def forward(self, audio_features):
        hidden_states = self.linear(audio_features)
        return hidden_states
# audio part end

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

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

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

    def _dynamic_frequency_update(self, position_ids, device):
        """
        dynamic RoPE layers should recompute `inv_freq` in the following situations:
        1 - growing beyond the cached sequence length (allow scaling)
        2 - the current sequence length is in the original scale (avoid losing precision with small sequences)
        """
        seq_len = torch.max(position_ids) + 1
        if seq_len > self.max_seq_len_cached:  # growth
            inv_freq, self.attention_scaling = self.rope_init_fn(
                self.config, device, seq_len=seq_len, **self.rope_kwargs
            )
            self.register_buffer("inv_freq", inv_freq, persistent=False)  # TODO joao: may break with compilation
            self.max_seq_len_cached = seq_len

        if seq_len < self.original_max_seq_len and self.max_seq_len_cached > self.original_max_seq_len:  # reset
            self.register_buffer("inv_freq", self.original_inv_freq, persistent=False)
            self.max_seq_len_cached = self.original_max_seq_len

    @torch.no_grad()
    def forward(self, x, position_ids):
        if "dynamic" in self.rope_type:
            self._dynamic_frequency_update(position_ids, device=x.device)

        # Core RoPE block. In contrast to other models, HithinkOmni has different position ids for thw grids
        # So we expand the inv_freq to shape (3, ...)
        inv_freq_expanded = self.inv_freq[None, None, :, None].float().expand(3, position_ids.shape[1], -1, 1)
        position_ids_expanded = position_ids[:, :, None, :].float()  # shape (3, bs, 1, positions)
        # Force float32 (see https://github.com/huggingface/transformers/pull/29285)
        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 = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(2, 3)
            emb = torch.cat((freqs, freqs), dim=-1)
            cos = emb.cos()
            sin = emb.sin()

        # Advanced RoPE types (e.g. yarn) apply a post-processing scaling factor, equivalent to scaling attention
        cos = cos * self.attention_scaling
        sin = sin * self.attention_scaling

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


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


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

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



def apply_multimodal_rotary_pos_emb(q, k, cos, sin, mrope_section, unsqueeze_dim=1):
    """Applies Rotary Position Embedding with Multimodal Sections to the query and key tensors (https://qwenlm.github.io/blog/qwen2-vl/).

    Explanation:
        Multimodal 3D rotary position embedding is an extension to 1D rotary position embedding. The input embedding
        sequence contains vision (images / videos) embedding and text embedding or just contains text embedding. For
        vision embedding part, we apply rotary position embedding on temporal, height and width dimension seperately.
        Here we split the channel dimension to 3 chunks for the temporal, height and width rotary position embedding.
        For text embedding part, we just apply 1D rotary position embedding. The three rotary position index (temporal,
        height and width) of text embedding is always the same, so the text embedding rotary position embedding has no
        difference with modern LLMs.

    Args:
        q (`torch.Tensor`): The query tensor.
        k (`torch.Tensor`): The key tensor.
        cos (`torch.Tensor`): The cosine part of the rotary embedding.
        sin (`torch.Tensor`): The sine part of the rotary embedding.
        position_ids (`torch.Tensor`):
            The position indices of the tokens corresponding to the query and key tensors. For example, this can be
            used to pass offsetted position ids when working with a KV-cache.
        mrope_section(`List(int)`):
            Multimodal rope section is for channel dimension of temporal, height and width in rope calculation.
        unsqueeze_dim (`int`, *optional*, defaults to 1):
            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
    Returns:
        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
    """
    mrope_section = mrope_section * 2
    cos = torch.cat([m[i % 3] for i, m in enumerate(cos.split(mrope_section, dim=-1))], dim=-1).unsqueeze(
        unsqueeze_dim
    )
    sin = torch.cat([m[i % 3] for i, m in enumerate(sin.split(mrope_section, dim=-1))], dim=-1).unsqueeze(
        unsqueeze_dim
    )

    if q.is_cuda:
        if apply_rotary_emb_func is not None:
            rot_dim = cos.shape[-1] // 2
            bs, qheads, seqlen, headdim = q.size()
            kheads = k.size(1)
            cos = cos[:, 0, :, :rot_dim].view(bs * seqlen, rot_dim)
            sin = sin[:, 0, :, :rot_dim].view(bs * seqlen, rot_dim)
            q = q.transpose(1, 2).view(1, bs * seqlen, qheads, headdim)
            k = k.transpose(1, 2).view(1, bs * seqlen, kheads, headdim)
            q_embed = apply_rotary_emb_func(q, cos, sin, False, False)  # interleaved=False, inplace=False
            k_embed = apply_rotary_emb_func(k, cos, sin, False, False)  # interleaved=False, inplace=False
            q_embed = q_embed.view(bs, seqlen, qheads, headdim).transpose(1, 2)
            k_embed = k_embed.view(bs, seqlen, kheads, headdim).transpose(1, 2)
            return q_embed, k_embed
        else:
            logger.warning_once("rotary_emb is not installed. If you want to accelerate training please install: "
                                "https://github.com/Dao-AILab/flash-attention/tree/main/csrc/rotary")

    q_embed = (q * cos) + (rotate_half(q) * sin)
    k_embed = (k * cos) + (rotate_half(k) * sin)
    return q_embed, k_embed


def apply_rotary_pos_emb_vision(tensor: torch.Tensor, freqs: torch.Tensor) -> torch.Tensor:
    orig_dtype = tensor.dtype
    tensor = tensor.float()
    cos = freqs.cos().float()
    sin = freqs.sin().float()
    if tensor.is_cuda:
        if apply_rotary_emb_func is not None:
            output = apply_rotary_emb_func(tensor, cos, sin, False, False)  # interleaved=False, inplace=False
            return output.to(orig_dtype)
        else:
            logger.warning_once("rotary_emb is not installed. If you want to accelerate training please install: "
                                "https://github.com/Dao-AILab/flash-attention/tree/main/csrc/rotary")

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

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


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

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

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        target_dtype = self.proj.weight.dtype
        hidden_states = hidden_states.view(
            -1, self.in_channels, self.temporal_patch_size, self.patch_size, self.patch_size
        )
        hidden_states = self.proj(hidden_states.to(dtype=target_dtype)).view(-1, self.embed_dim)
        return hidden_states


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

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


class VisionAttention(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.qkv = nn.Linear(dim, dim * 3, 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, k, v = self.qkv(hidden_states).reshape(seq_length, 3, self.num_heads, -1).permute(1, 0, 2, 3).unbind(0)
        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 VisionFlashAttention2(nn.Module):
    def __init__(self, dim: int, num_heads: int = 16) -> None:
        super().__init__()
        self.num_heads = num_heads
        self.qkv = nn.Linear(dim, dim * 3, 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, k, v = self.qkv(hidden_states).reshape(seq_length, 3, self.num_heads, -1).permute(1, 0, 2, 3).unbind(0)
        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)

        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 VisionSdpaAttention(nn.Module):
    def __init__(self, dim: int, num_heads: int = 16) -> None:
        super().__init__()
        self.num_heads = num_heads
        self.qkv = nn.Linear(dim, dim * 3, 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, k, v = self.qkv(hidden_states).reshape(seq_length, 3, self.num_heads, -1).permute(1, 0, 2, 3).unbind(0)
        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


VISION_ATTENTION_CLASSES = {
    "eager": VisionAttention,
    "flash_attention_2": VisionFlashAttention2,
    "sdpa": VisionSdpaAttention,
}


class HithinkOmniVisionBlock(nn.Module):
    def __init__(self, config, attn_implementation: str = "sdpa") -> None:
        super().__init__()
        self.norm1 = HithinkRMSNorm(config.hidden_size, eps=1e-6)
        self.norm2 = HithinkRMSNorm(config.hidden_size, eps=1e-6)
        self.attn = VISION_ATTENTION_CLASSES[attn_implementation](
            config.hidden_size, num_heads=config.num_heads
        )
        self.mlp = HithinkOmniMLP(config, bias=True)

    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 HithinkRMSNorm(nn.Module):
    def __init__(self, hidden_size, eps=1e-6):
        """
        HithinkRMSNorm is equivalent to T5LayerNorm
        """
        super().__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
        self.variance_epsilon = eps

    def forward(self, hidden_states, residual=None):
        if hidden_states.is_cuda:
            if dropout_add_rms_norm is not None:
                out, res = dropout_add_rms_norm(
                    hidden_states,
                    residual,
                    self.weight,
                    None,  # bias
                    0.,  # dropout_p
                    self.variance_epsilon,
                    prenorm=True,
                    residual_in_fp32=False,
                    return_dropout_mask=False,
                )
                return out if residual is None else (out, res)
            else:
                logger.warning_once("dropout_add_rms_norm is not installed. If you want to accelerate training please install: "
                                    "https://github.com/Dao-AILab/flash-attention/tree/main/csrc/layer_norm")

        if residual is not None:
            hidden_states = residual + hidden_states
            residual = hidden_states
        input_dtype = hidden_states.dtype
        hidden_states = hidden_states.to(torch.float32)
        variance = hidden_states.pow(2).mean(-1, keepdim=True)
        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
        hidden_states = self.weight * hidden_states.to(input_dtype)
        return hidden_states if residual is None else (hidden_states, residual)

    def extra_repr(self):
        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"


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

    def forward(self, x):
        if self.config.hidden_act == 'silu' and x.is_cuda:
            if swiglu is not None:
                down_proj = self.down_proj(swiglu(self.gate_proj(x), self.up_proj(x)))
            else:
                logger.warning_once("swiglu is not installed. If you want to accelerate training please install: "
                                    "https://github.com/Dao-AILab/flash-attention")
        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
        return down_proj


def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
    """
    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
    """
    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
    if n_rep == 1:
        return hidden_states
    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)


class HithinkOmniAttention(nn.Module):
    """
    Multi-headed attention from 'Attention Is All You Need' paper. Modified to use sliding window attention: Longformer
    and "Generating Long Sequences with Sparse Transformers".
    """

    def __init__(self, config: HithinkOmniConfig, layer_idx: Optional[int] = None):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx
        if layer_idx is None:
            logger.warning_once(
                f"Instantiating {self.__class__.__name__} without passing `layer_idx` is not recommended and will "
                "to errors during the forward call, if caching is used. Please make sure to provide a `layer_idx` "
                "when creating this class."
            )

        self.hidden_size = config.hidden_size
        self.num_heads = config.num_attention_heads
        self.head_dim = self.hidden_size // self.num_heads
        self.num_key_value_heads = config.num_key_value_heads
        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
        self.is_causal = True
        self.attention_dropout = config.attention_dropout
        self.rope_scaling = config.rope_scaling

        if (self.head_dim * self.num_heads) != self.hidden_size:
            raise ValueError(
                f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
                f" and `num_heads`: {self.num_heads})."
            )
        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=True)
        self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=True)
        self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=True)
        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)

        self.rotary_emb = HithinkOmniRotaryEmbedding(config=config)

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_value: Optional[Cache] = None,
        output_attentions: bool = False,
        use_cache: bool = False,
        cache_position: Optional[torch.LongTensor] = None,
        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        bsz, q_len, _ = hidden_states.size()

        query_states = self.q_proj(hidden_states)
        key_states = self.k_proj(hidden_states)
        value_states = self.v_proj(hidden_states)

        query_states = query_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2)
        key_states = key_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2)
        value_states = value_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2)

        cos, sin = position_embeddings
        query_states, key_states = apply_multimodal_rotary_pos_emb(
            query_states, key_states, cos, sin, self.rope_scaling["mrope_section"]
        )

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

        # repeat k/v heads if n_kv_heads < n_heads
        key_states = repeat_kv(key_states, self.num_key_value_groups)
        value_states = repeat_kv(value_states, self.num_key_value_groups)

        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)

        if attention_mask is not None:  # no matter the length, we just slice it
            causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
            attn_weights = attn_weights + causal_mask

        # Fix precision issues in HithinkOmni float16 inference
        # Replace inf values with zeros in attention weights to prevent NaN propagation
        if query_states.dtype == torch.float16:
            attn_weights = torch.where(torch.isinf(attn_weights), torch.zeros_like(attn_weights), attn_weights)

        # upcast attention to fp32
        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
        attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
        attn_output = torch.matmul(attn_weights, value_states)

        if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
            raise ValueError(
                f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
                f" {attn_output.size()}"
            )

        attn_output = attn_output.transpose(1, 2).contiguous()
        attn_output = attn_output.reshape(bsz, q_len, -1)

        attn_output = self.o_proj(attn_output)

        if not output_attentions:
            attn_weights = None

        return attn_output, attn_weights, past_key_value


class HithinkOmniFlashAttention2(HithinkOmniAttention):
    """
    HithinkOmni flash attention module, following HithinkOmni attention module. This module inherits from `HithinkOmniAttention`
    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. Additionally, for sliding window attention, we apply SWA only to the bottom
    config.max_window_layers layers.
    """

    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 alignement, 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,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_value: Optional[Cache] = None,
        output_attentions: bool = False,
        use_cache: bool = False,
        cache_position: Optional[torch.LongTensor] = None,
        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
    ):
        bsz, q_len, _ = hidden_states.size()

        query_states = self.q_proj(hidden_states)
        key_states = self.k_proj(hidden_states)
        value_states = self.v_proj(hidden_states)

        query_states = query_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2)
        key_states = key_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2)
        value_states = value_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2)

        # Because the input can be padded, the absolute sequence length depends on the max position id.
        cos, sin = position_embeddings
        query_states, key_states = apply_multimodal_rotary_pos_emb(
            query_states, key_states, cos, sin, self.rope_scaling["mrope_section"]
        )

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

        # repeat k/v heads if n_kv_heads < n_heads
        key_states = repeat_kv(key_states, self.num_key_value_groups)
        value_states = repeat_kv(value_states, self.num_key_value_groups)
        dropout_rate = 0.0 if not self.training else self.attention_dropout

        # In PEFT, usually we cast the layer norms in float32 for training stability reasons
        # therefore the input hidden states gets silently casted in float32. Hence, we need
        # cast them back in float16 just to be sure everything works as expected.
        input_dtype = query_states.dtype
        if input_dtype == torch.float32:
            if torch.is_autocast_enabled():
                target_dtype = torch.get_autocast_gpu_dtype()
            # Handle the case where the model is quantized
            elif hasattr(self.config, "_pre_quantization_dtype"):
                target_dtype = self.config._pre_quantization_dtype
            else:
                target_dtype = self.q_proj.weight.dtype

            logger.warning_once(
                f"The input hidden states seems to be silently casted in float32, this might be related to"
                f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
                f" {target_dtype}."
            )

            query_states = query_states.to(target_dtype)
            key_states = key_states.to(target_dtype)
            value_states = value_states.to(target_dtype)

        # Reashape to the expected shape for Flash Attention
        query_states = query_states.transpose(1, 2)
        key_states = key_states.transpose(1, 2)
        value_states = value_states.transpose(1, 2)

        if (
            self.config.use_sliding_window
            and getattr(self.config, "sliding_window", None) is not None
            and self.layer_idx >= self.config.max_window_layers
        ):
            sliding_window = self.config.sliding_window
        else:
            sliding_window = None

        attn_output = _flash_attention_forward(
            query_states,
            key_states,
            value_states,
            attention_mask,
            q_len,
            dropout=dropout_rate,
            sliding_window=sliding_window,
            is_causal=self.is_causal,
            use_top_left_mask=self._flash_attn_uses_top_left_mask,
        )
        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()

        attn_output = self.o_proj(attn_output)

        if not output_attentions:
            attn_weights = None

        return attn_output, attn_weights, past_key_value


class HithinkOmniSdpaAttention(HithinkOmniAttention):
    """
    Hithink attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
    `HithinkAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
    SDPA API.
    """

    # Adapted from HithinkAttention.forward
    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_value: Optional[Cache] = None,
        output_attentions: bool = False,
        use_cache: bool = False,
        cache_position: Optional[torch.LongTensor] = None,
        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        if output_attentions:
            # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
            logger.warning_once(
                "HithinkOmniModel is using HithinkOmniSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, "
                'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
            )
            return super().forward(
                hidden_states=hidden_states,
                attention_mask=attention_mask,
                position_ids=position_ids,
                past_key_value=past_key_value,
                output_attentions=output_attentions,
                use_cache=use_cache,
                cache_position=cache_position,
                position_embeddings=position_embeddings,
            )

        bsz, q_len, _ = hidden_states.size()

        query_states = self.q_proj(hidden_states)
        key_states = self.k_proj(hidden_states)
        value_states = self.v_proj(hidden_states)

        query_states = query_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2)
        key_states = key_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2)
        value_states = value_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2)

        cos, sin = position_embeddings
        query_states, key_states = apply_multimodal_rotary_pos_emb(
            query_states, key_states, cos, sin, self.rope_scaling["mrope_section"]
        )

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

        key_states = repeat_kv(key_states, self.num_key_value_groups)
        value_states = repeat_kv(value_states, self.num_key_value_groups)

        causal_mask = attention_mask
        if attention_mask is not None:  # no matter the length, we just slice it
            causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]

        # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
        # Reference: https://github.com/pytorch/pytorch/issues/112577.
        if query_states.device.type == "cuda" and attention_mask is not None:
            query_states = query_states.contiguous()
            key_states = key_states.contiguous()
            value_states = value_states.contiguous()

        # 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 q_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case q_len == 1.
        is_causal = True if causal_mask is None and q_len > 1 else False

        attn_output = torch.nn.functional.scaled_dot_product_attention(
            query_states,
            key_states,
            value_states,
            attn_mask=causal_mask,
            dropout_p=self.attention_dropout if self.training else 0.0,
            is_causal=is_causal,
        )

        attn_output = attn_output.transpose(1, 2).contiguous()
        attn_output = attn_output.view(bsz, q_len, self.hidden_size)

        attn_output = self.o_proj(attn_output)

        return attn_output, None, past_key_value


HITHINKOMNI_ATTENTION_CLASSES = {
    "eager": HithinkOmniAttention,
    "flash_attention_2": HithinkOmniFlashAttention2,
    "sdpa": HithinkOmniSdpaAttention,
}


class HithinkOmniDecoderLayer(nn.Module):
    def __init__(self, config: HithinkOmniConfig, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size

        if config.use_sliding_window and config._attn_implementation != "flash_attention_2":
            logger.warning_once(
                f"Sliding Window Attention is enabled but not implemented for `{config._attn_implementation}`; "
                "unexpected results may be encountered."
            )
        self.self_attn = HITHINKOMNI_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx)

        self.mlp = HithinkMLP(config)
        self.input_layernorm = HithinkRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = HithinkRMSNorm(config.hidden_size, eps=config.rms_norm_eps)

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

        residual = hidden_states

        hidden_states = self.input_layernorm(hidden_states)

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

        # Fully Connected
        hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
        hidden_states = self.mlp(hidden_states)
        hidden_states = residual + hidden_states

        outputs = (hidden_states,)

        if output_attentions:
            outputs += (self_attn_weights,)

        if use_cache:
            outputs += (present_key_value,)

        return outputs


HITHINKOMNI_START_DOCSTRING = r"""
    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
    etc.)

    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
    and behavior.

    Parameters:
        config ([`HithinkOmniConfig`]):
            Model configuration class with all the parameters of the model. Initializing with a config file does not
            load the weights associated with the model, only the configuration. Check out the
            [`~PreTrainedModel.from_pretrained`] method to load the model weights.
"""


@add_start_docstrings(
    "The bare HithinkOmni Model outputting raw hidden-states without any specific head on top.",
    HITHINKOMNI_START_DOCSTRING,
)
class HithinkOmniPreTrainedModel(PreTrainedModel):
    config_class = HithinkOmniConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["HithinkOmniDecoderLayer", "HithinkOmniVisionBlock"]
    _skip_keys_device_placement = "past_key_values"
    _supports_flash_attn_2 = True
    _supports_sdpa = True
    _supports_cache_class = True
    _supports_static_cache = False  # TODO (joao): fix. torch.compile failing probably due to `cache_positions`

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


class HithinkVisionTransformerPretrainedModel(HithinkOmniPreTrainedModel):
    config_class = HithinkOmniVisionConfig
    _no_split_modules = ["HithinkOmniVisionBlock"]

    def __init__(self, config, *inputs, **kwargs) -> None:
        super().__init__(config, *inputs, **kwargs)
        self.spatial_merge_size = config.spatial_merge_size
        self.patch_size = config.patch_size
        self.fullatt_block_indexes = config.fullatt_block_indexes
        self.window_size = config.window_size
        self.spatial_merge_unit = self.spatial_merge_size * self.spatial_merge_size

        self.patch_embed = PatchEmbed(
            patch_size=config.patch_size,
            temporal_patch_size=config.temporal_patch_size,
            in_channels=config.in_channels,
            embed_dim=config.hidden_size,
        )

        head_dim = config.hidden_size // config.num_heads
        self.rotary_pos_emb = VisionRotaryEmbedding(head_dim // 2)

        self.blocks = nn.ModuleList(
            [HithinkOmniVisionBlock(config, config._attn_implementation) for _ in range(config.depth)]
        )
        self.merger = PatchMerger(
            dim=config.out_hidden_size,
            context_dim=config.hidden_size,
            spatial_merge_size=config.spatial_merge_size,
        )
        self.gradient_checkpointing = False

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

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

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

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

        return window_index, cu_window_seqlens

    def forward(self, hidden_states: torch.Tensor, grid_thw: torch.Tensor) -> torch.Tensor:
        """
        Args:
            hidden_states (`torch.Tensor` of shape `(batch_size, 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)

        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


@add_start_docstrings(
    "The bare HithinkOmni Model outputting raw hidden-states without any specific head on top.",
    HITHINKOMNI_START_DOCSTRING,
)
class HithinkOmniModel(HithinkOmniPreTrainedModel):
    def __init__(self, config: HithinkOmniConfig):
        super().__init__(config)
        self.padding_idx = config.pad_token_id
        self.vocab_size = config.vocab_size

        if config.vocab_size:
            self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
        else:
            self.embed_tokens = None

        if config.vocab_size_ext:
            self.embed_tokens_ext = nn.Embedding(config.vocab_size_ext, config.hidden_size)
        else:
            self.embed_tokens_ext = None

        self.layers = nn.ModuleList(
            [HithinkOmniDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self._attn_implementation = config._attn_implementation
        self.norm = HithinkRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.rotary_emb = HithinkOmniRotaryEmbedding(config=config)

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

    def get_input_embeddings(self):
        return self.embed_tokens

    def set_input_embeddings(self, value):
        self.embed_tokens = value

    def convert_token_ids_to_embedding(self, input_ids):
        if self.embed_tokens_ext is None:
            inputs_embeds = self.embed_tokens(input_ids)
        else:
            ext_mask = (input_ids >= self.vocab_size)
            input_ids_base = input_ids.clone()
            input_ids_base.masked_fill_(ext_mask, 0)
            inputs_embeds = self.embed_tokens(input_ids_base)
            inputs_embeds_ext = self.embed_tokens_ext(input_ids[ext_mask] - self.vocab_size)
            ext_embed_mask = ext_mask.unsqueeze(-1).expand_as(inputs_embeds)
            inputs_embeds.masked_scatter_(ext_embed_mask, inputs_embeds_ext)
        return inputs_embeds

    def forward(
        self,
        input_ids: torch.LongTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[List[torch.FloatTensor]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
    ) -> Union[Tuple, BaseModelOutputWithPast]:
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        use_cache = use_cache if use_cache is not None else self.config.use_cache

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

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

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

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

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

        if (
            use_cache and not isinstance(past_key_values, Cache) and not self.training
        ):
            past_key_values = DynamicCache.from_legacy_cache(past_key_values)

        if cache_position is None:
            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
            cache_position = torch.arange(
                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
            )

        # the hard coded `3` is for temporal, height and width.
        if position_ids is None:
            position_ids = cache_position.view(1, 1, -1).expand(3, inputs_embeds.shape[0], -1)
        elif position_ids.dim() == 2:
            position_ids = position_ids[None, ...].expand(3, position_ids.shape[0], -1)

        causal_mask = self._update_causal_mask(
            attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions
        )

        hidden_states = inputs_embeds

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

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

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

            if self.gradient_checkpointing and self.training:
                layer_outputs = self._gradient_checkpointing_func(
                    decoder_layer.__call__,
                    hidden_states,
                    causal_mask,
                    position_ids,
                    past_key_values,
                    output_attentions,
                    use_cache,
                    cache_position,
                    position_embeddings,
                )
            else:
                layer_outputs = decoder_layer(
                    hidden_states,
                    attention_mask=causal_mask,
                    position_ids=position_ids,
                    past_key_value=past_key_values,
                    output_attentions=output_attentions,
                    use_cache=use_cache,
                    cache_position=cache_position,
                    position_embeddings=position_embeddings,
                )

            hidden_states = layer_outputs[0]

            if use_cache:
                next_decoder_cache = layer_outputs[2 if output_attentions else 1]

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

        hidden_states = self.norm(hidden_states)

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

        next_cache = next_decoder_cache if use_cache else None

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

    def _update_causal_mask(
        self,
        attention_mask: torch.Tensor,
        input_tensor: torch.Tensor,
        cache_position: torch.Tensor,
        past_key_values: Cache,
        output_attentions: bool,
    ):
        if self.config._attn_implementation == "flash_attention_2":
            if attention_mask is not None and 0.0 in attention_mask:
                return attention_mask
            return None

        # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in
        # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail
        # to infer the attention mask.
        past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
        using_static_cache = isinstance(past_key_values, StaticCache)
        using_sliding_window_cache = isinstance(past_key_values, SlidingWindowCache)

        # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
        if (
            self.config._attn_implementation == "sdpa"
            and not (using_static_cache or using_sliding_window_cache)
            and not output_attentions
        ):
            if AttentionMaskConverter._ignore_causal_mask_sdpa(
                attention_mask,
                inputs_embeds=input_tensor,
                past_key_values_length=past_seen_tokens,
                sliding_window=self.config.sliding_window,
                is_training=self.training,
            ):
                return None

        dtype, device = input_tensor.dtype, input_tensor.device
        min_dtype = torch.finfo(dtype).min
        sequence_length = input_tensor.shape[1]
        # SlidingWindowCache or StaticCache
        if using_sliding_window_cache or using_static_cache:
            target_length = past_key_values.get_max_cache_shape()
        # DynamicCache or no cache
        else:
            target_length = (
                attention_mask.shape[-1]
                if isinstance(attention_mask, torch.Tensor)
                else past_seen_tokens + sequence_length + 1
            )

        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
        causal_mask = self._prepare_4d_causal_attention_mask_with_cache_position(
            attention_mask,
            sequence_length=sequence_length,
            target_length=target_length,
            dtype=dtype,
            device=device,
            cache_position=cache_position,
            batch_size=input_tensor.shape[0],
            config=self.config,
            past_key_values=past_key_values,
        )

        if (
            self.config._attn_implementation == "sdpa"
            and attention_mask is not None
            and attention_mask.device.type == "cuda"
            and not output_attentions
        ):
            # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
            # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
            # Details: https://github.com/pytorch/pytorch/issues/110213
            causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)

        return causal_mask

    @staticmethod
    def _prepare_4d_causal_attention_mask_with_cache_position(
        attention_mask: torch.Tensor,
        sequence_length: int,
        target_length: int,
        dtype: torch.dtype,
        device: torch.device,
        cache_position: torch.Tensor,
        batch_size: int,
        config: HithinkOmniConfig,
        past_key_values: Cache,
    ):
        """
        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 plcae the 4D attention mask on.
            cache_position (`torch.Tensor`):
                Indices depicting the position of the input sequence tokens in the sequence.
            batch_size (`torch.Tensor`):
                Batch size.
            config (`HithinkOmniConfig`):
                The model's configuration class
            past_key_values (`Cache`):
                The cache class that is being used currently to generate
        """
        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:
            min_dtype = torch.finfo(dtype).min
            causal_mask = torch.full(
                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
            )
            diagonal_attend_mask = torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
            if config.sliding_window is not None:
                # if we have sliding window, we should not attend to tokens beyond sliding window length, so we mask them out also
                # the check is needed to verify is current checkpoint was trained with sliding window or not
                if not isinstance(past_key_values, SlidingWindowCache) or sequence_length > target_length:
                    sliding_attend_mask = torch.arange(target_length, device=device) <= (
                        cache_position.reshape(-1, 1) - config.sliding_window
                    )
                    diagonal_attend_mask.bitwise_or_(sliding_attend_mask)
            causal_mask *= diagonal_attend_mask
            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
                if attention_mask.shape[-1] > target_length:
                    attention_mask = attention_mask[:, :target_length]
                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


HITHINKOMNI_INPUTS_DOCSTRING = r"""
    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.

            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
            [`PreTrainedTokenizer.__call__`] for details.

            [What are input IDs?](../glossary#input-ids)
        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:

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

            [What are attention masks?](../glossary#attention-mask)

            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
            [`PreTrainedTokenizer.__call__`] for details.

            If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
            `past_key_values`).

            If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
            and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
            information on the default strategy.

            - 1 indicates the head is **not masked**,
            - 0 indicates the head is **masked**.
        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
            config.n_positions - 1]`. [What are position IDs?](../glossary#position-ids)
        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)`) and 2 additional tensors of shape
            `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.

            Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
            blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.

            If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
            don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
            `decoder_input_ids` of shape `(batch_size, sequence_length)`.
        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
            model's internal embedding lookup matrix.
        use_cache (`bool`, *optional*):
            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
            `past_key_values`).
        output_attentions (`bool`, *optional*):
            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
            tensors for more detail.
        output_hidden_states (`bool`, *optional*):
            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
            more detail.
        return_dict (`bool`, *optional*):
            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
        pixel_values (`torch.FloatTensor` of shape `(seq_length, num_channels * image_size * image_size)):
            The tensors corresponding to the input images. Pixel values can be obtained using
            [`AutoImageProcessor`]. See [`HithinkOmniImageProcessor.__call__`] for details. [`HithinkOmniProcessor`] uses
            [`HithinkOmniImageProcessor`] for processing images.
        pixel_values_videos (`torch.FloatTensor` of shape `(seq_length, num_channels * temporal_size * image_size * image_size)):
            The tensors corresponding to the input videos. Pixel values can be obtained using
            [`AutoImageProcessor`]. See [`HithinkOmniImageProcessor.__call__`] for details. [`HithinkOmniProcessor`] uses
            [`HithinkOmniImageProcessor`] for processing videos.
        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__`]
        feature_attention_mask (`torch.Tensor` of shape `(batch_size, feature_sequence_length)`):
            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**.
        image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
            The temporal, height and width of feature shape of each image in LLM.
        video_grid_thw (`torch.LongTensor` of shape `(num_videos, 3)`, *optional*):
            The temporal, height and width of feature shape of each video in LLM.
        rope_deltas (`torch.LongTensor` of shape `(batch_size, )`, *optional*):
            The rope index difference between sequence length and multimodal rope.
"""


class HithinkOmniForConditionalGeneration(HithinkOmniPreTrainedModel, GenerationMixin):
    _tied_weights_keys = ["lm_head.weight"]
    config_class = HithinkOmniConfig
    _no_split_modules = ["HithinkOmniDecoderLayer", "HithinkOmniVisionBlock"]

    def __init__(self, config: HithinkOmniConfig):
        super().__init__(config)
        self.visual = HithinkVisionTransformerPretrainedModel._from_config(config.vision_config)
        self.audio_tower = HithinkAudioEncoder._from_config(config.audio_config, attn_implementation=config._attn_implementation)
        self.multi_modal_projector = HithinkAudioMultiModalProjector(config)
        self.model = HithinkOmniModel(config)
        self.turn_taking_head = nn.Linear(config.hidden_size, 1, bias=False)
        self.vocab_size = config.vocab_size
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
        if config.vocab_size_ext:
            self.lm_head_ext = nn.Linear(config.hidden_size, config.vocab_size_ext, bias=False)
        else:
            self.lm_head_ext = None
        self.rope_deltas = None  # cache rope_deltas here

        if config.audio_decoder_config is None:
            self.audio_decoder = None
        else:
            audio_decoder_config = copy.deepcopy(config)
            audio_decoder_config.vocab_size = None
            audio_decoder_config.vocab_size_ext = None
            audio_decoder_config.num_hidden_layers = config.audio_decoder_config.num_hidden_layers
            self.audio_decoder = HithinkOmniModel(audio_decoder_config)

            speech_vocab_size = config.audio_decoder_config.codebook_size + 3  # +3 means bos, eos and pad
            num_codebooks = config.audio_decoder_config.num_codebooks
            self.codebook_embeddings = nn.ModuleList([
                nn.Embedding(speech_vocab_size, config.hidden_size) for _ in range(num_codebooks)
            ])
            self.codebook_heads = nn.Linear(config.hidden_size, speech_vocab_size * num_codebooks)
            self.num_codebooks = num_codebooks

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

        if CrossEntropyLoss.__module__.startswith('flash_attn'):
            self.z_loss = 0.0
        else:
            logger.warn("CrossEntropyLoss is not installed. If you want to accelerate training please install: "
                        "https://github.com/Dao-AILab/flash-attention")

    def audio_first_chunk_size(self, output_length):
        """流式推理时，输入音频第一个chunk的大小"""
        return 4 * output_length - 1

    def audio_next_chunk_size(self, output_length):
        """流式推理时，输入音频后续chunk的大小"""
        return 4 * output_length

    def audio_prev_chunk_overlap(self):
        """流式推理时，输入音频后续chunk与之前chunk重合部分的大小"""
        return 3

    def convert_audio_ids_to_embedding(self, audio_ids):
        audio_embedding = torch.stack([
            self.codebook_embeddings[i](audio_ids[:, :, i]) for i in range(self.num_codebooks)
        ]).sum(dim=0).to(self.model.embed_tokens.weight.dtype)  # (M, o_seq_len, hidden_size)
        audio_pad_token_id = self.config.audio_decoder_config.codebook_size + 2
        audio_mask = audio_ids.ne(audio_pad_token_id).any(dim=-1) # (M, o_seq_len,)
        return audio_embedding, audio_mask

    def get_input_embeddings(self):
        return self.model.embed_tokens

    def set_input_embeddings(self, value):
        self.model.embed_tokens = value

    def get_output_embeddings(self):
        return self.lm_head

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

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

    def get_decoder(self):
        return self.model

    def get_rope_index(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        image_grid_thw: Optional[torch.LongTensor] = None,
        video_grid_thw: Optional[torch.LongTensor] = None,
        second_per_grid_ts: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        """
        Calculate the 3D rope index based on image and video's temporal, height and width in LLM.

        Explanation:
            Each embedding sequence contains vision embedding and text embedding or just contains text embedding.

            For pure text embedding sequence, the rotary position embedding has no difference with modern LLMs.
            Examples:
                input_ids: [T T T T T], here T is for text.
                temporal position_ids: [0, 1, 2, 3, 4]
                height position_ids: [0, 1, 2, 3, 4]
                width position_ids: [0, 1, 2, 3, 4]

            For vision and text embedding sequence, we calculate 3D rotary position embedding for vision part
            and 1D rotary position embeddin for text part.
            Examples:
                Temporal (Time): 3 patches, representing different segments of the video in time.
                Height: 2 patches, dividing each frame vertically.
                Width: 2 patches, dividing each frame horizontally.
                We also have some important parameters:
                fps (Frames Per Second): The video's frame rate, set to 1. This means one frame is processed each second.
                tokens_per_second: This is a crucial parameter. It dictates how many "time-steps" or "temporal tokens" are conceptually packed into a one-second interval of the video. In this case, we have 25 tokens per second. So each second of the video will be represented with 25 separate time points. It essentially defines the temporal granularity.
                temporal_patch_size: The number of frames that compose one temporal patch. Here, it's 2 frames.
                interval: The step size for the temporal position IDs, calculated as tokens_per_second * temporal_patch_size / fps. In this case, 25 * 2 / 1 = 50. This means that each temporal patch will be have a difference of 50 in the temporal position IDs.
                input_ids: [V V V V V V V V V V V V T T T T T], here V is for vision.
                vision temporal position_ids: [0, 0, 0, 0, 50, 50, 50, 50, 100, 100, 100, 100]
                vision height position_ids: [0, 0, 1, 1, 0, 0, 1, 1, 0, 0, 1, 1]
                vision width position_ids: [0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1]
                text temporal position_ids: [101, 102, 103, 104, 105]
                text height position_ids: [101, 102, 103, 104, 105]
                text width position_ids: [101, 102, 103, 104, 105]
                Here we calculate the text start position_ids as the max vision position_ids plus 1.

        Args:
            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
                Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
                it.
            image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
                The temporal, height and width of feature shape of each image in LLM.
            video_grid_thw (`torch.LongTensor` of shape `(num_videos, 3)`, *optional*):
                The temporal, height and width of feature shape of each video in LLM.
            second_per_grid_ts (`torch.Tensor` of shape `(num_videos)`, *optional*):
                The time interval (in seconds) for each grid along the temporal dimension in the 3D position IDs.
            attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
                Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:

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

        Returns:
            position_ids (`torch.LongTensor` of shape `(3, batch_size, sequence_length)`)
            mrope_position_deltas (`torch.Tensor` of shape `(batch_size)`)
        """
        spatial_merge_size = self.config.vision_config.spatial_merge_size
        image_token_id = self.config.image_token_id
        video_token_id = self.config.video_token_id
        vision_start_token_id = self.config.vision_start_token_id
        mrope_position_deltas = []
        if input_ids is not None and (image_grid_thw is not None or video_grid_thw is not None):
            total_input_ids = input_ids
            if attention_mask is None:
                attention_mask = torch.ones_like(total_input_ids)
            position_ids = torch.ones(
                3,
                input_ids.shape[0],
                input_ids.shape[1],
                dtype=input_ids.dtype,
                device=input_ids.device,
            )
            image_index, video_index = 0, 0
            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 = 0, 0
                vision_start_indices = torch.argwhere(input_ids == vision_start_token_id).squeeze(1)
                vision_tokens = input_ids[vision_start_indices + 1]
                image_nums = (vision_tokens == image_token_id).sum()
                video_nums = (vision_tokens == video_token_id).sum()
                input_tokens = input_ids.tolist()
                llm_pos_ids_list: list = []
                st = 0
                remain_images, remain_videos = image_nums, video_nums
                for _ in range(image_nums + video_nums):
                    if image_token_id in input_tokens and remain_images > 0:
                        ed_image = input_tokens.index(image_token_id, st)
                    else:
                        ed_image = len(input_tokens) + 1
                    if video_token_id in input_tokens and remain_videos > 0:
                        ed_video = input_tokens.index(video_token_id, st)
                    else:
                        ed_video = len(input_tokens) + 1
                    if ed_image < ed_video:
                        t, h, w = (
                            image_grid_thw[image_index][0],
                            image_grid_thw[image_index][1],
                            image_grid_thw[image_index][2],
                        )
                        second_per_grid_t = 0
                        image_index += 1
                        remain_images -= 1
                        ed = ed_image

                    else:
                        t, h, w = (
                            video_grid_thw[video_index][0],
                            video_grid_thw[video_index][1],
                            video_grid_thw[video_index][2],
                        )
                        if second_per_grid_ts is not None:
                            second_per_grid_t = second_per_grid_ts[video_index]
                        else:
                            second_per_grid_t = 1.0
                        video_index += 1
                        remain_videos -= 1
                        ed = ed_video
                    llm_grid_t, llm_grid_h, llm_grid_w = (
                        t.item(),
                        h.item() // spatial_merge_size,
                        w.item() // spatial_merge_size,
                    )
                    text_len = ed - st

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

                    range_tensor = torch.arange(llm_grid_t).view(-1, 1)
                    expanded_range = range_tensor.expand(-1, llm_grid_h * llm_grid_w)

                    time_tensor = expanded_range * second_per_grid_t * self.config.vision_config.tokens_per_second

                    time_tensor_long = time_tensor.long()
                    t_index = time_tensor_long.flatten()

                    h_index = torch.arange(llm_grid_h).view(1, -1, 1).expand(llm_grid_t, -1, llm_grid_w).flatten()
                    w_index = torch.arange(llm_grid_w).view(1, 1, -1).expand(llm_grid_t, llm_grid_h, -1).flatten()
                    llm_pos_ids_list.append(torch.stack([t_index, h_index, w_index]) + text_len + st_idx)
                    st = ed + llm_grid_t * llm_grid_h * llm_grid_w

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

                llm_positions = torch.cat(llm_pos_ids_list, dim=1).reshape(3, -1)
                position_ids[..., i, attention_mask[i] == 1] = llm_positions.to(position_ids.device)
                mrope_position_deltas.append(llm_positions.max() + 1 - len(total_input_ids[i]))
            mrope_position_deltas = torch.tensor(mrope_position_deltas, device=input_ids.device).unsqueeze(1)
            return position_ids, mrope_position_deltas
        else:
            if attention_mask is not None:
                position_ids = attention_mask.long().cumsum(-1) - 1
                position_ids.masked_fill_(attention_mask == 0, 1)
                position_ids = position_ids.unsqueeze(0).expand(3, -1, -1).to(attention_mask.device)
                max_position_ids = position_ids.max(0, keepdim=False)[0].max(-1, keepdim=True)[0]
                mrope_position_deltas = max_position_ids + 1 - attention_mask.shape[-1]
            else:
                position_ids = (
                    torch.arange(input_ids.shape[1], device=input_ids.device)
                    .view(1, 1, -1)
                    .expand(3, input_ids.shape[0], -1)
                )
                mrope_position_deltas = torch.zeros(
                    [input_ids.shape[0], 1],
                    device=input_ids.device,
                    dtype=input_ids.dtype,
                )

            return position_ids, mrope_position_deltas

    @add_start_docstrings_to_model_forward(HITHINKOMNI_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=HithinkOmniCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
    def forward(
        self,
        input_ids: torch.LongTensor = None,
        input_features: torch.FloatTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        feature_attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[List[torch.FloatTensor]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        audio_past_key_values: Optional[Cache] = None,
        audio_use_cache: Optional[bool] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        output_logits: bool = True,
        return_dict: Optional[bool] = None,
        pixel_values: Optional[torch.Tensor] = None,
        pixel_values_videos: Optional[torch.FloatTensor] = None,
        image_grid_thw: Optional[torch.LongTensor] = None,
        video_grid_thw: Optional[torch.LongTensor] = None,
        rope_deltas: Optional[torch.LongTensor] = None,
        cache_position: Optional[torch.LongTensor] = None,
        second_per_grid_ts: Optional[torch.Tensor] = None,
        audio_ids: Optional[torch.LongTensor] = None,  # input of audio_decoder
        add_image: Optional[bool] = None,
        add_video: Optional[bool] = None,
        add_audio: Optional[bool] = None,
        is_turn_taking: Optional[bool] = None,
    ) -> Union[Tuple, HithinkOmniCausalLMOutputWithPast]:
        r"""
        Args:
            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]`.

        Returns:

        Example:

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

        >>> model = HithinkOmniForConditionalGeneration.from_pretrained("hithink-omni-qw25-audio-vocab-split")
        >>> processor = AutoProcessor.from_pretrained("hithink-omni-qw25-audio-vocab-split")

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

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

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

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

        # for audio part
        target_device = self.audio_tower.device
        if input_features is not None:
            # print(f'we have input audio feature')
            input_features = input_features.to(target_device)
            feature_attention_mask = feature_attention_mask.to(target_device)

        if inputs_embeds is None:
            inputs_embeds = self.model.convert_token_ids_to_embedding(input_ids)

            # print(f'inputs_embeds: {inputs_embeds.shape}')
            visual_dtype = self.visual.dtype
            if pixel_values is not None:
                pixel_values = pixel_values.type(visual_dtype)
                image_embeds = self.visual(pixel_values, grid_thw=image_grid_thw)
                n_image_tokens = (input_ids == self.config.image_token_id).sum().item()
                n_image_features = image_embeds.shape[0]
                if n_image_tokens != n_image_features:
                    raise ValueError(
                        f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}"
                    )

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

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

            elif add_image:  # 当前样本不包含图像，需要添加dummy图像（为保证多卡训练时模型参数/梯度同步一致性）
                # from PIL import Image
                # images = [Image.new('RGB', (32, 32), (0, 0, 0))]
                # media_inputs = processor.image_processor(images=images, videos=None, return_tensors='pt')
                m = self.visual.patch_embed
                d = 3 * m.temporal_patch_size * m.patch_size * m.patch_size
                pixels = torch.ones((16, d), dtype=visual_dtype, device=inputs_embeds.device)
                grid_thw = torch.tensor([[1, 4, 4]], device=inputs_embeds.device)
                image_embeds = self.visual(pixels, grid_thw)
                inputs_embeds += image_embeds.mean() * 0.

            if pixel_values_videos is not None:
                pixel_values_videos = pixel_values_videos.type(visual_dtype)
                video_embeds = self.visual(pixel_values_videos, grid_thw=video_grid_thw)
                n_video_tokens = (input_ids == self.config.video_token_id).sum().item()
                n_video_features = video_embeds.shape[0]
                if n_video_tokens != n_video_features:
                    raise ValueError(
                        f"Video features and video tokens do not match: tokens: {n_video_tokens}, features {n_video_features}"
                    )

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

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

            elif add_video:  # 当前样本不包含视频，需要添加dummy图像（为保证多卡训练时模型参数/梯度同步一致性）
                # from PIL import Image
                # images = [Image.new('RGB', (32, 32), (0, 0, 0))]
                # media_inputs = processor.image_processor(images=images, videos=None, return_tensors='pt')
                m = self.visual.patch_embed
                d = 3 * m.temporal_patch_size * m.patch_size * m.patch_size
                pixels = torch.ones((16, d), dtype=visual_dtype, device=inputs_embeds.device)
                grid_thw = torch.tensor([[1, 4, 4]], device=inputs_embeds.device)
                image_embeds = self.visual(pixels, grid_thw)
                inputs_embeds += image_embeds.mean() * 0.

            # merge audio and text embeddings
            if input_features is not None and (input_ids.shape[1] != 1 or audio_use_cache):
                audio_feat_lengths, audio_output_lengths = self.audio_tower._get_feat_extract_output_lengths(
                    feature_attention_mask.sum(-1)
                )
                # print(f'input_features: {input_features.shape}')
                audio_outputs = self.audio_tower(
                    input_features,
                    attention_mask=feature_attention_mask,
                    past_key_values=audio_past_key_values,
                    use_cache=audio_use_cache,
                )
                selected_audio_feature = audio_outputs.last_hidden_state
                audio_features = self.multi_modal_projector(selected_audio_feature)
                # print(f'audio_features: {audio_features.shape}')
                num_audios, max_audio_tokens, embed_dim = audio_features.shape
                audio_features_mask = torch.arange(max_audio_tokens, device=audio_features.device) \
                    .expand(num_audios, max_audio_tokens) < audio_output_lengths.unsqueeze(1)
                audio_features = audio_features[audio_features_mask].view(-1, embed_dim)
                audio_features = audio_features.to(inputs_embeds.device, inputs_embeds.dtype)
                audio_token_mask = (input_ids == self.config.audio_token_index).unsqueeze(-1).expand_as(inputs_embeds)
                inputs_embeds = inputs_embeds.masked_scatter(audio_token_mask, audio_features)
                # print(f'after mege audio and text embeddings: {inputs_embeds.shape}')

            elif add_audio:  # 当前样本不包含音频，需要添加dummy音频（为保证多卡训练时模型参数/梯度同步一致性）
                m = self.audio_tower
                expected_seq_length = 3000  # m.config.max_source_positions * m.conv1.stride[0] * m.conv2.stride[0]
                n_channels = 128  # m.conv1.weight.size(1)
                audio_outputs = m(torch.ones((1, n_channels, expected_seq_length), dtype=m.dtype, device=m.device))
                audio_features = self.multi_modal_projector(audio_outputs.last_hidden_state)
                inputs_embeds += audio_features.mean() * 0.

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

        # if we get 4D attention mask we cannot calculate rope deltas anymore. TODO @raushan fixme
        if position_ids is None and (attention_mask is None or attention_mask.ndim == 2):
            # calculate RoPE index once per generation in the pre-fill stage only
            if (
                (cache_position is not None and cache_position[0] == 0)
                or self.rope_deltas is None
                or (past_key_values is None or past_key_values.get_seq_length() == 0)
            ):
                position_ids, rope_deltas = self.get_rope_index(
                    input_ids,
                    image_grid_thw,
                    video_grid_thw,
                    second_per_grid_ts,
                    attention_mask,
                )
                self.rope_deltas = rope_deltas
            # then use the prev pre-calculated rope-deltas to get the correct position ids
            else:
                batch_size, seq_length, _ = inputs_embeds.shape
                delta = (
                    (cache_position[0] + self.rope_deltas).to(inputs_embeds.device)
                    if cache_position is not None
                    else 0
                )
                position_ids = torch.arange(seq_length, device=inputs_embeds.device)
                position_ids = position_ids.view(1, -1).expand(batch_size, -1)
                if cache_position is not None:  # otherwise `deltas` is an int `0`
                    delta = delta.repeat_interleave(batch_size // delta.shape[0], dim=0)
                position_ids = position_ids.add(delta)
                position_ids = position_ids.unsqueeze(0).expand(3, -1, -1)

        outputs = self.model(
            input_ids=None,
            position_ids=position_ids,
            attention_mask=attention_mask,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            cache_position=cache_position,
        )

        hidden_states = outputs[0]

        if is_turn_taking:  # 全双工语音分类（判断用户说话是否结束）
            logits = self.turn_taking_head(hidden_states)
            loss = None
            if labels is not None:
                loss_fct = BCEWithLogitsLoss()
                logits = logits.float()
                loss_mask = labels != -100
                loss = loss_fct(logits[loss_mask].squeeze(-1), labels[loss_mask].to(logits.dtype))

        elif self.training and labels is not None:  # 针对训练场景，降低显存占用
            is_train_flash = CrossEntropyLoss.__module__.startswith('flash_attn')
            if is_train_flash:
                loss_fct = CrossEntropyLoss(inplace_backward=True, lse_square_scale=self.z_loss)  # inplace_backward - saves memory
            else:
                loss_fct = CrossEntropyLoss()

            if audio_ids is not None:  # audio decoder 训练
                assert self.audio_decoder.config._attn_implementation != 'flash_attention_2', \
                        'Audio decoder 训练不支持 FlashAttention-2，建议使用 attn_implementation="sdpa"'

                audio_embeds, audio_attention_mask = self.convert_audio_ids_to_embedding(audio_ids)
                output_start_idx = (labels != -100).int().argmax(dim=1).unsqueeze(1)
                bsz, a_len, _ = audio_embeds.shape
                ii = torch.arange(bsz).unsqueeze(1).expand(-1, a_len)
                jj = torch.arange(a_len, device=audio_embeds.device).unsqueeze(0).expand(bsz, -1) + output_start_idx
                jj.masked_fill_(~audio_attention_mask, 0)
                final_embeds = (hidden_states[ii, jj] + audio_embeds) / (1 + self.num_codebooks)

                audio_hidden_states = self.audio_decoder(
                    inputs_embeds=final_embeds,
                    attention_mask=audio_attention_mask
                )[0]  # (batch_size, len, hidden_size)
                logits = self.codebook_heads(audio_hidden_states)  # (batch_size, len, codebook_num * (codebook_vocab_size + 3))

                codebook_size = self.config.audio_decoder_config.codebook_size
                audio_bos_token_id = codebook_size
                audio_pad_token_id = codebook_size + 2
                label_mask = (audio_ids == audio_bos_token_id) | (audio_ids == audio_pad_token_id)
                audio_labels = audio_ids.masked_fill(label_mask, loss_fct.ignore_index)
                logits = logits[:, :-1].reshape(-1, codebook_size + 3)
                labels = audio_labels[:, 1:].reshape(-1)

            else:  # 文本 labels 训练
                hidden_states = hidden_states[..., :-1, :]
                labels = labels[..., 1:]
                loss_mask = labels != loss_fct.ignore_index  # 只取需要计算loss的token，可节省大量显存！
                hidden_states = hidden_states[loss_mask].contiguous()
                logits = self.lm_head(hidden_states)
                if self.lm_head_ext is not None:
                    logits = torch.cat((logits, self.lm_head_ext(hidden_states)), dim=-1)
                if not is_train_flash:
                    logits = logits.float()
                labels = labels[loss_mask].contiguous().to(logits.device)

            loss = loss_fct(logits, labels)
            logits = None

        elif output_logits:
            logits = self.lm_head(hidden_states)
            if self.lm_head_ext is not None:
                logits = torch.cat((logits, self.lm_head_ext(hidden_states)), dim=-1)

            loss = None
            if labels is not None:
                # Upcast to float if we need to compute the loss to avoid potential precision issues
                logits = logits.float()
                # Shift so that tokens < n predict n
                shift_logits = logits[..., :-1, :].contiguous()
                shift_labels = labels[..., 1:].contiguous()
                # Flatten the tokens
                loss_fct = CrossEntropyLoss()
                shift_logits = shift_logits.view(-1, shift_logits.size(-1))
                shift_labels = shift_labels.view(-1)
                # Enable model parallelism
                shift_labels = shift_labels.to(shift_logits.device)
                loss = loss_fct(shift_logits, shift_labels)

        else:
            loss = None
            logits = None

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

        return HithinkOmniCausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            audio_past_key_values=audio_outputs.past_key_values if audio_use_cache else None,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
            rope_deltas=self.rope_deltas,
        )

    def prepare_inputs_for_generation(
        self,
        input_ids,
        input_features = None,
        feature_attention_mask = None,
        past_key_values=None,
        attention_mask=None,
        inputs_embeds=None,
        cache_position=None,
        position_ids=None,
        use_cache=True,
        pixel_values=None,
        pixel_values_videos=None,
        image_grid_thw=None,
        video_grid_thw=None,
        second_per_grid_ts=None,
        **kwargs,
    ):
        # Overwritten -- in specific circumstances we don't want to forward image inputs to the model

        # If we have cache: let's slice `input_ids` through `cache_position`, to keep only the unprocessed tokens
        # Exception 1: when passing input_embeds, input_ids may be missing entries
        # Exception 2: some generation methods do special slicing of input_ids, so we don't need to do it here
        if past_key_values is not None:
            if inputs_embeds is not None:  # Exception 1
                input_ids = input_ids[:, -cache_position.shape[0] :]
            elif input_ids.shape[1] != cache_position.shape[0]:  # Default case (the "else", a no op, is Exception 2)
                input_ids = input_ids[:, cache_position]

        if cache_position[0] != 0:
            pixel_values = None
            pixel_values_videos = None

        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
        if inputs_embeds is not None and cache_position[0] == 0:
            model_inputs = {"inputs_embeds": inputs_embeds, "input_ids": None}
        else:
            model_inputs = {"input_ids": input_ids, "inputs_embeds": None}

        if isinstance(past_key_values, StaticCache) and attention_mask.ndim == 2:
            if model_inputs["inputs_embeds"] is not None:
                batch_size, sequence_length, _ = inputs_embeds.shape
                device = inputs_embeds.device
            else:
                batch_size, sequence_length = input_ids.shape
                device = input_ids.device

            attention_mask = self.model._prepare_4d_causal_attention_mask_with_cache_position(
                attention_mask,
                sequence_length=sequence_length,
                target_length=past_key_values.get_max_cache_shape(),
                dtype=self.lm_head.weight.dtype,
                device=device,
                cache_position=cache_position,
                batch_size=batch_size,
                config=self.config,
                past_key_values=past_key_values,
            )

        model_inputs.update(
            {
                "input_features": input_features,
                "feature_attention_mask": feature_attention_mask,
                "position_ids": position_ids,
                "past_key_values": past_key_values,
                "use_cache": use_cache,
                "attention_mask": attention_mask,
                "pixel_values": pixel_values,
                "pixel_values_videos": pixel_values_videos,
                "image_grid_thw": image_grid_thw,
                "video_grid_thw": video_grid_thw,
                "cache_position": cache_position,
                "second_per_grid_ts": second_per_grid_ts,
            }
        )
        # import pdb; pdb.set_trace()
        return model_inputs

    @torch.no_grad()
    def stream_inference(
        self,
        input_ids: torch.LongTensor = None,
        input_features: torch.FloatTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
        position_ids: Optional[torch.Tensor] = None,
        feature_attention_mask: Optional[torch.Tensor] = None,
        pixel_values: Optional[torch.Tensor] = None,
        pixel_values_videos: Optional[torch.FloatTensor] = None,
        image_grid_thw: Optional[torch.LongTensor] = None,
        video_grid_thw: Optional[torch.LongTensor] = None,
        cache_position: Optional[torch.LongTensor] = None,
        output_txts: Optional[torch.LongTensor] = None,
        first_chunk_size: int = 25,
        chunk_size: int = 50,
        max_new_tokens: int = 512,
        top_k: int = 25,
        top_p: float = 0.95,
        temperature: float = 1.0,
        repeat_penalty: float = 1.0,
        length_penalty_params: Tuple[int, float] = (20, 1.01),
        **kwargs,
    ):
        """Streaming version of inference that yields partial results for batch_size=1"""
        assert input_ids.shape[0] == 1, "stream_inference only supports batch_size=1"

        # import pdb; pdb.set_trace()
        outputs = self(
            input_ids=input_ids,
            input_features=input_features,
            attention_mask=attention_mask,
            feature_attention_mask=feature_attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            use_cache=True,
            output_hidden_states=True,
            return_dict=True,
            pixel_values=pixel_values,
            pixel_values_videos=pixel_values_videos,
            image_grid_thw=image_grid_thw,
            video_grid_thw=video_grid_thw,
            cache_position=cache_position,
        )
        past_key_values = outputs.past_key_values
        last_hidden_state = outputs.hidden_states[-1][:, -1]

        # Initialize audio transformer states
        a_last_hidden_state = None
        a_past_key_values = None
        audio_attention_mask = None

        codebook_size = self.config.audio_decoder_config.codebook_size
        audio_bos_token_id = codebook_size
        audio_eos_token_id = codebook_size + 1
        audio_pad_token_id = codebook_size + 2

        # Initialize generation tensors
        hyps = torch.ones((1, self.num_codebooks + 1, max_new_tokens), dtype=torch.long, device=self.device) * audio_pad_token_id
        hyps[:, 0] = self.config.pad_token_id

        # if output_txts is not None, then the model is only responsible for generating the audio tokens
        if output_txts is not None:
            output_txts = output_txts.to(self.device)
            hyps[:, 0, :output_txts.shape[1]] = output_txts

        # Track generation completion
        end_flag = torch.zeros((1, self.num_codebooks + 1), dtype=torch.bool, device=self.device)

        # first_coden_idx = self.num_codebooks - 1 + 1 + chunk_size - 1
        first_coden_idx = self.num_codebooks - 1 + 1 + first_chunk_size - 1

        for i in range(max_new_tokens):

            if output_txts is None:
                text_logits = self.lm_head(last_hidden_state)
                if length_penalty_params[1] != 1.0:
                    text_logits = length_penalty(hyps[:, 0, :i], text_logits, length_penalty_params, self.config.eos_token_id)
                # text_token = text_logits.argmax(dim=-1)
                text_token = sample(hyps[:, :1, :i], text_logits.unsqueeze(1), top_p=0.6, temperature=0.6).squeeze(1)
                text_token = text_token.masked_fill(end_flag[:, 0], self.config.pad_token_id)
                hyps[:, 0, i] = text_token

            # generate the audio tokens
            if i == 0:
                hyps[:, 1:, i] = audio_bos_token_id
            else:
                audio_logits = self.codebook_heads(a_last_hidden_state.to(self.codebook_heads.weight.dtype))
                audio_logits = audio_logits.view(1, self.num_codebooks, codebook_size + 3)
                audio_token = sample(hyps[:, 1:, :i], audio_logits, top_k=top_k, top_p=top_p, temperature=temperature, repeat_penalty=repeat_penalty)
                # Apply masking
                audio_token = audio_token.masked_fill(end_flag[:, 1:], audio_pad_token_id)
                delay_mask = torch.arange(self.num_codebooks).expand(1, -1).to(self.device) > (i-1)
                audio_token = audio_token.masked_fill(delay_mask, audio_pad_token_id)
                hyps[:, 1:, i] = audio_token

            # Update end flag
            end_flag[:, 0] = end_flag[:, 0] | (hyps[:, 0, i] == self.config.eos_token_id)
            end_flag[:, 1:] = end_flag[:, 1:] | (hyps[:, 1:, i] == audio_eos_token_id)

            # Yield when buffer is full or generation is complete
            if i == first_coden_idx or (i > first_coden_idx and (i - first_coden_idx) % chunk_size == 0) or i == max_new_tokens - 1 or end_flag.all():
                if i <= first_coden_idx:
                    buffer_size = i - self.num_codebooks + 1
                elif (i - first_coden_idx) % chunk_size == 0:
                    buffer_size = chunk_size
                else:
                    buffer_size = (i - first_coden_idx) % chunk_size
                buffer = []
                for j in range(self.num_codebooks + 1):
                    st_idx = i - buffer_size + 1 - (self.num_codebooks - j)
                    buffer.append(hyps[:, j, st_idx:st_idx+buffer_size])  # (1, buffer_size)
                buffer = torch.cat(buffer, dim=0)  # (codebook_num + 1, buffer_size)
                text_tokens = buffer[0].unsqueeze(0)  # (1, buffer_size)
                audio_tokens = buffer[1:]  # (codebook_num, buffer_size)
                audio_token_len = (audio_tokens[0].ne(audio_pad_token_id) & audio_tokens[0].ne(audio_eos_token_id)).sum().item()
                audio_tokens = audio_tokens[:, :audio_token_len]  # (codebook_num, audio_token_len)
                pad_token_mask = audio_tokens.eq(audio_pad_token_id)
                eos_token_mask = audio_tokens.eq(audio_eos_token_id)
                if pad_token_mask.sum().item() > 0:
                    audio_tokens = audio_tokens.masked_fill(pad_token_mask, torch.randint(0, codebook_size, (1,)).item())
                if eos_token_mask.sum().item() > 0:
                    audio_tokens = audio_tokens.masked_fill(eos_token_mask, torch.randint(0, codebook_size, (1,)).item())
                yield text_tokens, [audio_tokens.cpu().numpy()], past_key_values, attention_mask

            # Stop if all batch ended
            if end_flag.all():
                break

            # update the last_hidden_state and past_key_values
            this_input_ids = hyps[:, :, i]
            this_input_ids = this_input_ids.unsqueeze(1)
            this_inputs_text_embeds = self.model.convert_token_ids_to_embedding(this_input_ids[:, :, 0])
            this_inputs_audio_embeds, audio_att_mask = self.convert_audio_ids_to_embedding(this_input_ids[:, :, 1:])
            attention_mask = torch.cat([attention_mask, this_input_ids[:, 0, 0].unsqueeze(-1).ne(self.config.pad_token_id)], dim=1)
            last_hidden_state, past_key_values = self.model(
                attention_mask=attention_mask,
                past_key_values=past_key_values,
                inputs_embeds=this_inputs_text_embeds,
                use_cache=True,
                return_dict=False,
            )
            last_hidden_state = last_hidden_state[:, -1]
            this_inputs_audio_embeds = (this_inputs_audio_embeds + last_hidden_state.unsqueeze(1)) / (1 + self.num_codebooks)
            if audio_attention_mask is None:
                audio_attention_mask = audio_att_mask
            else:
                audio_attention_mask = torch.cat([audio_attention_mask, audio_att_mask], dim=1)
            # if the audio tokens are too long, then we need to truncate the audio tokens, it's wrong, it will affect the prediction
            # if a_past_key_values is not None and a_past_key_values[0][0].shape[2] >= audio_chunk_max_tokens:
            #     audio_attention_mask = audio_attention_mask[:, audio_chunk_max_tokens//2+1:]
            #     # concat the first one and the last half
            #     a_past_key_values = tuple(
            #         tuple(torch.cat([kv[:, :, :1, :], kv[:, :, audio_chunk_max_tokens//2:, :]], dim=2) for kv in layer)
            #         for layer in a_past_key_values
            #     )

            a_last_hidden_state, a_past_key_values = self.audio_decoder(
                inputs_embeds=this_inputs_audio_embeds,
                attention_mask=audio_attention_mask.to(torch.int64), 
                past_key_values=a_past_key_values, 
                use_cache=True,
                return_dict=False,
            )
            a_last_hidden_state = a_last_hidden_state[:, -1]


def length_penalty(input_ids, scores, length_penalty_params=(20, 1.01), eos_token_id=None):
    """
        Args:
            input_ids: torch.LongTensor of shape (batch_size, seq_len)
            scores: torch.FloatTensor of shape (batch_size, vocab_size)
            length_penalty_params: Tuple[int, float]
            eos_token_id: int
    """
    processor = ExponentialDecayLengthPenalty(length_penalty_params, eos_token_id=eos_token_id, input_ids_seq_length=1)
    scores = processor(input_ids, scores)
    return scores


def sample(
        input_ids: torch.LongTensor,
        scores: torch.Tensor,
        top_k: int = 50,
        top_p: float = 0.95,
        temperature: float = 1.0,
        repeat_penalty: float = 1.0,
        ):
    """
        Args:
            input_ids: torch.LongTensor of shape (batch_size, C, seq_len)
            scores: torch.FloatTensor of shape (batch_size, C, vocab_size)
            top_k: int
            top_p: float
            temperature: float
            repeat_penalty: float
            C is the number of channels
    """
    repeat_logits_processor = RepetitionPenaltyLogitsProcessor(repeat_penalty)
    top_p_logits_processor = TopPLogitsWarper(top_p)
    top_k_logits_processor = TopKLogitsWarper(top_k)
    temperature_logits_processor = TemperatureLogitsWarper(temperature)

    assert len(input_ids.shape) == 3, f"input_ids shape is {input_ids.shape}"
    assert len(scores.shape) == 3, f"scores shape is {scores.shape}"
    assert input_ids.shape[1] == scores.shape[1], f"input_ids shape is {input_ids.shape}, scores shape is {scores.shape}"
    batch_size, c, seq_len = input_ids.shape
    for i in range(c):
        input_id = input_ids[:, i]
        score = scores[:, i]
        # Apply repetition penalty
        score = repeat_logits_processor(input_id, score)
        # Apply top-p sampling
        score = top_p_logits_processor(input_id, score)
        # Apply top-k sampling
        # score = top_k_logits_processor(input_id, score)
        # Apply temperature
        score = temperature_logits_processor(input_id, score)
        # replace to the scores
        scores[:, i] = score
    
    # sample
    probs = scores.softmax(dim=-1) # (batch_size, C, vocab_size)
    q = torch.empty_like(probs).exponential_(1)
    return torch.argmax((probs / q), dim=-1) # (batch_size, C)