modeling_bailing_moe_v2.py

# coding=utf-8
# Copyright 2023 Antgroup and The HuggingFace Inc. team. All rights reserved.
#
# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
# and OPT implementations in this library. It has been modified from its
# original forms to accommodate minor architectural differences compared
# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""PyTorch BailingMoE model."""
import math
import warnings
import os
from typing import List, Optional, Tuple, Union

import torch
import torch.nn.functional as F
import torch.utils.checkpoint
from torch import nn
from torch.nn import CrossEntropyLoss
# import transformer_engine.pytorch as te
from transformers.activations import ACT2FN
from transformers.cache_utils import Cache, DynamicCache
from transformers.modeling_attn_mask_utils import (
    AttentionMaskConverter,
    _prepare_4d_attention_mask,
    _prepare_4d_causal_attention_mask,
    _prepare_4d_causal_attention_mask_for_sdpa,
)
from transformers.modeling_outputs import (
    MoeModelOutputWithPast,
    MoeCausalLMOutputWithPast,
)
from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
from transformers.modeling_utils import PreTrainedModel
from transformers.pytorch_utils import ALL_LAYERNORM_LAYERS, is_torch_greater_or_equal_than_1_13
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 transformers.utils.import_utils import is_torch_fx_available
from configuration_bailing_moe_v2 import BailingMoeV2Config
from transformers.generation.utils import GenerationMixin
from modeling_utils import patch_continuous_features, build_modality_mask



if is_flash_attn_2_available():
    from flash_attn import flash_attn_func, flash_attn_varlen_func
    from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input  # noqa


# This makes `_prepare_4d_causal_attention_mask` a leaf function in the FX graph.
# It means that the function will not be traced through and simply appear as a node in the graph.
if is_torch_fx_available():
    if not is_torch_greater_or_equal_than_1_13:
        import torch.fx

    _prepare_4d_causal_attention_mask = torch.fx.wrap(_prepare_4d_causal_attention_mask)


logger = logging.get_logger(__name__)

_CONFIG_FOR_DOC = "BailingMoeV2Config"


def _get_unpad_data(attention_mask):
    seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
    indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
    max_seqlen_in_batch = seqlens_in_batch.max().item()
    cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.torch.int32), (1, 0))
    return (
        indices,
        cu_seqlens,
        max_seqlen_in_batch,
    )


def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
    warnings.warn(
        "Calling `transformers.models.BailingMoeV2.modeling_BailingMoeV2._prepare_4d_attention_mask` is deprecated and will be removed in v4.37. Use `transformers.modeling_attn_mask_utils._prepare_4d_attention_mask"
    )
    return _prepare_4d_attention_mask(mask=mask, dtype=dtype, tgt_len=tgt_len)


def _make_causal_mask(
    input_ids_shape: torch.Size, dtype: torch.dtype, device: torch.device, past_key_values_length: int = 0
):
    warnings.warn(
        "Calling `transformers.models.BailingMoeV2.modeling_BailingMoeV2._make_causal_mask` is deprecated and will be removed in v4.37. Use `transformers.models.BailingMoeV2.modeling_BailingMoeV2.AttentionMaskConverter._make_causal_mask"
    )
    return AttentionMaskConverter._make_causal_mask(
        input_ids_shape=input_ids_shape, dtype=dtype, device=device, past_key_values_length=past_key_values_length
    )


class BailingMoeV2RMSNorm(nn.Module):
    def __init__(self, hidden_size, eps=1e-6):
        """
        BailingMoeV2RMSNorm is equivalent to T5LayerNorm
        """
        super().__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
        self.variance_epsilon = eps

    def forward(self, hidden_states):
        input_dtype = hidden_states.dtype
        hidden_states = hidden_states.to(torch.float32)
        variance = hidden_states.pow(2).mean(-1, keepdim=True)
        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
        return self.weight * hidden_states.to(input_dtype)

    def reset_parameters(self):
        nn.init.ones_(self.weight)  # 显式重置 for enabling fsdp


ALL_LAYERNORM_LAYERS.append(BailingMoeV2RMSNorm)


class BailingMoeV2RotaryEmbedding(nn.Module):
    def __init__(self, config: BailingMoeV2Config, device=None):
        super().__init__()
        # BC: "rope_type" was originally "type"
        if hasattr(config, "rope_scaling") and config.rope_scaling is not None:
            self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
        else:
            self.rope_type = "default"
        self.max_seq_len_cached = config.max_position_embeddings
        self.original_max_seq_len = config.max_position_embeddings

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

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

    @torch.no_grad()
    @dynamic_rope_update  # power user: used with advanced RoPE types (e.g. dynamic rope)
    def forward(self, x, position_ids):
        position_ids = position_ids[:, 0, :] 
        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
        position_ids_expanded = position_ids[:, None, :].float()


        device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
        with torch.autocast(device_type=device_type, enabled=False):  # Force float32
            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
            emb = torch.cat((freqs, freqs), dim=-1)
            cos = emb.cos() * self.attention_scaling
            sin = emb.sin() * self.attention_scaling

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

    def reset_parameters(self) -> None:
        new_inv_freq, self.attention_scaling = self.rope_init_fn(self.config, None)
        self.inv_freq.copy_(new_inv_freq)
    
    
class BailingMoeV2RotaryEmbedding3D(nn.Module):
    def __init__(self, config: BailingMoeV2Config, device=None):
        super().__init__()
        # BC: "rope_type" was originally "type"
        self.rope_init_type = "default"
        self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
        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_init_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
        self.config = config

    @torch.no_grad()
    @dynamic_rope_update  # power user: used with advanced RoPE types (e.g. dynamic rope)
    def forward(self, x, position_ids):
        if self.rope_type == "3D" or self.rope_type == "video_rope":
            inv_freq_expanded = self.inv_freq[None, None, :, None].float().expand(3, position_ids.shape[1], -1, 1).to(x.device)
            position_ids_expanded = position_ids[:, :, None, :].float()
        else:
            inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
            position_ids_expanded = position_ids[:, None, :].float()

        device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
        with torch.autocast(device_type=device_type, enabled=False):  # Force float32
            if self.rope_type == "3D" or self.rope_type == "video_rope":
                freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(2, 3)
            else:
                freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
            emb = torch.cat((freqs, freqs), dim=-1)
            cos = emb.cos() * self.attention_scaling
            sin = emb.sin() * self.attention_scaling

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

    def reset_parameters(self) -> None:
        new_inv_freq, self.attention_scaling = self.rope_init_fn(self.config, None)
        self.inv_freq.copy_(new_inv_freq)


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


def apply_3d_rotary_pos_emb(
    q, 
    k, 
    cos, 
    sin, 
    mrope_section=[8, 12, 12], 
    unsqueeze_dim=1, 
    rope_type="m_rope", 
    rotary_half=True
):
    """Applies Rotary Position Embedding with Multimodal Sections to the query and key tensors (https://qwenlm.github.io/blog/qwen2-vl/).
    Explanation:
        Multimodal 3D rotary position embedding is an extension to 1D rotary position embedding. The input embedding
        sequence contains vision (images / videos) embedding and text embedding or just contains text embedding. For
        vision embedding part, we apply rotary position embedding on temporal, height and width dimension separately.
        Here we split the channel dimension to 3 chunks for the temporal, height and width rotary position embedding.
        For text embedding part, we just apply 1D rotary position embedding. The three rotary position index (temporal,
        height and width) of text embedding is always the same, so the text embedding rotary position embedding has no
        difference with modern LLMs.
    Args:
        q (`torch.Tensor`): The query tensor.
        k (`torch.Tensor`): The key tensor.
        cos (`torch.Tensor`): The cosine part of the rotary embedding.
        sin (`torch.Tensor`): The sine part of the rotary embedding.
        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.
        rope_type (`str`, *optional*, defaults to "m_rope"):
        rotary_half (`bool`, *optional*, defaults to `False`): Keep half or full tensor for later concatenation
    Returns:
        `tuple(torch.Tensor)` comprising the query and key tensors rotated using the Rotary Position Embedding.
    """
    if rope_type == "3D":  # rename rope_type
        rope_type = "m_rope"

    if rope_type == "m_rope":
        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).to(q.device)
        sin = torch.cat([m[i % 3] for i, m in enumerate(sin.split(mrope_section, dim=-1))], dim=-1).unsqueeze(unsqueeze_dim).to(q.device)
    elif rope_type == "video_rope":
        mrope_section = list(mrope_section)
        mrope_section = [mrope_section[0], mrope_section[1] + mrope_section[2]]
        mrope_section = mrope_section * 2
        # adjust t last -> (48, 16, 48, 16)
        mrope_section = mrope_section[::-1]
        index = 0
        result_cos = []
        result_sin = []
        # get x1, y1, x2, y2, ..., t1, t2, ...
        for i, section in enumerate(mrope_section):
            if i % 2 == 0:
                for j in range(section):
                    row = 1 if j % 2 == 0 else 2
                    result_cos.append(cos[row, ..., index: index + 1])
                    result_sin.append(sin[row, ..., index: index + 1])
                    index += 1
            else:
                result_cos.append(cos[0, ..., index:index + section])
                result_sin.append(sin[0, ..., index:index + section])
                index += section
        cos, sin = torch.cat(result_cos, dim=-1).unsqueeze(dim=unsqueeze_dim).to(q.device), torch.cat(result_sin, dim=-1).unsqueeze(dim=unsqueeze_dim).to(q.device)
    else:  # vanilla rope for llm
        cos = cos.unsqueeze(unsqueeze_dim).to(q.device)
        sin = sin.unsqueeze(unsqueeze_dim).to(q.device)

    if rotary_half:
        rotary_dim = cos.shape[-1]
        q_rot, q_pass = q[..., :rotary_dim], q[..., rotary_dim:]
        k_rot, k_pass = k[..., :rotary_dim], k[..., rotary_dim:]

        # Apply rotary embeddings on the first half
        q_embed = (q_rot * cos) + (rotate_half(q_rot) * sin)
        k_embed = (k_rot * cos) + (rotate_half(k_rot) * sin)

        # Concatenate back to full shape
        q_embed = torch.cat([q_embed, q_pass], dim=-1)
        k_embed = torch.cat([k_embed, k_pass], dim=-1)
    else:
        q_embed = (q * cos) + (rotate_half(q) * sin)
        k_embed = (k * cos) + (rotate_half(k) * sin)
    return q_embed, k_embed

def get_t_scale_rope_index(
    config,
    input_ids: torch.LongTensor,
    image_grid_thw: Optional[torch.LongTensor] = None,
    video_grid_thw: Optional[torch.LongTensor] = None,
    attention_mask: Optional[torch.Tensor] = None,
    scale_factor: float = 1.0,
    second_per_grid_ts: Optional[torch.Tensor] = None,
    use_interleaved_frame_timestamp: Optional = True,
) -> Tuple[torch.Tensor, torch.Tensor]:
    spatial_merge_size = config.spatial_merge_size
    image_token_id = config.image_patch_token
    video_token_id = config.video_patch_token
    image_start_token_id = config.image_start_token
    video_start_token_id = config.video_start_token
    use_abs_time_pos = second_per_grid_ts is not None


    mrope_position_deltas = []
    if 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,
        )

        if video_grid_thw is not None and use_interleaved_frame_timestamp:
            video_grid_thw = torch.repeat_interleave(
                video_grid_thw, video_grid_thw[:, 0], dim=0
            )
            video_grid_thw[:, 0] = 1

        image_index, video_index = 0, 0
        attention_mask = attention_mask.to(total_input_ids.device)

        for i, input_ids in enumerate(total_input_ids):
            if attention_mask is not None:
                input_ids = input_ids[attention_mask[i] == 1]
            image_nums, video_nums = 0, 0
            if image_grid_thw is not None:
                vision_start_indices = torch.argwhere(input_ids == image_start_token_id).squeeze(1)
                vision_tokens = input_ids[vision_start_indices + 1]
                image_nums = (vision_tokens == image_token_id).sum()
            if video_grid_thw is not None:
                if use_interleaved_frame_timestamp:
                    vision_start_indices = torch.argwhere(
                        input_ids == image_start_token_id
                    ).squeeze(1)
                    vision_tokens = input_ids[vision_start_indices + 1]
                    video_nums = (vision_tokens == video_token_id).sum()
                else:
                    vision_start_indices = torch.argwhere(
                        input_ids == video_start_token_id
                    ).squeeze(1)
                    vision_tokens = input_ids[vision_start_indices + 1]
                    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][0].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)

                # body-diagonal symmetry
                t_index = torch.arange(llm_grid_t).view(-1, 1).expand(
                    -1, llm_grid_h * llm_grid_w).flatten()
                h_index = torch.arange(llm_grid_h).view(1, -1, 1).expand(
                    llm_grid_t, -1, llm_grid_w).flatten() - (llm_grid_h - 1) // 2
                w_index = torch.arange(llm_grid_w).view(1, 1, -1).expand(
                    llm_grid_t, llm_grid_h, -1).flatten() - (llm_grid_w - 1) // 2

                # time dim adjust step size
                if use_abs_time_pos:
                    t_index = t_index * second_per_grid_t * scale_factor
                else:
                    t_index = t_index * scale_factor
                t_index = t_index + text_len + st_idx

                h_index = h_index + t_index
                w_index = w_index + t_index
                llm_pos_ids_list.append(
                    torch.stack([t_index, h_index, w_index]))
                st = ed + llm_grid_t * llm_grid_h * llm_grid_w

            if st < len(input_tokens):
                # next text token near last video token position = last t + 1
                st_idx = llm_pos_ids_list[-1][0].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)
            llm_positions = llm_positions.to(dtype=position_ids.dtype, device=position_ids.device)
            position_ids[..., i, attention_mask[i] == 1] = llm_positions
            # generate first token = last t + 1
            mrope_position_deltas.append(llm_positions[0].max() + 1 - len(total_input_ids[i]))
        mrope_position_deltas = torch.tensor(mrope_position_deltas, device=input_ids.device).unsqueeze(1)
    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(input_ids.device)
            max_position_ids = position_ids.max(0, keepdim=False)[0].max(-1, keepdim=True)[0]
            mrope_position_deltas = max_position_ids + 1 - attention_mask.shape[-1]
        else:
            position_ids = (
                torch.arange(input_ids.shape[1], device=input_ids.device)
                .view(1, 1, -1)
                .expand(3, input_ids.shape[0], -1)
            )
            mrope_position_deltas = torch.zeros(
                [input_ids.shape[0], 1],
                device=input_ids.device,
                dtype=input_ids.dtype,
            )

    return position_ids, mrope_position_deltas

def get_rope_index(
    config,
    input_ids: Optional[torch.LongTensor] = None,
    image_grid_thw: Optional[torch.LongTensor] = None,
    video_grid_thw: Optional[torch.LongTensor] = None,
    attention_mask: Optional[torch.Tensor] = None,
    second_per_grid_ts: Optional[torch.Tensor] = None,
    use_interleaved_frame_timestamp: Optional = True,
) -> 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 = config.spatial_merge_size
        image_token_id = config.image_patch_token
        video_token_id = config.video_patch_token
        image_start_token_id = config.image_start_token
        video_start_token_id = config.video_start_token

        use_abs_time_pos = second_per_grid_ts is not None

        mrope_position_deltas = []
        if 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,
            )

            if video_grid_thw is not None and use_interleaved_frame_timestamp:
                video_grid_thw = torch.repeat_interleave(
                    video_grid_thw, video_grid_thw[:, 0], dim=0
                )
                video_grid_thw[:, 0] = 1

            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
                if image_grid_thw is not None:
                    vision_start_indices = torch.argwhere(input_ids == image_start_token_id).squeeze(1)
                    vision_tokens = input_ids[vision_start_indices + 1]
                    image_nums = (vision_tokens == image_token_id).sum()
                if video_grid_thw is not None:
                    if use_interleaved_frame_timestamp:
                        vision_start_indices = torch.argwhere(
                            input_ids == image_start_token_id
                        ).squeeze(1)
                        vision_tokens = input_ids[vision_start_indices + 1]
                        video_nums = (vision_tokens == video_token_id).sum()
                    else:
                        vision_start_indices = torch.argwhere(
                            input_ids == video_start_token_id
                        ).squeeze(1)
                        vision_tokens = input_ids[vision_start_indices + 1]
                        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)
                    if use_abs_time_pos:
                        time_tensor = expanded_range * second_per_grid_t * config.tokens_per_second
                        time_tensor_long = time_tensor.long()
                    else:
                        time_tensor_long = expanded_range.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)
        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(input_ids.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


class BailingMoeV2MLP(nn.Module):
    def __init__(self, config: BailingMoeV2Config, intermediate_size: int):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.intermediate_size = 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, image_mask=None, audio_mask=None):
        return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))


class BailingMoeV2Gate(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.top_k = config.num_experts_per_tok
        self.num_experts = config.num_experts

        self.n_group = config.n_group
        self.topk_group = config.topk_group

        # topk selection algorithm
        self.gating_dim = config.hidden_size
        self.weight = nn.Parameter(torch.empty((self.num_experts, self.gating_dim)))
        self.routed_scaling_factor = config.routed_scaling_factor
        self.bias_update_coeff = 0.001

        self.expert_bias = torch.nn.Parameter(torch.zeros(self.num_experts), requires_grad=False)
        self.reset_parameters()

    def reset_parameters(self) -> None:
        import torch.nn.init as init

        init.kaiming_uniform_(self.weight, a=math.sqrt(5))

    def update_bias(self, distribution: torch.Tensor):
        with torch.no_grad():
            delta_bias = (distribution.mean() - distribution).sign()
        self.expert_bias.data = self.expert_bias.data + self.bias_update_coeff * delta_bias

    def group_limited_topk(
        self,
        scores: torch.Tensor,
    ):
        num_tokens, _ = scores.size()
        # Organize the experts into groups
        group_scores = scores.view(num_tokens, self.n_group, -1).topk(2, dim=-1)[0].sum(dim=-1)
        group_idx = torch.topk(group_scores, k=self.topk_group, dim=-1, sorted=False)[1]
        group_mask = torch.zeros_like(group_scores)
        group_mask.scatter_(1, group_idx, 1)

        # Mask the experts based on selection groups
        score_mask = (
            group_mask.unsqueeze(-1)
            .expand(num_tokens, self.n_group, self.num_experts // self.n_group)
            .reshape(num_tokens, -1)
        )

        masked_scores = scores.masked_fill(~score_mask.bool(), float('-inf'))
        probs, top_indices = torch.topk(masked_scores, k=self.top_k, dim=-1, sorted=False)

        return probs, top_indices

    def forward(self, hidden_states):
        # compute gating score
        hidden_states = hidden_states.view(-1, hidden_states.shape[-1])
        logits = F.linear(hidden_states.type(torch.float32), self.weight.type(torch.float32))

        scores = torch.sigmoid(logits)

        scores_for_routing = scores + self.expert_bias
        _, topk_idx = self.group_limited_topk(scores_for_routing)

        scores = torch.gather(scores, dim=1, index=topk_idx).type_as(logits)

        topk_weight = scores / (scores.sum(dim=-1, keepdim=True) + 1e-20) if self.top_k > 1 else scores
        topk_weight = topk_weight * self.routed_scaling_factor

        return topk_idx, topk_weight, logits


class BailingMoeV2SparseMoeBlock(nn.Module):
    """
    A mixed expert module containing shared experts.
    """

    def __init__(self, config: BailingMoeV2Config):
        super().__init__()
        self.config = config
        self.num_experts_per_tok = config.num_experts_per_tok
        self._setup_experts()
        self.router_type = config.router_type
        if self.router_type == "topN":
            logger.info(f"use topN Router")
            self.gate = BailingMoeV2Gate(config)
        elif self.router_type == "MultiRouter":
            logger.info(f"use MultiRouter")
            self.gate = BailingMoeV2Gate(config)
            self.image_gate = BailingMoeV2Gate(config)
            self.audio_gate = BailingMoeV2Gate(config)
        if config.num_shared_experts is not None:
            self.shared_experts = BailingMoeV2MLP(
                config=config, intermediate_size=config.moe_intermediate_size * config.num_shared_experts
            )

    def _setup_experts(self):
        self.experts = nn.ModuleList(
            [
                BailingMoeV2MLP(config=self.config, intermediate_size=self.config.moe_intermediate_size)
                for _ in range(self.config.num_experts)
            ]
        )

    def forward(self, hidden_states, image_mask, audio_mask):
        identity = hidden_states
        bsz, seq_len, h = hidden_states.shape
        if self.router_type == "MultiRouter":
            if image_mask is not None:
                if len(image_mask.shape) == 3:
                    assert image_mask.shape[-1] == 1
                elif len(image_mask.shape) == 2:
                    assert image_mask.shape == hidden_states.shape[:2]
                    image_mask = image_mask.unsqueeze(-1)
                else:
                    print('image mask shape, sum', image_mask.shape, image_mask.sum())
                    assert False
            if audio_mask is not None:
                if len(audio_mask.shape) == 3:
                    assert audio_mask.shape[-1] == 1
                elif len(audio_mask.shape) == 2:
                    assert audio_mask.shape == hidden_states.shape[:2]
                    audio_mask = audio_mask.unsqueeze(-1)
                else:
                    print('audio mask shape, sum', audio_mask.shape, audio_mask.sum())
                    assert False
            if image_mask is not None and audio_mask is not None:
                assert torch.logical_and(image_mask, audio_mask).sum() == 0
            
            image_topk_idx, image_topk_weight, image_router_logits = self.image_gate(hidden_states)
            audio_topk_idx, audio_topk_weight, audio_router_logits = self.audio_gate(hidden_states)
            topk_idx, topk_weight, router_logits = self.gate(hidden_states)

            if image_mask is not None:
                image_mask = image_mask.view(-1, 1)
                topk_idx = image_topk_idx * image_mask + topk_idx * torch.logical_not(image_mask)
                topk_weight = image_topk_weight * image_mask + topk_weight * torch.logical_not(image_mask)
                router_logits = image_router_logits * image_mask + router_logits * torch.logical_not(image_mask)
            if audio_mask is not None:
                audio_mask = audio_mask.view(-1, 1)
                audio_mask = audio_mask.to(router_logits.device)
                topk_idx = audio_topk_idx * audio_mask + topk_idx * torch.logical_not(audio_mask)
                topk_weight = audio_topk_weight * audio_mask + topk_weight * torch.logical_not(audio_mask)
                router_logits = audio_router_logits * audio_mask + router_logits * torch.logical_not(audio_mask)
        else:
            topk_idx, topk_weight, router_logits = self.gate(hidden_states)

        hidden_states = hidden_states.view(-1, hidden_states.shape[-1])
        flat_topk_idx = topk_idx.view(-1)
        y = self.moe_infer(hidden_states, topk_idx, topk_weight).view(bsz, seq_len, h)
        if self.config.num_shared_experts is not None:
            y = y + self.shared_experts(identity)
        return y, (router_logits.view(bsz, seq_len, -1), topk_idx.view(bsz, seq_len, -1))

    @torch.no_grad()
    def moe_infer(self, x, topk_ids, topk_weight):
        cnts = topk_ids.new_zeros((topk_ids.shape[0], len(self.experts)))
        cnts.scatter_(1, topk_ids, 1)
        tokens_per_expert = cnts.sum(dim=0)
        idxs = topk_ids.view(-1).argsort()
        sorted_tokens = x[idxs // topk_ids.shape[1]]
        sorted_tokens_shape = sorted_tokens.shape
        tokens_per_expert = tokens_per_expert.cpu().numpy()
        outputs = []
        start_idx = 0
        for i, num_tokens in enumerate(tokens_per_expert):
            end_idx = start_idx + num_tokens
            if num_tokens == 0:
                continue
            expert = self.experts[i]
            tokens_for_this_expert = sorted_tokens[start_idx:end_idx]
            expert_out = expert(tokens_for_this_expert)
            outputs.append(expert_out.to(x.device))
            start_idx = end_idx

        outs = torch.cat(outputs, dim=0) if len(outputs) else sorted_tokens.new_empty(0)
        new_x = torch.empty_like(outs)
        new_x[idxs] = outs
        final_out = (
            new_x.view(*topk_ids.shape, -1)
            .type(topk_weight.dtype)
            .mul_(topk_weight.unsqueeze(dim=-1))
            .sum(dim=1)
            .type(new_x.dtype)
        )
        return final_out


# Copied from transformers.models.llama.modeling_llama.repeat_kv
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)


# Copied from transformers.models.llama.modeling_llama.LlamaAttention with Llama->BailingMoeV2
class BailingMoeV2Attention(nn.Module):
    """Multi-headed attention from 'Attention Is All You Need' paper"""

    def __init__(self, config: BailingMoeV2Config, 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.attention_dropout = config.attention_dropout
        self.hidden_size = config.hidden_size
        self.num_heads = config.num_attention_heads
        self.head_dim = config.head_dim or self.hidden_size // self.num_heads
        partial_rotary_factor = config.partial_rotary_factor if hasattr(config, "partial_rotary_factor") else 1.0
        self.rope_dim = int(self.head_dim * partial_rotary_factor)
        self.num_key_value_heads = config.num_key_value_heads
        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
        self.max_position_embeddings = config.max_position_embeddings
        self.rope_theta = config.rope_theta
        self.is_causal = True

        self.query_key_value = nn.Linear(
            self.hidden_size,
            (self.num_heads + 2 * self.num_key_value_heads) * self.head_dim,
            bias=config.use_qkv_bias,
        )

        self.q_norm = BailingMoeV2RMSNorm(self.head_dim, eps=config.rms_norm_eps)
        self.k_norm = BailingMoeV2RMSNorm(self.head_dim, eps=config.rms_norm_eps)
        self.dense = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=config.use_bias)

        if self.config.rope_scaling is not None:
            self.rope_scaling = {"type": "mrope", "mrope_section": [8, 12, 12]}
            self.mrope_section = self.rope_scaling["mrope_section"]

    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,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_value: Optional[Cache] = None,
        output_attentions: bool = False,
        use_cache: bool = False,
        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
        cache_position: Optional[torch.LongTensor] = None,
        **kwargs,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        if "padding_mask" in kwargs:
            warnings.warn(
                "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
            )

        bsz, q_len, _ = hidden_states.size()

        qkv = self.query_key_value(hidden_states)
        qkv = qkv.view(bsz, q_len, self.num_heads + 2 * self.num_key_value_heads, self.head_dim)

        query_states, key_states, value_states = qkv.split(
            [self.num_heads, self.num_key_value_heads, self.num_key_value_heads], dim=-2
        )
        query_states = query_states.transpose(1, 2)
        key_states = key_states.transpose(1, 2)
        value_states = value_states.transpose(1, 2)

        if not self.training:
            query_states = query_states.contiguous()
            key_states = key_states.contiguous()
            value_states = value_states.contiguous()
            del qkv

        query_states = self.q_norm(query_states)
        key_states = self.k_norm(key_states)

        kv_seq_len = key_states.shape[-2]
        if past_key_value is not None:
            if self.layer_idx is None:
                raise ValueError(
                    f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
                    "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
                    "with a layer index."
                )
            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
        cos, sin = position_embeddings
        query_states, key_states = apply_3d_rotary_pos_emb(
            query_states, 
            key_states, 
            cos, 
            sin, 
            mrope_section=self.rope_scaling["mrope_section"], 
            rope_type=self.config.rope_scaling["type"]
        )
        
        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)

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

        # if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
        #     raise ValueError(
        #         f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
        #         f" {attn_weights.size()}"
        #     )

        if attention_mask is not None: # no matter the length, we just slice it
            # if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
            #     raise ValueError(
            #         f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
            #     )
            causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
            attn_weights = attn_weights + causal_mask

        # 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.dense(attn_output)

        if not output_attentions:
            attn_weights = None
        
        if not self.training:
            del query_states, key_states, value_states

        return attn_output, attn_weights, past_key_value


# Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2 with Llama->BailingMoeV2
class BailingMoeV2FlashAttention2(BailingMoeV2Attention):
    """
    BailingMoeV2 flash attention module. This module inherits from `BailingMoeV2Attention` as the weights of the module stays
    untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
    flash attention and deal with padding tokens in case the input contains any of them.
    """

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

        # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
        # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right 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.LongTensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_value: Optional[Cache] = None,
        output_attentions: bool = False,
        use_cache: bool = False,
        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
        cache_position: Optional[torch.LongTensor] = None,
        **kwargs,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        # BailingMoeV2FlashAttention2 attention does not support output_attentions
        if "padding_mask" in kwargs:
            warnings.warn(
                "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
            )

            # overwrite attention_mask with padding_mask
            attention_mask = kwargs.pop("padding_mask")

        output_attentions = False

        bsz, q_len, _ = hidden_states.size()

        # Flash attention requires the input to have the shape
        # batch_size x seq_length x head_dim x hidden_dim
        # therefore we just need to keep the original shape

        qkv = self.query_key_value(hidden_states)
        qkv = qkv.view(bsz, q_len, self.num_heads + 2 * self.num_key_value_heads, self.head_dim)

        query_states, key_states, value_states = qkv.split(
            [self.num_heads, self.num_key_value_heads, self.num_key_value_heads], dim=-2
        )
        query_states = query_states.transpose(1, 2)
        key_states = key_states.transpose(1, 2)
        value_states = value_states.transpose(1, 2)

        if not self.training:
            query_states = query_states.contiguous()
            key_states = key_states.contiguous()
            value_states = value_states.contiguous()
            del qkv

        query_states = self.q_norm(query_states)
        key_states = self.k_norm(key_states)

        # kv_seq_len = key_states.shape[-2]
        # if past_key_value is not None:
        #     kv_seq_len += past_key_value.get_seq_length(layer_idx=self.layer_idx)
        cos, sin = position_embeddings
        if self.config.rope_scaling is not None:
            rope_type = self.config.rope_scaling.get("rope_type", self.config.rope_scaling.get("type"))
        else:
            rope_type = "default"
        query_states, key_states = apply_3d_rotary_pos_emb(
            query_states,
            key_states,
            cos,
            sin,
            rope_type=rope_type
        )

        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)

        # 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.
        query_states = query_states.transpose(1, 2)
        key_states = key_states.transpose(1, 2)
        value_states = value_states.transpose(1, 2)

        dropout_rate = self.attention_dropout if self.training else 0.0

        # In PEFT, usually we cast the layer norms in float32 for training stability reasons
        # therefore the input hidden states gets silently cast in float32. Hence, we need
        # cast them back in the correct dtype just to be sure everything works as expected.
        # This might slow down training & inference so it is recommended to not cast the LayerNorms
        # in fp32. (BailingMoeV2RMSNorm handles it correctly)

        input_dtype = query_states.dtype
        if input_dtype == torch.float32:
            # Handle the case where the model is quantized
            if hasattr(self.config, "_pre_quantization_dtype"):
                target_dtype = self.config._pre_quantization_dtype
            elif torch.is_autocast_enabled():
                target_dtype = torch.get_autocast_gpu_dtype()
            else:
                target_dtype = self.query_key_value.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 = self._flash_attention_forward(
            query_states, key_states, value_states, attention_mask, q_len, dropout=dropout_rate
        )

        attn_output = attn_output.reshape(bsz, q_len, -1).contiguous()
        attn_output = self.dense(attn_output)

        if not output_attentions:
            attn_weights = None

        if not self.training:
            del query_states, key_states, value_states

        return attn_output, attn_weights, past_key_value

    def _flash_attention_forward(
        self, query_states, key_states, value_states, attention_mask, query_length, dropout=0.0, softmax_scale=None
    ):
        """
        Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token
        first unpad the input, then computes the attention scores and pad the final attention scores.

        Args:
            query_states (`torch.Tensor`):
                Input query states to be passed to Flash Attention API
            key_states (`torch.Tensor`):
                Input key states to be passed to Flash Attention API
            value_states (`torch.Tensor`):
                Input value states to be passed to Flash Attention API
            attention_mask (`torch.Tensor`):
                The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
                position of padding tokens and 1 for the position of non-padding tokens.
            dropout (`int`, *optional*):
                Attention dropout
            softmax_scale (`float`, *optional*):
                The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
            query_length (`int`):
                The length of the query sequence in terms of tokens. This represents the number of tokens in the
                `query_states` tensor along the sequence dimension. It is used to determine the effective sequence
                length for attention computations.
        """
        if not self._flash_attn_uses_top_left_mask:
            causal = self.is_causal
        else:
            # TODO: Remove the `query_length != 1` check once Flash Attention for RoCm is bumped to 2.1. For details, please see the comment in BailingMoeV2FlashAttention2 __init__.
            causal = self.is_causal and query_length != 1

        # Contains at least one padding token in the sequence
        if attention_mask is not None:
            batch_size = query_states.shape[0]
            query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input(
                query_states, key_states, value_states, attention_mask, query_length
            )

            cu_seqlens_q, cu_seqlens_k = cu_seq_lens
            max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens

            attn_output_unpad = flash_attn_varlen_func(
                query_states,
                key_states,
                value_states,
                cu_seqlens_q=cu_seqlens_q,
                cu_seqlens_k=cu_seqlens_k,
                max_seqlen_q=max_seqlen_in_batch_q,
                max_seqlen_k=max_seqlen_in_batch_k,
                dropout_p=dropout,
                softmax_scale=softmax_scale,
                causal=causal,
            )

            attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
        else:
            attn_output = flash_attn_func(
                query_states, key_states, value_states, dropout, softmax_scale=softmax_scale, causal=causal
            )

        return attn_output

    def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length):
        indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)
        batch_size, kv_seq_len, num_key_value_heads, head_dim = key_layer.shape

        key_layer = index_first_axis(
            key_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
        )
        value_layer = index_first_axis(
            value_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
        )
        if query_length == kv_seq_len:
            query_layer = index_first_axis(
                query_layer.reshape(batch_size * kv_seq_len, self.num_heads, head_dim), indices_k
            )
            cu_seqlens_q = cu_seqlens_k
            max_seqlen_in_batch_q = max_seqlen_in_batch_k
            indices_q = indices_k
        elif query_length == 1:
            max_seqlen_in_batch_q = 1
            cu_seqlens_q = torch.arange(
                batch_size + 1, dtype=torch.int32, device=query_layer.device
            )  # There is a memcpy here, that is very bad.
            indices_q = cu_seqlens_q[:-1]
            query_layer = query_layer.squeeze(1)
        else:
            # The -q_len: slice assumes left padding.
            attention_mask = attention_mask[:, -query_length:]
            query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask)

        return (
            query_layer,
            key_layer,
            value_layer,
            indices_q,
            (cu_seqlens_q, cu_seqlens_k),
            (max_seqlen_in_batch_q, max_seqlen_in_batch_k),
        )


ATTENTION_CLASSES = {
    "eager": BailingMoeV2Attention,
    "flash_attention_2": BailingMoeV2FlashAttention2,
}


class BailingMoeV2DecoderLayer(nn.Module):
    def __init__(self, config: BailingMoeV2Config, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size

        self.attention = ATTENTION_CLASSES[config._attn_implementation](config=config, layer_idx=layer_idx)

        self.mlp = (
            BailingMoeV2SparseMoeBlock(config)
            if (config.num_experts is not None and layer_idx >= config.first_k_dense_replace)
            else BailingMoeV2MLP(config=config, intermediate_size=config.intermediate_size)
        )
        self.input_layernorm = BailingMoeV2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = BailingMoeV2RMSNorm(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,
        cache_position: Optional[torch.LongTensor] = None,
        image_mask: Optional[torch.Tensor] = None,
        audio_mask: Optional[torch.Tensor] = None,
        past_key_value: Optional[Tuple[torch.Tensor]] = None,
        output_attentions: Optional[bool] = False,
        output_router_logits: Optional[bool] = False,
        use_cache: Optional[bool] = False,
        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_size, sequence_length)` if flash attention is used or `(batch_size, 1,
                query_sequence_length, key_sequence_length)` if default attention is used.
            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]`.
            past_key_value (`Tuple(torch.FloatTensor)`, *optional*):
                cached past key and value projection states
            output_attentions (`bool`, *optional*):
                Whether to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
            output_router_logits (`bool`, *optional*):
                Whether or not to return the logits of all the routers. They are useful for computing the router loss,
                and should not be returned during inference.
            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`).
            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.
        """
        if "padding_mask" in kwargs:
            warnings.warn(
                "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
            )
        residual = hidden_states

        hidden_states = self.input_layernorm(hidden_states)

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

        # Fully Connected
        residual = hidden_states
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = self.mlp(hidden_states, image_mask, audio_mask)
        #hidden_states = self.mlp(hidden_states)
        if isinstance(hidden_states, tuple):
            hidden_states, router_logits = hidden_states
        else:
            router_logits = None
        hidden_states = residual + hidden_states.to(residual.device)

        outputs = (hidden_states,)

        if output_attentions:
            outputs += (self_attn_weights,)

        if use_cache:
            outputs += (present_key_value,)

        if output_router_logits:
            outputs += (router_logits,)

        return outputs


BAILINGMOEV2_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 ([`BailingMoeV2Config`]):
            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 BailingMoeV2 Model outputting raw hidden-states without any specific head on top.",
    BAILINGMOEV2_START_DOCSTRING,
)
class BailingMoeV2PreTrainedModel(PreTrainedModel):
    config_class = BailingMoeV2Config
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["BailingMoeV2DecoderLayer"]
    _skip_keys_device_placement = "past_key_values"
    _supports_flash_attn_2 = True
    _supports_sdpa = True
    _supports_cache_class = True

    def _init_weights(self, module):
        std = self.config.initializer_range
        if isinstance(module, nn.Linear):
            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_()


BAILINGMOEV2_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 `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 (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*):
            Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
            blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values`
            returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`.

            Two formats are allowed:
            - a [`~cache_utils.Cache`] instance;
            - 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)`). This is also known as the legacy
            cache format.

            The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the
            legacy cache format will be returned.

            If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't
            have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `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.
"""


@add_start_docstrings(
    "The bare BailingMoeV2 Model outputting raw hidden-states without any specific head on top.",
    BAILINGMOEV2_START_DOCSTRING,
)
class BailingMoeV2Model(BailingMoeV2PreTrainedModel):
    """
    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`BailingMoeV2DecoderLayer`]

    Args:
        config: BailingMoeV2Config
    """

    def __init__(self, config: BailingMoeV2Config):
        super().__init__(config)
        self.padding_idx = config.pad_token_id
        self.vocab_size = config.vocab_size

        self.word_embeddings = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
        self.layers = nn.ModuleList(
            [BailingMoeV2DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self._use_sdpa = config._attn_implementation == "sdpa"
        self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"
        self.norm = BailingMoeV2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        if self.config.rope_scaling is not None:
            self.rotary_emb = BailingMoeV2RotaryEmbedding3D(config=config)
        else:
            self.rotary_emb = BailingMoeV2RotaryEmbedding(config=config)
        self.gradient_checkpointing = False
        config.spatial_merge_size = 2 
        config.tokens_per_second = 2
        self.rope_deltas = None
        # Initialize weights and apply final processing
        self.post_init()

    def get_input_embeddings(self):
        return self.word_embeddings

    def set_input_embeddings(self, value):
        self.word_embeddings = value
 
    def prompt_wrap_vision(self, input_ids, inputs_embeds, vision_embeds, vision_token_id):
        if vision_embeds is None or input_ids is None:
            return inputs_embeds

        if len(vision_embeds.shape) == 3:
            vision_embeds = vision_embeds.reshape(-1, vision_embeds.shape[-1])

        n_image_tokens = (input_ids == vision_token_id).sum().item()
        n_image_features = vision_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}"
            )
        vision_mask = (input_ids == vision_token_id).unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device)
        
        image_embeds = vision_embeds.to(inputs_embeds.device, inputs_embeds.dtype)
        inputs_embeds = inputs_embeds.masked_scatter(vision_mask, image_embeds)

        return inputs_embeds

    def prompt_wrap_audio(self, input_ids, inputs_embeds, audio_embeds, audio_embeds_lengths, placeholder_audio_loc_lens):
        inputs_embeds = patch_continuous_features(
           input_embeddings=inputs_embeds, placeholder_loc_lens=placeholder_audio_loc_lens,
           encoded_feats=audio_embeds, encoded_feat_lens=audio_embeds_lengths,
        )
        router_mask_audio = build_modality_mask(placeholder_audio_loc_lens, inputs_embeds.shape[:-1])
        router_mask_audio = router_mask_audio.to(inputs_embeds.device)
        return inputs_embeds, router_mask_audio
     
    def prompt_wrap_navit(self, input_ids, config, query_embeds_image=None, query_embeds_video=None, query_embeds_audio=None,
        query_embeds_audio_lengths=None, placeholder_audio_loc_lens=None, target_embeds=None):
        inputs_embeds = self.word_embeddings(input_ids)
        vision_mask = None
        audio_mask = None
        if query_embeds_image is None and query_embeds_video is None and query_embeds_audio is None and target_embeds is None:
            return inputs_embeds, vision_mask, audio_mask

        if query_embeds_image is not None:
            inputs_embeds = self.prompt_wrap_vision(input_ids, inputs_embeds, query_embeds_image, config.image_patch_token)
        if query_embeds_video is not None:
            inputs_embeds = self.prompt_wrap_vision(input_ids, inputs_embeds, query_embeds_video, config.video_patch_token)

        image_mask = input_ids == config.image_patch_token
        video_mask = input_ids == config.video_patch_token
        vision_mask = (image_mask + video_mask) > 0
        vision_mask = vision_mask.unsqueeze(-1).to(input_ids.device)

        if query_embeds_audio is not None:
            inputs_embeds, audio_mask = self.prompt_wrap_audio(
                input_ids, inputs_embeds, query_embeds_audio, query_embeds_audio_lengths, placeholder_audio_loc_lens,
            )

        return inputs_embeds, vision_mask, audio_mask


    @add_start_docstrings_to_model_forward(BAILINGMOEV2_INPUTS_DOCSTRING)
    def forward(
        self,
        input_ids: torch.LongTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        query_embeds_image: Optional[torch.Tensor] = None,
        query_embeds_video: Optional[torch.Tensor] = None,
        query_embeds_audio: Optional[torch.Tensor] = None,
        query_embeds_audio_lengths: Optional[torch.Tensor] = None,
        placeholder_audio_loc_lens: Optional[torch.Tensor] = None,
        target_embeds: Optional[torch.Tensor] = None, 
        past_key_values: Optional[List[torch.FloatTensor]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        image_grid_thw: Optional[torch.Tensor] = None,
        image_grid_thw_video: Optional[torch.Tensor] = None, 
        use_cache: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        output_router_logits: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        second_per_grid_ts: Optional[torch.Tensor] = None,
        image_mask=None,
        audio_mask=None,
        **kwargs,
    ) -> Union[Tuple, MoeModelOutputWithPast]:

        if inputs_embeds is not None:
            words_embeddings = inputs_embeds
            input_shape = inputs_embeds.size()[:2]
        else:
            if (query_embeds_image is None and query_embeds_video is None and query_embeds_audio is None and target_embeds is None) or input_ids.size(1) == 1:
                words_embeddings = self.word_embeddings(input_ids.clip(0, self.word_embeddings.weight.shape[0] - 1))
                input_shape = input_ids.size()
                image_mask = None
                audio_mask = None
            else:
                words_embeddings, image_mask, audio_mask = self.prompt_wrap_navit(
                    input_ids.clip(0, self.word_embeddings.weight.shape[0] - 1), self.config, query_embeds_image, query_embeds_video, query_embeds_audio,
                    query_embeds_audio_lengths, placeholder_audio_loc_lens, target_embeds,  # noqa
                )

                input_shape = words_embeddings.size()[:2]
        embeddings = words_embeddings

        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
        )
        output_router_logits = (
            output_router_logits if output_router_logits is not None else self.config.output_router_logits
        )
        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

        # retrieve input_ids and inputs_embeds
        if input_ids is not None and inputs_embeds is not None:
            assert input_ids.size(1) == inputs_embeds.size(1), "{} vs {}".format(
                input_ids.size,
                inputs_embeds.size,
            )
            batch_size, seq_length = inputs_embeds.shape[:2]
            #raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
        elif input_ids is not None:
            batch_size, seq_length = input_ids.shape[:2]
        elif inputs_embeds is not None:
            batch_size, seq_length = inputs_embeds.shape[:2]
        else:
            raise ValueError("You have to specify either 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`transformers."
                )
                use_cache = False

        # if use_cache and past_key_values is None:
        #     past_key_values = DynamicCache()
        if use_cache:
            use_legacy_cache = not isinstance(past_key_values, Cache)
            if use_legacy_cache:
                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.Tensor = torch.arange(
                past_seen_tokens, past_seen_tokens + input_ids.shape[1], device=input_ids.device
            )

        if self.rotary_emb.rope_type == "video_rope" and input_ids.size(1) != 1:
            position_ids, rope_deltas = get_t_scale_rope_index(
                self.config,
                input_ids,
                image_grid_thw,
                image_grid_thw_video,
                attention_mask,
                scale_factor=2.0,
                second_per_grid_ts=second_per_grid_ts,
                use_interleaved_frame_timestamp=self.config.use_interleaved_frame_timestamp,
            )
            self.rope_deltas = rope_deltas
        elif self.rotary_emb.rope_type == "3D" and input_ids.size(1) != 1:
            position_ids, rope_deltas = get_rope_index(
                self.config,
                input_ids,
                image_grid_thw,
                image_grid_thw_video,
                attention_mask,
                second_per_grid_ts,
                use_interleaved_frame_timestamp=self.config.use_interleaved_frame_timestamp,
            )
            self.rope_deltas = rope_deltas
        elif self.rotary_emb.rope_type == "3D" or self.rotary_emb.rope_type == "video_rope": # decode stage
            batch_size, seq_length = input_ids.shape
            if cache_position is not None:
                delta = (cache_position[0] + self.rope_deltas).to(input_ids.device)
            else:
                delta = torch.tensor(0).to(input_ids.device)
            position_ids = torch.arange(seq_length, device=input_ids.device)
            position_ids = position_ids.view(1, 1, -1).expand(3, batch_size, -1)
            # delta = (past_key_values_length + self.rope_deltas).to(input_ids.device)
            delta = delta.repeat_interleave(batch_size // delta.shape[0], dim=1)
            position_ids = position_ids + delta.to(position_ids.device)

        inputs_embeds = embeddings
        if position_ids is None:
            position_ids = cache_position.unsqueeze(0)

        
        
        causal_mask = self._update_causal_mask(
            attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions
        )
        # if self._use_flash_attention_2:
        #     # 2d mask is passed through the layers
        #     attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None
        # elif self._use_sdpa and not output_attentions:
        #     # output_attentions=True can not be supported when using SDPA, and we fall back on
        #     # the manual implementation that requires a 4D causal mask in all cases.
        #     attention_mask = _prepare_4d_causal_attention_mask_for_sdpa(
        #         attention_mask,
        #         (batch_size, seq_length),
        #         inputs_embeds,
        #         past_key_values_length,
        #     )
        # else:
        #     # 4d mask is passed through the layers
        #     attention_mask = _prepare_4d_causal_attention_mask(
        #         attention_mask, (batch_size, seq_length), inputs_embeds, past_key_values_length
        #     )

        # embed positions
        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
        all_router_logits = () if output_router_logits 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,
                    image_mask,
                    audio_mask,
                    past_key_values,
                    output_attentions,
                    output_router_logits,
                    use_cache,
                    position_embeddings,
                    cache_position,
                )
            else:
                layer_outputs = decoder_layer(
                    hidden_states,
                    attention_mask=causal_mask,
                    position_ids=position_ids,
                    image_mask=image_mask,
                    audio_mask=audio_mask,
                    past_key_value=past_key_values,
                    output_attentions=output_attentions,
                    output_router_logits=output_router_logits,
                    use_cache=use_cache,
                    position_embeddings=position_embeddings,
                    cache_position=cache_position,
                )
            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],)

            if output_router_logits and layer_outputs[-1] is not None:
                all_router_logits += (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 = None
        # use_legacy_cache = False
        if use_cache:
            next_cache = next_decoder_cache.to_legacy_cache() if use_legacy_cache else next_decoder_cache
        if not return_dict:
            return tuple(
                v
                for v in [hidden_states, next_cache, all_hidden_states, all_self_attns, all_router_logits]
                if v is not None
            )
        return MoeModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=next_cache,
            hidden_states=all_hidden_states,
            attentions=all_self_attns,
            router_logits=all_router_logits,
        )

    # Copied from transformers.models.phimoe.modeling_phimoe.PhimoeModel._update_causal_mask with Phimoe->BailingMoeV2
    def _update_causal_mask(
        self,
        attention_mask: Union[torch.Tensor],
        input_tensor: torch.Tensor,
        cache_position: torch.Tensor,
        past_key_values: Cache,
        output_attentions: bool = False,
    ):
        if self.config._attn_implementation == "flash_attention_2":
            if attention_mask is not None and past_key_values is not None:
                is_padding_right = attention_mask[:, -1].sum().item() != input_tensor.size()[0]
                if is_padding_right:
                    raise ValueError(
                        "You are attempting to perform batched generation with padding_side='right'"
                        " this may lead to unexpected behaviour for Flash Attention version of BailingV2Moe. Make sure to "
                        " call `tokenizer.padding_side  = 'left'` before tokenizing the input. "
                    )
            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)
        # 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 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=None,
                is_training=self.training,
            ):
                return None
        dtype = input_tensor.dtype
        min_dtype = torch.finfo(dtype).min
        sequence_length = input_tensor.shape[1]
        # StaticCache
        if 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,
            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 in ["cuda", "xpu", "npu"]
            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
    # Copied from transformers.models.phimoe.modeling_phimoe.PhimoeModel._prepare_4d_causal_attention_mask_with_cache_position with Phimoe->BailingV2Moe
    def _prepare_4d_causal_attention_mask_with_cache_position(
        attention_mask: torch.Tensor,
        sequence_length: int,
        target_length: int,
        dtype: torch.dtype,
        cache_position: torch.Tensor,
        batch_size: int,
        config: BailingMoeV2Config,
        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.
            cache_position (`torch.Tensor`):
                Indices depicting the position of the input sequence tokens in the sequence.
            batch_size (`torch.Tensor`):
                Batch size.
            config (`BailingV2MoeConfig`):
                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=cache_position.device
            )
            diagonal_attend_mask = torch.arange(target_length, device=cache_position.device) > cache_position.reshape(
                -1, 1
            )
            text_config = config.get_text_config()
            if (
                getattr(text_config, "use_sliding_window", True)
                and getattr(text_config, "sliding_window", None) 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
                is_static_sliding_cache = isinstance(past_key_values, StaticCache) and all(past_key_values.is_sliding)
                if not is_static_sliding_cache or sequence_length > target_length:
                    sliding_attend_mask = torch.arange(target_length, device=cache_position.device) <= (
                        cache_position.reshape(-1, 1) - text_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, :].to(
                    causal_mask.device
                )
                padding_mask = padding_mask == 0
                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
                    padding_mask, min_dtype
                )
        return causal_mask


class BailingMoeV2ForCausalLM(BailingMoeV2PreTrainedModel, GenerationMixin):
    _tied_weights_keys = ["lm_head.weight"]

    def __init__(self, config: BailingMoeV2Config):
        super().__init__(config)
        self.model = BailingMoeV2Model(config)
        self.vocab_size = config.vocab_size
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)

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

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

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

    def get_output_embeddings(self):
        return self.lm_head

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

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

    def get_decoder(self):
        return self.model

    @add_start_docstrings_to_model_forward(BAILINGMOEV2_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=MoeCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
    def forward(
        self,
        input_ids: torch.LongTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        query_embeds_image: Optional[torch.Tensor] = None,
        query_embeds_video: Optional[torch.Tensor] = None,
        query_embeds_audio: Optional[torch.Tensor] = None,
        query_embeds_audio_lengths: Optional[torch.Tensor] = None,
        placeholder_audio_loc_lens: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[List[torch.FloatTensor]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        image_grid_thw: Optional[torch.Tensor] = None,
        image_grid_thw_video: Optional[torch.Tensor] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        output_router_logits: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        second_per_grid_ts: Optional[torch.Tensor] = None,
        num_logits_to_keep: Optional[int] = 0,
        image_mask=None,
        audio_mask=None,
        **kwargs,
    ) -> Union[Tuple, MoeCausalLMOutputWithPast]:
        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 transformers import AutoTokenizer

        >>> model = BailingMoeV2ForCausalLM.from_pretrained(PATH_TO_CONVERTED_WEIGHTS)
        >>> tokenizer = AutoTokenizer.from_pretrained(PATH_TO_CONVERTED_TOKENIZER)

        >>> prompt = "Hey, are you conscious? Can you talk to me?"
        >>> inputs = tokenizer(prompt, return_tensors="pt")

        >>> # 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]
        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
        ```"""
        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
        )
        output_router_logits = (
            output_router_logits if output_router_logits is not None else self.config.output_router_logits
        )
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
        outputs = self.model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            query_embeds_image=query_embeds_image,
            query_embeds_video=query_embeds_video,
            query_embeds_audio=query_embeds_audio,
            query_embeds_audio_lengths=query_embeds_audio_lengths,
            placeholder_audio_loc_lens=placeholder_audio_loc_lens,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            image_grid_thw=image_grid_thw,
            image_grid_thw_video=image_grid_thw_video,
            use_cache=use_cache,
            cache_position=cache_position,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            output_router_logits=output_router_logits,
            return_dict=return_dict,
            second_per_grid_ts=second_per_grid_ts,
            image_mask=image_mask,
            audio_mask=audio_mask,
            **kwargs,
        )

        hidden_states = outputs[0]

        logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :])
        # logits = logits.float()

        loss = None
        aux_loss = None

        if labels is not None:
            # Shift so that tokens < n predict n
            shift_logits = logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            # Flatten the tokens
            loss_fct = CrossEntropyLoss(inplace_backward=True)
            shift_logits = shift_logits.view(-1, self.config.vocab_size)
            shift_labels = shift_labels.view(-1)
            # Enable model parallelism
            shift_labels = shift_labels.to(shift_logits.device)
            loss = loss_fct(shift_logits, shift_labels)

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

        return MoeCausalLMOutputWithPast(
            loss=loss,
            aux_loss=aux_loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
            router_logits=outputs.router_logits,
        )

    def prepare_inputs_for_generation(
        self,
        input_ids,
        query_embeds_image: torch.Tensor = None,
        query_embeds_video: torch.Tensor = None,
        query_embeds_audio: torch.Tensor = None,
        query_embeds_audio_lengths: torch.Tensor = None,
        placeholder_audio_loc_lens: torch.Tensor = None,
        image_mask=None,
        audio_mask=None,
        past_key_values=None,
        attention_mask=None,
        inputs_embeds=None,
        cache_position=None,
        position_ids=None,
        use_cache=True,
        image_grid_thw=None,
        image_grid_thw_video=None,
        second_per_grid_ts=None,
        is_audio_generation_mode=False,
        num_logits_to_keep=0,
        **kwargs,
    ):
        if past_key_values is not None:
            if isinstance(past_key_values, Cache):
                cache_length = past_key_values.get_seq_length()
                past_length = cache_length
                # 获取最大缓存长度（DynamicCache 使用 max_cache_len 属性）
                if hasattr(past_key_values, "max_cache_len") and past_key_values.max_cache_len is not None:
                    max_cache_length = past_key_values.max_cache_len
                else:
                    # 兼容其他 Cache 子类（如 StaticCache）
                    max_cache_length = (
                        past_key_values.get_max_length()
                        if hasattr(past_key_values, "get_max_length")
                        else getattr(past_key_values, "get_max_cache_shape", lambda: None)()
                    )
            else:
                cache_length = past_length = past_key_values[0][0].shape[2]
                max_cache_length = None

            # Keep only the unprocessed tokens:
            # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where
            # some of the inputs are exclusivelly passed as part of the cache (e.g. when passing input_embeds as input)
            if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]:
                input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :]
            # 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard
            # input_ids based on the past_length.
            elif past_length < input_ids.shape[1]:
                input_ids = input_ids[:, past_length:]
            # 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens.

            # If we are about to go beyond the maximum cache length, we need to crop the input attention mask.
            if (
                max_cache_length is not None
                and attention_mask is not None
                and cache_length + input_ids.shape[1] > max_cache_length
            ):
                attention_mask = attention_mask[:, -max_cache_length:]

        position_ids = kwargs.get("position_ids", None)
        if attention_mask is not None and position_ids is None:
            # create position_ids on the fly for batch generation
            position_ids = attention_mask.long().cumsum(-1) - 1
            position_ids.masked_fill_(attention_mask == 0, 1)
            if past_key_values:
                position_ids = position_ids[:, -input_ids.shape[1] :]
            position_ids = position_ids.unsqueeze(0)

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

        model_inputs.update(
            {
                "query_embeds_image": query_embeds_image,
                "query_embeds_video": query_embeds_video,
                "query_embeds_audio": query_embeds_audio,
                "position_ids": position_ids,
                "past_key_values": past_key_values,
                "use_cache": use_cache,
                "attention_mask": attention_mask,
                "image_grid_thw": image_grid_thw,
                "image_grid_thw_video": image_grid_thw_video,
                "image_mask": image_mask,
                "audio_mask": audio_mask,
                "query_embeds_audio_lengths": query_embeds_audio_lengths,
                "placeholder_audio_loc_lens": placeholder_audio_loc_lens,
                "num_logits_to_keep": num_logits_to_keep,
            }
        )
        return model_inputs

    @staticmethod
    def _reorder_cache(past_key_values, beam_idx):
        reordered_past = ()
        for layer_past in past_key_values:
            reordered_past += (
                tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past),
            )
        return reordered_past