modeling_qwen2_flash.py

# coding=utf-8
# transformers==4.39.2 NOTE
# Borrows some implementations from https://github.com/Cooperx521/PyramidDrop, thanks!
# Copyright 2024 The Qwen team, Alibaba Group 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 Qwen2 model."""
import inspect
import math
import warnings
from typing import List, Optional, Tuple, Union, Dict, Any, Iterable

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

from transformers.activations import ACT2FN
from transformers.cache_utils import Cache, DynamicCache
from transformers.modeling_attn_mask_utils import _prepare_4d_causal_attention_mask, _prepare_4d_causal_attention_mask_for_sdpa
from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast, SequenceClassifierOutputWithPast
from transformers.modeling_utils import PreTrainedModel
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.models.qwen2.configuration_qwen2 import Qwen2Config
from .constants import IGNORE_INDEX


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

    _flash_supports_window_size = "window_size" in list(inspect.signature(flash_attn_func).parameters)


logger = logging.get_logger(__name__)


_CHECKPOINT_FOR_DOC = "Qwen/Qwen2-7B-beta"
_CONFIG_FOR_DOC = "Qwen2Config"

QWEN2_PRETRAINED_MODEL_ARCHIVE_LIST = [
    "Qwen/Qwen2-7B-beta",
    # See all Qwen2 models at https://huggingface.co/models?filter=qwen2
]


# Copied from transformers.models.llama.modeling_llama._get_unpad_data
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.int32), (1, 0))
    return (
        indices,
        cu_seqlens,
        max_seqlen_in_batch,
    )


# Copied from transformers.models.llama.modeling_llama.LlamaRMSNorm with Llama->Qwen2
class Qwen2RMSNorm(nn.Module):
    def __init__(self, hidden_size, eps=1e-6):
        """
        Qwen2RMSNorm 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)


# Copied from transformers.models.mistral.modeling_mistral.MistralRotaryEmbedding with Mistral->Qwen2
class Qwen2RotaryEmbedding(nn.Module):
    def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
        super().__init__()
        self.dim = dim
        self.max_position_embeddings = max_position_embeddings
        self.base = base
        inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64).float().to(device) / self.dim))
        self.register_buffer("inv_freq", inv_freq, persistent=False)

    @torch.no_grad()
    def forward(self, x, position_ids):
        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):
            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
            emb = torch.cat((freqs, freqs), dim=-1)
            cos = emb.cos()
            sin = emb.sin()
        
        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)


# 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)


# Copied from transformers.models.mistral.modeling_mistral.apply_rotary_pos_emb
def apply_rotary_pos_emb(q, k, cos, sin, position_ids, unsqueeze_dim=1):
    """Applies Rotary Position Embedding to the query and key tensors.

    Args:
        q (`torch.Tensor`): The query tensor.
        k (`torch.Tensor`): The key tensor.
        cos (`torch.Tensor`): The cosine part of the rotary embedding.
        sin (`torch.Tensor`): The sine part of the rotary embedding.
        position_ids (`torch.Tensor`): Not used in dynamic RoPE, kept for compatibility.
        unsqueeze_dim (`int`, *optional*, defaults to 1):
            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos and sin.
    Returns:
        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
    """
    cos = cos.unsqueeze(unsqueeze_dim)
    sin = sin.unsqueeze(unsqueeze_dim)
    q_embed = (q * cos) + (rotate_half(q) * sin)
    k_embed = (k * cos) + (rotate_half(k) * sin)
    return q_embed, k_embed


def apply_rotary_pos_emb2(q_pe, k_pe, cos, sin, position_ids=None, unsqueeze_dim=1):
    cos = cos.unsqueeze(unsqueeze_dim)
    sin = sin.unsqueeze(unsqueeze_dim)

    # dim alignment
    step = cos.shape[-1] // q_pe.shape[-1]
    indices = torch.arange(0, cos.size(-1), step, device=cos.device)    
    cos = cos[..., indices]
    sin = sin[..., indices]

    q_embed = (q_pe * cos) + (rotate_half(q_pe) * sin)
    k_embed = (k_pe * cos) + (rotate_half(k_pe) * sin)
    return q_embed, k_embed


def apply_rotary_pos_emb2_single(tensor, cos, sin,  position_ids=None, unsqueeze_dim=1):

    cos = cos.unsqueeze(unsqueeze_dim)
    sin = sin.unsqueeze(unsqueeze_dim)

    # dim alignment
    step = cos.shape[-1] // tensor.shape[-1]
    indices = torch.arange(0, cos.size(-1), step, device=cos.device)    
    cos = cos[..., indices]
    sin = sin[..., indices]

    tensor_embed = (tensor * cos) + (rotate_half(tensor) * sin)
    return tensor_embed


def apply_rotary_pos_emb2_separate(q_pe, vision_k_pe, text_key_pe, cos, sin, vision_indices, text_indices, position_ids=None, unsqueeze_dim=1):
    cos = cos.unsqueeze(unsqueeze_dim)
    sin = sin.unsqueeze(unsqueeze_dim)
    step = cos.shape[-1] // q_pe.shape[-1]
    indices = torch.arange(0, cos.size(-1), step, device=cos.device)    
    cos = cos[..., indices]
    sin = sin[..., indices]
    
    # 对整个q_pe应用位置编码
    q_pe = (q_pe * cos) + (rotate_half(q_pe) * sin)
    
    # 为vision_k_pe选择对应位置的cos和sin
    if vision_indices[0].numel() > 0:
        vision_cos = cos[vision_indices[0], :, vision_indices[1], :]  # 直接用索引选择
        vision_sin = sin[vision_indices[0], :, vision_indices[1], :]
        vision_cos = vision_cos.squeeze(1).unsqueeze(0)
        vision_sin = vision_sin.squeeze(1).unsqueeze(0)
        vision_k_pe = (vision_k_pe * vision_cos) + (rotate_half(vision_k_pe) * vision_sin)
    # 为text_key_pe选择对应位置的cos和sin
    if text_indices[0].numel() > 0:
        text_cos = cos[text_indices[0], :, text_indices[1], :]  # 直接用索引选择
        text_sin = sin[text_indices[0], :, text_indices[1], :]
        text_cos = text_cos.squeeze(1).unsqueeze(0)
        text_sin = text_sin.squeeze(1).unsqueeze(0)
        text_key_pe = (text_key_pe * text_cos) + (rotate_half(text_key_pe) * text_sin)
    
    return q_pe, vision_k_pe, text_key_pe



# Copied from transformers.models.mistral.modeling_mistral.MistralMLP with Mistral->Qwen2
class Qwen2MLP(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size
        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
        self.act_fn = ACT2FN[config.hidden_act]

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


# 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)


class Qwen2Attention(nn.Module):
    """
    Multi-headed attention from 'Attention Is All You Need' paper. Modified to use sliding window attention: Longformer
    and "Generating Long Sequences with Sparse Transformers".
    """

    def __init__(self, config: Qwen2Config, layer_idx: Optional[int] = None):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx
        if layer_idx is None:
            logger.warning_once(
                f"Instantiating {self.__class__.__name__} without passing `layer_idx` is not recommended and will "
                "to errors during the forward call, if caching is used. Please make sure to provide a `layer_idx` "
                "when creating this class."
            )

        self.hidden_size = config.hidden_size
        self.num_heads = config.num_attention_heads
        self.head_dim = self.hidden_size // self.num_heads
        self.num_key_value_heads = config.num_key_value_heads
        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
        self.max_position_embeddings = config.max_position_embeddings
        self.rope_theta = config.rope_theta
        self.is_causal = True
        self.attention_dropout = config.attention_dropout
        
        self.softmax_temperature = nn.Parameter(torch.tensor(1.0))
        self.k_nope_scale_factor = nn.Parameter(torch.tensor(1.0))

        if (self.head_dim * self.num_heads) != self.hidden_size:
            raise ValueError(
                f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
                f" and `num_heads`: {self.num_heads})."
            )

        # 控制变量：处理模式
        # "mixed": 多模态分离处理（视觉压缩，文本完整）
        # "compress_all": 统一压缩处理
        # "no_compress": 统一不压缩处理（但仍使用部分RoPE形式）
        self.processing_mode = getattr(config, 'attention_processing_mode', 'compress_all')
        
        # 基础投影层
        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=True)
        self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=True)
        self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=True)
        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)

        self.rotary_emb = Qwen2RotaryEmbedding(
            self.head_dim,
            max_position_embeddings=self.max_position_embeddings,
            base=self.rope_theta,
        )

        # 共同的参数设置（所有模式都使用部分RoPE）
        config.attention_bias = False
        self.interval = 2 
        self.q_rope_dim = self.num_heads * (self.head_dim//self.interval)
        self.q_nope_dim = self.num_heads * self.head_dim - self.q_rope_dim
        self.k_rope_dim = self.num_key_value_heads * (self.head_dim//self.interval)
        self.k_nope_dim = self.num_key_value_heads*self.head_dim - self.k_rope_dim
        self.kv_lora_rank = 128      #4个kv头每个头的latentsize
        self.v_head_dim = self.head_dim

        self.kv_a_proj_nope = nn.Linear(self.hidden_size, self.num_key_value_heads*self.kv_lora_rank, bias=False)
        self.k_proj_pe = nn.Linear(self.hidden_size,self.k_rope_dim,bias=True)
        self.k_b_proj_nope = nn.Linear(self.num_key_value_heads*self.kv_lora_rank, self.num_key_value_groups*self.k_nope_dim, bias=True)
        self.v_b_proj = nn.Linear(self.num_key_value_heads*self.kv_lora_rank, self.num_key_value_groups*self.num_key_value_heads*self.head_dim, bias=True)
        # 新增属性存储当前处理的索引
        self.prefill_vision_indices = None
        self.prefill_text_indices = None
        self.prefill_seq_len = None


    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,
        vision_text_mask: Optional[torch.Tensor] = 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.`"
            )

        # 根据处理模式选择不同的forward逻辑
        if self.processing_mode == 'no_compress':
            return self._forward_no_compress(
                hidden_states, attention_mask, position_ids, past_key_value,
                output_attentions, use_cache, vision_text_mask, **kwargs
            )
        elif self.processing_mode == 'compress_all':
            return self._forward_compress_all(
                hidden_states, attention_mask, position_ids, past_key_value,
                output_attentions, use_cache, vision_text_mask, **kwargs
            )
        elif self.processing_mode == 'mixed':
            return self._forward_mixed(
                hidden_states, attention_mask, position_ids, past_key_value,
                output_attentions, use_cache, vision_text_mask, **kwargs
            )
        else:
            raise ValueError(f"Unsupported processing mode: {self.processing_mode}")

    def _forward_no_compress(
        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,
        vision_text_mask: Optional[torch.Tensor] = None,
        **kwargs,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        """不压缩处理，但使用部分RoPE形式"""
        bsz, q_len, _ = hidden_states.size()

        # 计算query states
        query_states = self.q_proj(hidden_states)
        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)

        # 分离query的pe和nope部分
        mask = torch.zeros(self.head_dim, dtype=torch.bool, device=hidden_states.device)
        mask[::self.interval] = True 
        q_pe = query_states[..., mask]
        q_nope = query_states[..., ~mask]

        # 计算key和value states（完整的，不压缩）
        key_states = self.k_proj(hidden_states)
        value_states = self.v_proj(hidden_states)
        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)

        # 从完整的key中分离pe部分（类似query的处理方式）
        k_pe = key_states[..., mask]  # 直接裁剪，不使用k_proj_pe
        k_nope = key_states[..., ~mask]

        # 处理缓存
        kv_seq_len = q_len
        if past_key_value is not None:
            if self.layer_idx is None:
                raise ValueError("layer_idx is required when using cache")
            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)

        # 应用旋转位置编码
        cos, sin = self.rotary_emb(k_pe, position_ids)
        q_pe, k_pe = apply_rotary_pos_emb2(q_pe, k_pe, cos, sin, position_ids)

        # 重新组合key states（pe部分使用RoPE后的结果，nope部分保持原样）
        key_states[..., mask] = k_pe
        key_states[..., ~mask] = k_nope

        # 更新缓存 - 统一使用MixedDynamicCache接口，用None占位不需要的部分
        if past_key_value is not None and use_cache:
            cache_kwargs = {"sin": sin, "cos": cos}
            # no_compress模式：文本用完整缓存，视觉部分用None占位
            _, _, final_key_states, final_value_states = past_key_value.update(
                vision_k_pe=None,  # 不压缩模式不需要
                vision_compressed_kv=None,  # 不压缩模式不需要
                text_key_states=key_states,  # 所有token都作为"文本"处理
                text_value_states=value_states,
                layer_idx=self.layer_idx,
                cache_kwargs=cache_kwargs
            )
        else:
            final_key_states = key_states
            final_value_states = value_states

        # 重复到所有头
        final_key_states = repeat_kv(final_key_states, self.num_key_value_groups)
        final_value_states = repeat_kv(final_value_states, self.num_key_value_groups)

        # 计算attention
        effective_scale_factor = self.softmax_temperature.to(q_pe.dtype) / math.sqrt(self.head_dim)
        
        # 分别计算pe和nope的attention权重
        attn_weights_pe = torch.matmul(q_pe, final_key_states[:, :, :, mask].transpose(2, 3)) * effective_scale_factor
        attn_weights_nope = torch.matmul(q_nope, final_key_states[:, :, :, ~mask].transpose(2, 3)) * effective_scale_factor
        attn_weights = attn_weights_pe + attn_weights_nope

        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:
            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()}"
                )
            attn_weights = attn_weights + attention_mask

        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(final_value_states.dtype)
        attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
        attn_output = torch.matmul(attn_weights, final_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, self.num_heads * self.head_dim)
        attn_output = self.o_proj(attn_output)

        if not output_attentions:
            attn_weights = None

        return attn_output, attn_weights, past_key_value

    def _forward_compress_all(
        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,
        vision_text_mask: Optional[torch.Tensor] = None,
        **kwargs,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        """统一压缩处理，所有token都使用MLA"""
        bsz, q_len, _ = hidden_states.size()

        # 计算query states
        query_states = self.q_proj(hidden_states)
        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)

        # 分离query的pe和nope部分
        mask = torch.zeros(self.head_dim, dtype=torch.bool, device=hidden_states.device)
        mask[::self.interval] = True 
        q_pe = query_states[..., mask]
        q_nope = query_states[..., ~mask]

        # ========== 统一压缩计算（所有token都使用MLA） ==========
        compressed_kv = self.kv_a_proj_nope(hidden_states)  # [bsz, q_len, num_kv_heads*lora_rank]
        k_pe = self.k_proj_pe(hidden_states).view(bsz, q_len, self.num_key_value_heads, (self.head_dim//self.interval)).transpose(1, 2)
        
        # 为缓存准备数据
        k_pe_for_cache = k_pe
        compressed_kv_for_cache = compressed_kv.view(bsz, q_len, self.num_key_value_heads, self.kv_lora_rank).transpose(1, 2)

        # ========== 处理缓存和位置编码 ==========
        kv_seq_len = q_len
        if past_key_value is not None:
            if self.layer_idx is None:
                raise ValueError("layer_idx is required when using cache")
            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)

        # 应用旋转位置编码
        cos, sin = self.rotary_emb(k_pe_for_cache, position_ids)
        q_pe, k_pe_for_cache = apply_rotary_pos_emb2(q_pe, k_pe_for_cache, cos, sin, position_ids)

        # 更新缓存 - 统一使用MixedDynamicCache接口，用None占位不需要的部分
        if past_key_value is not None and use_cache:
            cache_kwargs = {"sin": sin, "cos": cos}
            # compress_all模式：所有token都用压缩缓存，文本部分用None占位
            cached_k_pe, cached_compressed_kv, _, _ = past_key_value.update(
                vision_k_pe=k_pe_for_cache,  # 所有token都作为"视觉"处理（使用压缩）
                vision_compressed_kv=compressed_kv_for_cache,
                text_key_states=None,  # 统一压缩模式不需要
                text_value_states=None,  # 统一压缩模式不需要
                layer_idx=self.layer_idx,
                cache_kwargs=cache_kwargs
            )
        else:
            # 没有缓存，直接使用当前计算的结果
            cached_k_pe = k_pe_for_cache
            cached_compressed_kv = compressed_kv_for_cache

        # ========== 恢复原始维度 ==========
        seq_len = cached_k_pe.shape[-2]
        compressed_kv_reshaped = cached_compressed_kv.transpose(1, 2).reshape(bsz, seq_len, self.num_key_value_heads * self.kv_lora_rank)
        k_nope = self.k_b_proj_nope(compressed_kv_reshaped).view(bsz, seq_len, self.num_heads, self.head_dim - (self.head_dim//self.interval)).transpose(1, 2)
        value_states = self.v_b_proj(compressed_kv_reshaped).view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2)
        
        # 重复k_pe到所有头（只有k_pe需要repeat）
        k_pe_repeated = repeat_kv(cached_k_pe, self.num_key_value_groups)
        
        # k_nope和value_states已经是所有头的维度，不需要repeat
        k_nope = k_nope * self.k_nope_scale_factor.to(k_nope.dtype)
        
        # 合并key states
        final_key_states = torch.zeros(bsz, self.num_heads, seq_len, self.head_dim, device=hidden_states.device, dtype=hidden_states.dtype)
        final_key_states[:, :, :, mask] = k_pe_repeated
        final_key_states[:, :, :, ~mask] = k_nope
        final_value_states = value_states

        # ========== 计算attention ==========
        effective_scale_factor = self.softmax_temperature.to(q_pe.dtype) / math.sqrt(self.head_dim)
        
        # 分别计算pe和nope的attention权重
        attn_weights_pe = torch.matmul(q_pe, final_key_states[:, :, :, mask].transpose(2, 3)) * effective_scale_factor
        attn_weights_nope = torch.matmul(q_nope, final_key_states[:, :, :, ~mask].transpose(2, 3)) * effective_scale_factor
        attn_weights = attn_weights_pe + attn_weights_nope

        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:
            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()}"
                )
            attn_weights = attn_weights + attention_mask

        # 应用softmax和dropout
        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(final_value_states.dtype)
        attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
        attn_output = torch.matmul(attn_weights, final_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, self.num_heads * self.head_dim)
        attn_output = self.o_proj(attn_output)

        if not output_attentions:
            attn_weights = None

        return attn_output, attn_weights, past_key_value


    def _forward_mixed(
        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,
        vision_text_mask: Optional[torch.Tensor] = None,
        **kwargs,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        """
        混合处理模式的分发函数：根据是否有缓存和视觉token来决定使用prefill还是decode
        判断逻辑：
        1. 如果没有缓存(past_key_value is None)，说明是第一次forward，必然是prefill阶段
        2. 如果有缓存但当前输入包含视觉token，也是prefill阶段（虽然这种情况在实际使用中不太常见）
        3. 如果有缓存且当前输入只有文本token，则是decode阶段
        """
        
        # 检查当前输入是否包含视觉token
        has_vision_tokens = (vision_text_mask is not None and vision_text_mask.any())
        
        # 判断是prefill还是decode阶段
        if past_key_value is None or has_vision_tokens:
            # print("-------------------@prefilling------------------------")
            # prefill阶段：第一次forward或包含视觉token
            return self._forward_mixed_prefill(
                hidden_states=hidden_states,
                attention_mask=attention_mask,
                position_ids=position_ids,
                past_key_value=past_key_value,
                output_attentions=output_attentions,
                use_cache=use_cache,
                vision_text_mask=vision_text_mask,
                **kwargs
            )
        else:
            # decode阶段：有缓存且当前输入只有文本token
            return self._forward_mixed_decode(
                hidden_states=hidden_states,
                attention_mask=attention_mask,
                position_ids=position_ids,
                past_key_value=past_key_value,
                output_attentions=output_attentions,
                use_cache=use_cache,
                vision_text_mask=vision_text_mask,
                **kwargs
            )

    def _forward_mixed_prefill(
        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,
        vision_text_mask: Optional[torch.Tensor] = None,
        **kwargs,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        """
        混合模式的prefill阶段处理：
        - 处理视觉token（使用压缩投影）和文本token（使用完整投影）
        - 分别存储到不同的缓存中
        """
        bsz, q_len, _ = hidden_states.size()
        self.prefill_seq_len = q_len

        # ========== 1. 统一计算query states ==========
        query_states = self.q_proj(hidden_states)
        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
        
        mask = torch.zeros(self.head_dim, dtype=torch.bool, device=hidden_states.device)
        mask[::self.interval] = True 
        q_pe = query_states[..., mask]
        q_nope = query_states[..., ~mask]
        # ========== 2. 分离视觉和文本token ==========
        assert vision_text_mask is not None, "vision_text_mask is required in mixed mode"
        
        vision_mask = vision_text_mask  # [bsz, q_len], True表示视觉token
        text_mask = ~vision_text_mask   # [bsz, q_len], True表示文本token
        

        # 获取视觉token的索引和数据
        vision_indices = vision_text_mask.nonzero(as_tuple=True)  # (batch_idx, seq_idx)
        vision_tokens = hidden_states[vision_indices]  # [num_vision_tokens, hidden_size]
        vision_batch_indices = vision_indices[0]       # 每个视觉token属于哪个batch
        vision_seq_indices = vision_indices[1]         # 每个视觉token在序列中的位置
        num_vision_tokens = vision_tokens.shape[0]
        
        # 获取文本token的索引和数据
        text_indices = (~vision_text_mask).nonzero(as_tuple=True)
        text_tokens = hidden_states[text_indices]      # [num_text_tokens, hidden_size]
        text_batch_indices = text_indices[0]           # 每个文本token属于哪个batch
        text_seq_indices = text_indices[1]             # 每个文本token在序列中的位置
        num_text_tokens = text_tokens.shape[0]
        # 存储到当前层的属性中
        self.prefill_vision_indices = vision_indices
        self.prefill_text_indices = text_indices
        # ========== 3. 分别处理视觉和文本token ==========
        # 视觉token：使用压缩投影
        vision_compressed_kv = self.kv_a_proj_nope(vision_tokens).view(num_vision_tokens, self.num_key_value_heads, self.kv_lora_rank).transpose(0,1)
        vision_k_pe = self.k_proj_pe(vision_tokens).view(num_vision_tokens, self.num_key_value_heads,self.head_dim//self.interval).transpose(0,1)  
        
        # 文本token：使用原始投影
        text_key_states = self.k_proj(text_tokens).view(num_text_tokens, self.num_key_value_heads, self.head_dim).transpose(0,1)  
        text_value_states = self.v_proj(text_tokens).view(num_text_tokens, self.num_key_value_heads, self.head_dim).transpose(0,1)  

        # ========== 4. 处理position encoding ==========
        kv_seq_len = q_len
        if past_key_value is not None:
            if self.layer_idx is None:
                raise ValueError("layer_idx is required when using cache")
            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)

        cos, sin = self.rotary_emb(vision_k_pe, position_ids)
        text_key_pe = text_key_states[..., mask] 
        q_pe, vision_k_pe, text_key_pe = apply_rotary_pos_emb2_separate(q_pe, vision_k_pe, text_key_pe, cos, sin,vision_indices,text_indices)
        text_key_states[..., mask] = text_key_pe #text_key_states (num_key_value_heads,num_text_tokens,  self.head_dim)

        # ========== 5. 更新缓存（分离存储） ==========
        if past_key_value is not None and use_cache:
            cache_kwargs = {"sin": sin, "cos": cos}
            vision_k_pe, vision_compressed_kv, text_key_states, text_value_states = past_key_value.update(
                vision_k_pe=vision_k_pe,
                vision_compressed_kv=vision_compressed_kv,
                text_key_states=text_key_states,
                text_value_states=text_value_states,
                layer_idx=self.layer_idx,
                cache_kwargs=cache_kwargs
            )#(num_key_value_heads,num_text_tokens,  self.head_dim)，可以正常在MixedDynamicCache中工作

        # ========== 6. 根据当前序列的mask按需组装key/value ==========
        final_key_pe = torch.zeros(bsz, self.num_heads, kv_seq_len, self.head_dim//self.interval, 
                                device=hidden_states.device, dtype=hidden_states.dtype)
        final_key_nope = torch.zeros(bsz, self.num_heads, kv_seq_len, self.head_dim - self.head_dim//self.interval,
                                    device=hidden_states.device, dtype=hidden_states.dtype)
        final_value = torch.zeros(bsz, self.num_heads, kv_seq_len, self.head_dim,
                                device=hidden_states.device, dtype=hidden_states.dtype)
        
        # 组装视觉token的K/V（从压缩格式恢复）
        if num_vision_tokens > 0 and vision_k_pe is not None:
            # 从压缩状态恢复nope和value
            # vision_compressed_kv: [num_key_value_heads, num_vision_tokens, lora_rank]
            # 需要重新reshape为 [num_vision_tokens, num_key_value_heads * lora_rank]

            cache_num_vision_tokens = vision_compressed_kv.shape[1]
            vision_compressed_reshaped = vision_compressed_kv.transpose(0, 1).reshape(
                cache_num_vision_tokens, self.num_key_value_heads * self.kv_lora_rank)  # [num_vision_tokens, num_key_value_heads * lora_rank]
            
            # 恢复k_nope和value
            vision_k_nope = self.k_b_proj_nope(vision_compressed_reshaped).view(
                cache_num_vision_tokens, self.num_heads, self.head_dim - self.head_dim // self.interval)
            vision_value = self.v_b_proj(vision_compressed_reshaped).view(
                cache_num_vision_tokens, self.num_heads, self.head_dim)
            vision_k_nope = vision_k_nope * self.k_nope_scale_factor.to(vision_k_nope.dtype)
            
            # 重复k_pe到所有头: [num_key_value_heads, num_vision_tokens, pe_dim] -> [num_heads, num_vision_tokens, pe_dim]
            vision_k_pe_repeated = repeat_kv(vision_k_pe.unsqueeze(0), self.num_key_value_groups).squeeze(0)

            # 使用存储的索引填充到final tensors
            final_key_pe[self.prefill_vision_indices[0], :, self.prefill_vision_indices[1], :] = vision_k_pe_repeated.transpose(0, 1)
            final_key_nope[self.prefill_vision_indices[0], :, self.prefill_vision_indices[1], :] = vision_k_nope
            final_value[self.prefill_vision_indices[0], :, self.prefill_vision_indices[1], :] = vision_value

        # 组装文本token的K/V（直接使用完整格式）
        if num_text_tokens > 0 and text_key_states is not None:
            # text_key_states: [num_key_value_heads, num_text_tokens, head_dim]
            # text_value_states: [num_key_value_heads, num_text_tokens, head_dim]
            
            # 重复k/v到所有头
            text_key_repeated = repeat_kv(text_key_states.unsqueeze(0), self.num_key_value_groups).squeeze(0)
            text_value_repeated = repeat_kv(text_value_states.unsqueeze(0), self.num_key_value_groups).squeeze(0)
            
            # 分离pe和nope
            mask = torch.zeros(self.head_dim, dtype=torch.bool, device=hidden_states.device)
            mask[::self.interval] = True
            text_key_pe_repeated = text_key_repeated[..., mask]
            text_key_nope_repeated = text_key_repeated[..., ~mask]

            # 使用存储的索引填充到final tensors
            final_key_pe[self.prefill_text_indices[0], :, self.prefill_text_indices[1], :] = text_key_pe_repeated.transpose(0, 1)
            final_key_nope[self.prefill_text_indices[0], :, self.prefill_text_indices[1], :] = text_key_nope_repeated.transpose(0, 1)
            final_value[self.prefill_text_indices[0], :, self.prefill_text_indices[1], :] = text_value_repeated.transpose(0, 1)


        # ========== 7. 计算attention ==========
        effective_scale_factor = self.softmax_temperature.to(q_pe.dtype) / math.sqrt(self.head_dim)
        
        attn_weights_pe = torch.matmul(q_pe, final_key_pe.transpose(2, 3)) * effective_scale_factor
        attn_weights_nope = torch.matmul(q_nope, final_key_nope.transpose(2, 3)) * effective_scale_factor
        attn_weights = attn_weights_pe + attn_weights_nope

        if attention_mask is not None:
            attn_weights = attn_weights + attention_mask

        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(final_value.dtype)
        attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
        attn_output = torch.matmul(attn_weights, final_value)

        attn_output = attn_output.transpose(1, 2).contiguous()
        attn_output = attn_output.reshape(bsz, q_len, self.num_heads * self.head_dim)
        attn_output = self.o_proj(attn_output)

        if not output_attentions:
            attn_weights = None

        return attn_output, attn_weights, past_key_value


    def _forward_mixed_decode(
        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,
        vision_text_mask: Optional[torch.Tensor] = None,
        **kwargs,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        """
        混合模式的decode阶段处理：
        - 当前输入只有文本token（使用完整投影）
        - 从缓存中读取历史的视觉token（压缩格式）和文本token（完整格式）
        - 使用存储的索引正确恢复历史token位置，新文本token追加到末尾
        """
        bsz, q_len, _ = hidden_states.size()
        assert past_key_value is not None, "past_key_value is required in decode stage"
        assert q_len == 1, "Decode stage should have q_len=1"  # decode阶段通常一次只处理一个token
        
        # ========== 1. 统一计算query states ==========
        query_states = self.q_proj(hidden_states)
        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
        
        mask = torch.zeros(self.head_dim, dtype=torch.bool, device=hidden_states.device)
        mask[::self.interval] = True 
        q_pe = query_states[..., mask]
        q_nope = query_states[..., ~mask]

        # ========== 2. 处理当前文本token（使用完整投影） ==========
        # 当前输入全是文本token，使用原始投影
        current_text_key = self.k_proj(hidden_states).view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
        current_text_value = self.v_proj(hidden_states).view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)

        # ========== 3. 处理position encoding ==========
        kv_seq_len = q_len
        if past_key_value is not None:
            if self.layer_idx is None:
                raise ValueError("layer_idx is required when using cache")
            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)

        cos, sin = self.rotary_emb(current_text_key, position_ids)
        q_pe = apply_rotary_pos_emb2_single(q_pe, cos, sin)
        
        # 对当前文本token的key应用位置编码
        current_text_key_pe = current_text_key[..., mask]
        current_text_key_pe = apply_rotary_pos_emb2_single(current_text_key_pe, cos, sin)
        current_text_key[..., mask] = current_text_key_pe

        # ========== 4. 更新缓存（按batch×seqlen格式存储新文本） ==========
        # 直接transpose和view，更简洁的维度变换
        current_text_key_for_cache = current_text_key.transpose(0, 1).contiguous().view(self.num_key_value_heads, bsz * q_len, self.head_dim)
        current_text_value_for_cache = current_text_value.transpose(0, 1).contiguous().view(self.num_key_value_heads, bsz * q_len, self.head_dim)

        if use_cache:
            cache_kwargs = {"sin": sin, "cos": cos}
            cached_vision_k_pe, cached_vision_compressed_kv, cached_text_key, cached_text_value = past_key_value.update(
                vision_k_pe=None,  # decode阶段没有新的视觉token
                vision_compressed_kv=None,
                text_key_states=current_text_key_for_cache,  # 按统一格式添加当前文本token
                text_value_states=current_text_value_for_cache,
                layer_idx=self.layer_idx,
                cache_kwargs=cache_kwargs
            )#(num_key_value_heads , num_tokens , head_dim)
        # ========== 5. 使用存储的索引组装完整的key/value ==========
        vision_token_len = cached_vision_k_pe.shape[-2]  # prefill阶段的vision tokens数量
        prefill_text_token_len = len(self.prefill_text_indices[0])  # prefill阶段的text tokens数量
        total_cached_text_len = cached_text_key.shape[-2]  # 缓存中所有text tokens数量
        decode_text_token_len = total_cached_text_len - prefill_text_token_len  # decode阶段累积的text tokens数量
        prefill_total_token_len = vision_token_len + prefill_text_token_len#prefill阶段的tokens数量
        decode_text_seq_len = decode_text_token_len//bsz #decode阶段的序列长度，不是token数量
        total_seq_len = self.prefill_seq_len + decode_text_seq_len #总共的序列长度，不是token数量
        # 初始化最终的key/value tensors
        final_key_pe = torch.zeros(bsz, self.num_heads, total_seq_len, self.head_dim//self.interval, device=hidden_states.device, dtype=hidden_states.dtype)
        final_key_nope = torch.zeros(bsz, self.num_heads, total_seq_len, self.head_dim - self.head_dim//self.interval,device=hidden_states.device, dtype=hidden_states.dtype)
        final_value = torch.zeros(bsz, self.num_heads, total_seq_len, self.head_dim,device=hidden_states.device, dtype=hidden_states.dtype)

        # 1. 恢复历史vision tokens（使用prefill索引）
        if vision_token_len > 0:
            vision_compressed_reshaped = cached_vision_compressed_kv.transpose(0, 1).reshape(vision_token_len, self.num_key_value_heads * self.kv_lora_rank)
            
            vision_k_nope = self.k_b_proj_nope(vision_compressed_reshaped).view(vision_token_len, self.num_heads, self.head_dim - self.head_dim // self.interval)
            vision_value = self.v_b_proj(vision_compressed_reshaped).view(vision_token_len, self.num_heads, self.head_dim)
            vision_k_nope = vision_k_nope * self.k_nope_scale_factor.to(vision_k_nope.dtype)
            
            vision_k_pe_repeated = repeat_kv(cached_vision_k_pe.unsqueeze(0), self.num_key_value_groups).squeeze(0)#[num_heads, vision_token_len, pe_dim]

            # 使用prefill的vision索引填充
            final_key_pe[self.prefill_vision_indices[0], :, self.prefill_vision_indices[1], :] = vision_k_pe_repeated.transpose(0, 1)
            final_key_nope[self.prefill_vision_indices[0], :, self.prefill_vision_indices[1], :] = vision_k_nope
            final_value[self.prefill_vision_indices[0], :, self.prefill_vision_indices[1], :] = vision_value

        # 2. 恢复历史prefill text tokens（使用prefill索引）
        if prefill_text_token_len > 0:
            # 取缓存中前面的prefill text tokens
            prefill_text_key = cached_text_key[:, :prefill_text_token_len, :]  # [num_key_value_heads, prefill_text_len, head_dim]
            prefill_text_value = cached_text_value[:, :prefill_text_token_len, :]
            
            # 重复到所有头 - 现在都是 [num_heads, prefill_text_len, head_dim]
            prefill_text_key_repeated = repeat_kv(prefill_text_key.unsqueeze(0), self.num_key_value_groups).squeeze(0)
            prefill_text_value_repeated = repeat_kv(prefill_text_value.unsqueeze(0), self.num_key_value_groups).squeeze(0)
            
            # 分离pe和nope
            prefill_text_key_pe = prefill_text_key_repeated[..., mask]
            prefill_text_key_nope = prefill_text_key_repeated[..., ~mask]

            # 使用prefill的text索引填充 
            final_key_pe[self.prefill_text_indices[0], :, self.prefill_text_indices[1], :] = prefill_text_key_pe.transpose(0, 1)
            final_key_nope[self.prefill_text_indices[0], :, self.prefill_text_indices[1], :] = prefill_text_key_nope.transpose(0, 1)
            final_value[self.prefill_text_indices[0], :, self.prefill_text_indices[1], :] = prefill_text_value_repeated.transpose(0, 1)

        # 3. 添加decode阶段的text tokens（按顺序追加在后面）
        if decode_text_token_len > 0:
            # 取缓存中后面的decode text tokens
            decode_text_key = cached_text_key[:, prefill_text_token_len:, :]  # [num_key_value_heads, decode_text_len, head_dim]
            decode_text_value = cached_text_value[:, prefill_text_token_len:, :]
            
            # 重复到所有头并reshape为batch格式
            decode_text_key_repeated = repeat_kv(decode_text_key.unsqueeze(0), self.num_key_value_groups).squeeze(0)  # [num_heads, decode_text_len, head_dim]
            decode_text_value_repeated = repeat_kv(decode_text_value.unsqueeze(0), self.num_key_value_groups).squeeze(0)
            # 将decode tokens reshape回batch格式: [num_heads, bsz*decode_text_seq_len, head_dim] -> [bsz, num_heads, decode_text_seq_len, head_dim]
            decode_text_key_batch = decode_text_key_repeated.view(self.num_heads, bsz, decode_text_seq_len, self.head_dim).transpose(0, 1)
            decode_text_value_batch = decode_text_value_repeated.view(self.num_heads, bsz, decode_text_seq_len, self.head_dim).transpose(0, 1)
            
            # 分离pe和nope
            decode_text_key_pe = decode_text_key_batch[..., mask]
            decode_text_key_nope = decode_text_key_batch[..., ~mask]

            # 追加在序列末尾
            final_key_pe[:, :, self.prefill_seq_len :, :] = decode_text_key_pe
            final_key_nope[:, :, self.prefill_seq_len :, :] = decode_text_key_nope
            final_value[:, :, self.prefill_seq_len :, :] = decode_text_value_batch


        # ========== 6. 计算attention ==========
        effective_scale_factor = self.softmax_temperature.to(q_pe.dtype) / math.sqrt(self.head_dim)
        
        attn_weights_pe = torch.matmul(q_pe, final_key_pe.transpose(2, 3)) * effective_scale_factor
        attn_weights_nope = torch.matmul(q_nope, final_key_nope.transpose(2, 3)) * effective_scale_factor
        attn_weights = attn_weights_pe + attn_weights_nope

        if attention_mask is not None:
            if attention_mask.size() != (bsz, 1, q_len, total_seq_len):
                raise ValueError(
                    f"Attention mask should be of size {(bsz, 1, q_len, total_seq_len)}, but is {attention_mask.size()}"
                )
            attn_weights = attn_weights + attention_mask

        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(final_value.dtype)
        attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
        attn_output = torch.matmul(attn_weights, final_value)

        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, self.num_heads * self.head_dim)
        attn_output = self.o_proj(attn_output)

        if not output_attentions:
            attn_weights = None

        return attn_output, attn_weights, past_key_value



class Qwen2FlashAttention2(Qwen2Attention):
    """
    Qwen2 flash attention module, following Qwen2 attention module. This module inherits from `Qwen2Attention`
    as the weights of the module stays untouched. The only required change would be on the forward pass
    where it needs to correctly call the public API of flash attention and deal with padding tokens
    in case the input contains any of them. Additionally, for sliding window attention, we apply SWA only to the bottom
    config.max_window_layers layers.
    """

    # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2.__init__
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)

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

    def forward(
        self,
        hidden_states: torch.Tensor,
        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,
        vision_text_mask: Optional[torch.Tensor] = 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.`"
            )

        # 根据处理模式选择不同的forward逻辑
        if self.processing_mode == 'no_compress':
            print('*********************************no_compress*********************************')
            return self._forward_no_compress(
                hidden_states, attention_mask, position_ids, past_key_value,
                output_attentions, use_cache, vision_text_mask, **kwargs
            )
        elif self.processing_mode == 'compress_all':
            print('*********************************compress_all*********************************')
            # print('flash_compress_all')
            return self._forward_compress_all(
                hidden_states, attention_mask, position_ids, past_key_value,
                output_attentions, use_cache, vision_text_mask, **kwargs
            )
        elif self.processing_mode == 'mixed':
            print('*********************************   mixed   *********************************')
            return self._forward_mixed(
                hidden_states, attention_mask, position_ids, past_key_value,
                output_attentions, use_cache, vision_text_mask, **kwargs
            )
        else:
            raise ValueError(f"Unsupported processing mode: {self.processing_mode}")

    def _forward_no_compress(
        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,
        vision_text_mask: Optional[torch.Tensor] = None,
        **kwargs,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        """不压缩处理，但使用部分RoPE形式"""
        bsz, q_len, _ = hidden_states.size()

        # 计算query states
        query_states = self.q_proj(hidden_states)
        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)

        # 分离query的pe和nope部分
        mask = torch.zeros(self.head_dim, dtype=torch.bool, device=hidden_states.device)
        mask[::self.interval] = True 
        q_pe = query_states[..., mask]

        # 计算key和value states（完整的，不压缩）
        key_states = self.k_proj(hidden_states)
        value_states = self.v_proj(hidden_states)
        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)

        # 从完整的key中分离pe部分（类似query的处理方式）
        k_pe = key_states[..., mask]  # 直接裁剪，不使用k_proj_pe
        k_nope = key_states[..., ~mask]

        # 处理缓存
        kv_seq_len = q_len
        if past_key_value is not None:
            if self.layer_idx is None:
                raise ValueError("layer_idx is required when using cache")
            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)

        # 应用旋转位置编码
        cos, sin = self.rotary_emb(k_pe, position_ids)
        q_pe, k_pe = apply_rotary_pos_emb2(q_pe, k_pe, cos, sin, position_ids)

        # 重新组合key states（pe部分使用RoPE后的结果，nope部分保持原样）
        key_states[..., mask] = k_pe
        key_states[..., ~mask] = k_nope

        # 更新缓存 - 统一使用MixedDynamicCache接口，用None占位不需要的部分
        if past_key_value is not None and use_cache:
            cache_kwargs = {"sin": sin, "cos": cos}
            # no_compress模式：文本用完整缓存，视觉部分用None占位
            _, _, final_key_states, final_value_states = past_key_value.update(
                vision_k_pe=None,  # 不压缩模式不需要
                vision_compressed_kv=None,  # 不压缩模式不需要
                text_key_states=key_states,  # 所有token都作为"文本"处理
                text_value_states=value_states,
                layer_idx=self.layer_idx,
                cache_kwargs=cache_kwargs
            )
        else:
            final_key_states = key_states
            final_value_states = value_states

        # 重复到所有头
        final_key_states = repeat_kv(final_key_states, self.num_key_value_groups)
        final_value_states = repeat_kv(final_value_states, self.num_key_value_groups)

        # ========== Flash Attention计算 ==========
        # 重新组合query states
        query_states = q_pe.new_empty(bsz, self.num_heads, q_len, self.head_dim)
        query_states[:, :, :, mask] = q_pe
        
        # 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 = final_key_states.transpose(1, 2)
        value_states = final_value_states.transpose(1, 2)

        dropout_rate = 0.0 if not self.training else self.attention_dropout

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

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

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

        current_softmax_scale = self.head_dim ** (-0.5)
        effective_softmax_scale = current_softmax_scale * self.softmax_temperature.to(query_states.dtype)

        attn_output = self._flash_attention_forward(
            query_states,
            key_states,
            value_states,
            attention_mask,
            q_len,
            dropout=dropout_rate,
            softmax_scale=effective_softmax_scale.item(),
        )
        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
        attn_output = self.o_proj(attn_output)

        if not output_attentions:
            attn_weights = None

        return attn_output, attn_weights, past_key_value

    def _forward_compress_all(
        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,
        vision_text_mask: Optional[torch.Tensor] = None,
        **kwargs,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        """统一压缩处理，所有token都使用MLA"""
        bsz, q_len, _ = hidden_states.size()

        # 计算query states
        query_states = self.q_proj(hidden_states)
        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)

        # 分离query的pe和nope部分
        mask = torch.zeros(self.head_dim, dtype=torch.bool, device=hidden_states.device)
        mask[::self.interval] = True 
        q_pe = query_states[..., mask]

        # ========== 统一压缩计算（所有token都使用MLA） ==========
        compressed_kv = self.kv_a_proj_nope(hidden_states)  # [bsz, q_len, num_kv_heads*lora_rank]
        k_pe = self.k_proj_pe(hidden_states).view(bsz, q_len, self.num_key_value_heads, (self.head_dim//self.interval)).transpose(1, 2)
        
        # 为缓存准备数据
        k_pe_for_cache = k_pe
        compressed_kv_for_cache = compressed_kv.view(bsz, q_len, self.num_key_value_heads, self.kv_lora_rank).transpose(1, 2)

        # ========== 处理缓存和位置编码 ==========
        kv_seq_len = q_len
        if past_key_value is not None:
            if self.layer_idx is None:
                raise ValueError("layer_idx is required when using cache")
            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)

        # 应用旋转位置编码
        cos, sin = self.rotary_emb(k_pe_for_cache, position_ids)
        q_pe, k_pe_for_cache = apply_rotary_pos_emb2(q_pe, k_pe_for_cache, cos, sin, position_ids)

        # 更新缓存 - 统一使用MixedDynamicCache接口，用None占位不需要的部分
        if past_key_value is not None and use_cache:
            cache_kwargs = {"sin": sin, "cos": cos}
            # compress_all模式：所有token都用压缩缓存，文本部分用None占位
            cached_k_pe, cached_compressed_kv, _, _ = past_key_value.update(
                vision_k_pe=k_pe_for_cache,  # 所有token都作为"视觉"处理（使用压缩）
                vision_compressed_kv=compressed_kv_for_cache,
                text_key_states=None,  # 统一压缩模式不需要
                text_value_states=None,  # 统一压缩模式不需要
                layer_idx=self.layer_idx,
                cache_kwargs=cache_kwargs
            )
        else:
            # 没有缓存，直接使用当前计算的结果
            cached_k_pe = k_pe_for_cache
            cached_compressed_kv = compressed_kv_for_cache

        # ========== 恢复原始维度 ==========
        seq_len = cached_k_pe.shape[-2]
        compressed_kv_reshaped = cached_compressed_kv.transpose(1, 2).reshape(bsz, seq_len, self.num_key_value_heads * self.kv_lora_rank)
        k_nope = self.k_b_proj_nope(compressed_kv_reshaped).view(bsz, seq_len, self.num_heads, self.head_dim - (self.head_dim//self.interval)).transpose(1, 2)
        value_states = self.v_b_proj(compressed_kv_reshaped).view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2)
        
        # 重复k_pe到所有头（只有k_pe需要repeat）
        k_pe_repeated = repeat_kv(cached_k_pe, self.num_key_value_groups)
        
        # k_nope和value_states已经是所有头的维度，不需要repeat
        k_nope = k_nope * self.k_nope_scale_factor.to(k_nope.dtype)
        
        # 合并key states
        final_key_states = torch.zeros(bsz, self.num_heads, seq_len, self.head_dim, device=hidden_states.device, dtype=hidden_states.dtype)
        final_key_states[:, :, :, mask] = k_pe_repeated
        final_key_states[:, :, :, ~mask] = k_nope
        final_value_states = value_states

        # ========== Flash Attention计算 ==========
        # 重新组合query states
        query_states = q_pe.new_empty(bsz, self.num_heads, q_len, self.head_dim)
        query_states[:, :, :, mask] = q_pe
        
        # 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).contiguous()
        key_states = final_key_states.transpose(1, 2).contiguous()
        value_states = final_value_states.transpose(1, 2).contiguous()

        dropout_rate = 0.0 if not self.training else self.attention_dropout

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

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

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

        current_softmax_scale = self.head_dim ** (-0.5)
        effective_softmax_scale = current_softmax_scale * self.softmax_temperature.to(query_states.dtype)

        attn_output = self._flash_attention_forward(
            query_states,
            key_states,
            value_states,
            attention_mask,
            q_len,
            dropout=dropout_rate,
            softmax_scale=effective_softmax_scale.item(),
        )
        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
        attn_output = self.o_proj(attn_output)

        if not output_attentions:
            attn_weights = None

        return attn_output, attn_weights, past_key_value


    def _forward_mixed(
        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,
        vision_text_mask: Optional[torch.Tensor] = None,
        **kwargs,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        """
        混合处理模式的分发函数：根据是否有缓存和视觉token来决定使用prefill还是decode
        判断逻辑：
        1. 如果没有缓存(past_key_value is None)，说明是第一次forward，必然是prefill阶段
        2. 如果有缓存但当前输入包含视觉token，也是prefill阶段（虽然这种情况在实际使用中不太常见）
        3. 如果有缓存且当前输入只有文本token，则是decode阶段
        """
        
        # 检查当前输入是否包含视觉token
        has_vision_tokens = (vision_text_mask is not None and vision_text_mask.any())
        
        # 判断是prefill还是decode阶段
        if past_key_value is None or has_vision_tokens:
            # print("-------------------@prefilling------------------------")
            # prefill阶段：第一次forward或包含视觉token
            return self._forward_mixed_prefill(
                hidden_states=hidden_states,
                attention_mask=attention_mask,
                position_ids=position_ids,
                past_key_value=past_key_value,
                output_attentions=output_attentions,
                use_cache=use_cache,
                vision_text_mask=vision_text_mask,
                **kwargs
            )
        else:
            # decode阶段：有缓存且当前输入只有文本token
            return self._forward_mixed_decode(
                hidden_states=hidden_states,
                attention_mask=attention_mask,
                position_ids=position_ids,
                past_key_value=past_key_value,
                output_attentions=output_attentions,
                use_cache=use_cache,
                vision_text_mask=vision_text_mask,
                **kwargs
            )

    def _forward_mixed_prefill(
        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,
        vision_text_mask: Optional[torch.Tensor] = None,
        **kwargs,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        """
        混合模式的prefill阶段处理：
        - 处理视觉token（使用压缩投影）和文本token（使用完整投影）
        - 分别存储到不同的缓存中
        """
        bsz, q_len, _ = hidden_states.size()
        self.prefill_seq_len = q_len

        # ========== 1. 统一计算query states ==========
        query_states = self.q_proj(hidden_states)
        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
        
        mask = torch.zeros(self.head_dim, dtype=torch.bool, device=hidden_states.device)
        mask[::self.interval] = True 
        q_pe = query_states[..., mask]
        # ========== 2. 分离视觉和文本token ==========
        assert vision_text_mask is not None, "vision_text_mask is required in mixed mode"
        
        vision_mask = vision_text_mask  # [bsz, q_len], True表示视觉token
        text_mask = ~vision_text_mask   # [bsz, q_len], True表示文本token
        
        # 获取视觉token的索引和数据
        vision_indices = vision_text_mask.nonzero(as_tuple=True)  # (batch_idx, seq_idx)
        vision_tokens = hidden_states[vision_indices]  # [num_vision_tokens, hidden_size]
        vision_batch_indices = vision_indices[0]       # 每个视觉token属于哪个batch
        vision_seq_indices = vision_indices[1]         # 每个视觉token在序列中的位置
        num_vision_tokens = vision_tokens.shape[0]
        
        # 获取文本token的索引和数据
        text_indices = (~vision_text_mask).nonzero(as_tuple=True)
        text_tokens = hidden_states[text_indices]      # [num_text_tokens, hidden_size]
        text_batch_indices = text_indices[0]           # 每个文本token属于哪个batch
        text_seq_indices = text_indices[1]             # 每个文本token在序列中的位置
        num_text_tokens = text_tokens.shape[0]
        # 存储到当前层的属性中
        self.prefill_vision_indices = vision_indices
        self.prefill_text_indices = text_indices
        # ========== 3. 分别处理视觉和文本token ==========
        # 视觉token：使用压缩投影
        vision_compressed_kv = self.kv_a_proj_nope(vision_tokens).view(num_vision_tokens, self.num_key_value_heads, self.kv_lora_rank).transpose(0,1)
        vision_k_pe = self.k_proj_pe(vision_tokens).view(num_vision_tokens, self.num_key_value_heads,self.head_dim//self.interval).transpose(0,1)  
        
        # 文本token：使用原始投影
        text_key_states = self.k_proj(text_tokens).view(num_text_tokens, self.num_key_value_heads, self.head_dim).transpose(0,1)  
        text_value_states = self.v_proj(text_tokens).view(num_text_tokens, self.num_key_value_heads, self.head_dim).transpose(0,1)  

        # ========== 4. 处理position encoding ==========
        kv_seq_len = q_len
        if past_key_value is not None:
            if self.layer_idx is None:
                raise ValueError("layer_idx is required when using cache")
            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)

        cos, sin = self.rotary_emb(vision_k_pe, position_ids)
        text_key_pe = text_key_states[..., mask]
        q_pe, vision_k_pe, text_key_pe = apply_rotary_pos_emb2_separate(q_pe, vision_k_pe, text_key_pe, cos, sin,vision_indices,text_indices)
        text_key_states[..., mask] = text_key_pe

        # ========== 5. 更新缓存（分离存储） ==========
        if past_key_value is not None and use_cache:
            cache_kwargs = {"sin": sin, "cos": cos}
            vision_k_pe, vision_compressed_kv, text_key_states, text_value_states = past_key_value.update(
                vision_k_pe=vision_k_pe.contiguous(),
                vision_compressed_kv=vision_compressed_kv.contiguous(),
                text_key_states=text_key_states.contiguous(),
                text_value_states=text_value_states.contiguous(),
                layer_idx=self.layer_idx,
                cache_kwargs=cache_kwargs
            )#(num_key_value_heads,num_text_tokens,  self.head_dim)，可以正常在MixedDynamicCache中工作

        # ========== 6. 基于切片和cat的简洁重组方法 ==========
        # 处理视觉和文本token数据
        if num_vision_tokens > 0:
            cache_num_vision_tokens = vision_compressed_kv.shape[1]
            vision_compressed_reshaped = vision_compressed_kv.transpose(0, 1).reshape(
                cache_num_vision_tokens, self.num_key_value_heads * self.kv_lora_rank)
            
            vision_k_nope = self.k_b_proj_nope(vision_compressed_reshaped).view(
                cache_num_vision_tokens, self.num_heads, self.head_dim - self.head_dim // self.interval)
            vision_value = self.v_b_proj(vision_compressed_reshaped).view(
                cache_num_vision_tokens, self.num_heads, self.head_dim)
            vision_k_nope = vision_k_nope * self.k_nope_scale_factor.to(vision_k_nope.dtype)
            
            vision_k_pe_repeated = repeat_kv(vision_k_pe.unsqueeze(0), self.num_key_value_groups).squeeze(0)
            
            vision_key_full = torch.empty(cache_num_vision_tokens, self.num_heads, self.head_dim, device=hidden_states.device, dtype=hidden_states.dtype)
            vision_key_full[..., mask] = vision_k_pe_repeated.transpose(0, 1)
            vision_key_full[..., ~mask] = vision_k_nope

        if num_text_tokens > 0:
            text_key_repeated = repeat_kv(text_key_states.unsqueeze(0), self.num_key_value_groups).squeeze(0).transpose(0, 1)
            text_value_repeated = repeat_kv(text_value_states.unsqueeze(0), self.num_key_value_groups).squeeze(0).transpose(0, 1)

        # 按batch循环，用切片和cat拼装
        batch_keys = []
        batch_values = []

        for batch_idx in range(bsz):
            # 找到当前batch中视频token的起始和结束位置
            batch_vision_mask = self.prefill_vision_indices[0] == batch_idx

            vision_start = self.prefill_vision_indices[1][batch_vision_mask].min().item()
            vision_end = self.prefill_vision_indices[1][batch_vision_mask].max().item() + 1
            
            # 找到当前batch的文本token，分离视频前后的文本token
            batch_text_mask = self.prefill_text_indices[0] == batch_idx
            batch_text_positions = self.prefill_text_indices[1][batch_text_mask]
            pre_video_text_mask = batch_text_positions < vision_start
            post_video_text_mask = batch_text_positions >= vision_end
            
            # 构建当前batch的key和value序列
            seq_parts_key = []
            seq_parts_value = []
            
            # 视频前的文本token
            if pre_video_text_mask.any():
                pre_text_indices = batch_text_positions[pre_video_text_mask]
                pre_text_token_indices = torch.where(batch_text_mask)[0][pre_video_text_mask]
                seq_parts_key.append(text_key_repeated[pre_text_token_indices])
                seq_parts_value.append(text_value_repeated[pre_text_token_indices])
            
            # 视频token
            vision_token_indices = torch.where(batch_vision_mask)[0]
            seq_parts_key.append(vision_key_full[vision_token_indices])
            seq_parts_value.append(vision_value[vision_token_indices])
            
            # 视频后的文本token
            if post_video_text_mask.any():
                post_text_indices = batch_text_positions[post_video_text_mask]
                post_text_token_indices = torch.where(batch_text_mask)[0][post_video_text_mask]
                seq_parts_key.append(text_key_repeated[post_text_token_indices])
                seq_parts_value.append(text_value_repeated[post_text_token_indices])
            
            # cat拼接当前batch的完整序列
            batch_key = torch.cat(seq_parts_key, dim=0)  # [seq_len, num_heads, head_dim]
            batch_value = torch.cat(seq_parts_value, dim=0)
        
        batch_keys.append(batch_key.unsqueeze(0))  # [1, seq_len, num_heads, head_dim]
        batch_values.append(batch_value.unsqueeze(0))

        final_key_states = torch.cat(batch_keys, dim=0)  # [bsz, seq_len, num_heads, head_dim]
        final_value = torch.cat(batch_values, dim=0)
        # ========== 7. Flash Attention计算 ==========
        # 重新组合query states
        query_states[:, :, :, mask] = q_pe
        
        # 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).contiguous()
        # final_key_states和final_value已经是目标形状，无需transpose
        key_states = final_key_states.contiguous()
        value_states = final_value.contiguous()

        dropout_rate = 0.0 if not self.training else self.attention_dropout

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

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

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

        current_softmax_scale = self.head_dim ** (-0.5)
        effective_softmax_scale = current_softmax_scale * self.softmax_temperature.to(query_states.dtype)

        attn_output = self._flash_attention_forward(
            query_states,
            key_states,
            value_states,
            attention_mask,
            q_len,
            dropout=dropout_rate,
            softmax_scale=effective_softmax_scale.item(),
        )
        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
        attn_output = self.o_proj(attn_output)

        if not output_attentions:
            attn_weights = None

        return attn_output, attn_weights, past_key_value


    def _forward_mixed_decode(
        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,
        vision_text_mask: Optional[torch.Tensor] = None,
        **kwargs,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        """
        混合模式的decode阶段处理：
        - 当前输入只有文本token（使用完整投影）
        - 从缓存中读取历史的视觉token（压缩格式）和文本token（完整格式）
        - 使用存储的索引正确恢复历史token位置，新文本token追加到末尾
        """
        bsz, q_len, _ = hidden_states.size()
        assert past_key_value is not None, "past_key_value is required in decode stage"
        assert q_len == 1, "Decode stage should have q_len=1"  # decode阶段通常一次只处理一个token
        
        # ========== 1. 统一计算query states ==========
        query_states = self.q_proj(hidden_states)
        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
        
        mask = torch.zeros(self.head_dim, dtype=torch.bool, device=hidden_states.device)
        mask[::self.interval] = True 
        q_pe = query_states[..., mask]

        # ========== 2. 处理当前文本token（使用完整投影） ==========
        # 当前输入全是文本token，使用原始投影
        current_text_key = self.k_proj(hidden_states).view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
        current_text_value = self.v_proj(hidden_states).view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)

        # ========== 3. 处理position encoding ==========
        kv_seq_len = q_len
        if past_key_value is not None:
            if self.layer_idx is None:
                raise ValueError("layer_idx is required when using cache")
            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)

        cos, sin = self.rotary_emb(current_text_key, position_ids)
        q_pe = apply_rotary_pos_emb2_single(q_pe, cos, sin)
        
        # 对当前文本token的key应用位置编码
        current_text_key_pe = current_text_key[..., mask]
        current_text_key_pe = apply_rotary_pos_emb2_single(current_text_key_pe, cos, sin)
        current_text_key[..., mask] = current_text_key_pe

        # ========== 4. 更新缓存（按batch×seqlen格式存储新文本） ==========
        # 直接transpose和view，更简洁的维度变换
        current_text_key_for_cache = current_text_key.transpose(0, 1).contiguous().view(self.num_key_value_heads, bsz * q_len, self.head_dim)
        current_text_value_for_cache = current_text_value.transpose(0, 1).contiguous().view(self.num_key_value_heads, bsz * q_len, self.head_dim)

        if use_cache:
            cache_kwargs = {"sin": sin, "cos": cos}
            cached_vision_k_pe, cached_vision_compressed_kv, cached_text_key, cached_text_value = past_key_value.update(
                vision_k_pe=None,  # decode阶段没有新的视觉token
                vision_compressed_kv=None,
                text_key_states=current_text_key_for_cache.contiguous(),  # 按统一格式添加当前文本token
                text_value_states=current_text_value_for_cache.contiguous(),
                layer_idx=self.layer_idx,
                cache_kwargs=cache_kwargs
            )#(num_key_value_heads , num_tokens , head_dim)
        # ========== 5. 使用存储的索引组装完整的key/value ==========
        vision_token_len = cached_vision_k_pe.shape[-2]  # prefill阶段的vision tokens数量
        prefill_text_token_len = len(self.prefill_text_indices[0])  # prefill阶段的text tokens数量
        total_cached_text_len = cached_text_key.shape[-2]  # 缓存中所有text tokens数量
        decode_text_token_len = total_cached_text_len - prefill_text_token_len  # decode阶段累积的text tokens数量
        prefill_total_token_len = vision_token_len + prefill_text_token_len#prefill阶段的tokens数量
        decode_text_seq_len = decode_text_token_len//bsz #decode阶段的序列长度，不是token数量
        total_seq_len = self.prefill_seq_len + decode_text_seq_len #总共的序列长度，不是token数量
        # 准备历史vision tokens数据
        if vision_token_len > 0:
            vision_compressed_reshaped = cached_vision_compressed_kv.transpose(0, 1).reshape(vision_token_len, self.num_key_value_heads * self.kv_lora_rank)
            
            vision_k_nope = self.k_b_proj_nope(vision_compressed_reshaped).view(vision_token_len, self.num_heads, self.head_dim - self.head_dim // self.interval)
            vision_value = self.v_b_proj(vision_compressed_reshaped).view(vision_token_len, self.num_heads, self.head_dim)
            vision_k_nope = vision_k_nope * self.k_nope_scale_factor.to(vision_k_nope.dtype)
            
            vision_k_pe_repeated = repeat_kv(cached_vision_k_pe.unsqueeze(0), self.num_key_value_groups).squeeze(0)
            
            vision_key_full = torch.empty(vision_token_len, self.num_heads, self.head_dim, device=hidden_states.device, dtype=hidden_states.dtype)
            vision_key_full[..., mask] = vision_k_pe_repeated.transpose(0, 1)
            vision_key_full[..., ~mask] = vision_k_nope

        # 准备历史text tokens数据
        if prefill_text_token_len > 0:
            prefill_text_key = cached_text_key[:, :prefill_text_token_len, :]
            prefill_text_value = cached_text_value[:, :prefill_text_token_len, :]
            prefill_text_key_repeated = repeat_kv(prefill_text_key.unsqueeze(0), self.num_key_value_groups).squeeze(0).transpose(0, 1)
            prefill_text_value_repeated = repeat_kv(prefill_text_value.unsqueeze(0), self.num_key_value_groups).squeeze(0).transpose(0, 1)

        # 准备decode阶段的text tokens数据
        if decode_text_token_len > 0:
            decode_text_key = cached_text_key[:, prefill_text_token_len:, :]
            decode_text_value = cached_text_value[:, prefill_text_token_len:, :]
            decode_text_key_repeated = repeat_kv(decode_text_key.unsqueeze(0), self.num_key_value_groups).squeeze(0)
            decode_text_value_repeated = repeat_kv(decode_text_value.unsqueeze(0), self.num_key_value_groups).squeeze(0)

        # 按batch循环，用切片和cat拼装
        batch_keys = []
        batch_values = []

        for batch_idx in range(bsz):
            # 找到当前batch中视频token的位置
            batch_vision_mask = self.prefill_vision_indices[0] == batch_idx

            vision_start = self.prefill_vision_indices[1][batch_vision_mask].min().item()
            vision_end = self.prefill_vision_indices[1][batch_vision_mask].max().item() + 1
            vision_token_indices = torch.where(batch_vision_mask)[0]
            
            # 找到当前batch的prefill text token位置
            batch_text_mask = self.prefill_text_indices[0] == batch_idx
            batch_text_positions = self.prefill_text_indices[1][batch_text_mask]
            pre_video_text_mask = batch_text_positions < vision_start
            post_video_text_mask = batch_text_positions >= vision_end
            
            # 构建当前batch的完整序列
            seq_parts_key = []
            seq_parts_value = []
            
            # prefill阶段：视频前的文本token
            if pre_video_text_mask.any():
                pre_text_token_indices = torch.where(batch_text_mask)[0][pre_video_text_mask]
                seq_parts_key.append(prefill_text_key_repeated[pre_text_token_indices])
                seq_parts_value.append(prefill_text_value_repeated[pre_text_token_indices])
            
            # prefill阶段：视频token
            seq_parts_key.append(vision_key_full[vision_token_indices])
            seq_parts_value.append(vision_value[vision_token_indices])
            
            # prefill阶段：视频后的文本token
            if post_video_text_mask.any():
                post_text_token_indices = torch.where(batch_text_mask)[0][post_video_text_mask]
                seq_parts_key.append(prefill_text_key_repeated[post_text_token_indices])
                seq_parts_value.append(prefill_text_value_repeated[post_text_token_indices])
            
            # decode阶段：历史decode文本tokens
            if decode_text_token_len > 0:
                decode_start_idx = batch_idx * decode_text_seq_len
                decode_end_idx = decode_start_idx + decode_text_seq_len
                batch_decode_keys = decode_text_key_repeated[:, decode_start_idx:decode_end_idx].transpose(0, 1)
                batch_decode_values = decode_text_value_repeated[:, decode_start_idx:decode_end_idx].transpose(0, 1)
                seq_parts_key.append(batch_decode_keys)
                seq_parts_value.append(batch_decode_values)
            
            # cat拼接当前batch的完整序列
            batch_key = torch.cat(seq_parts_key, dim=0)
            batch_value = torch.cat(seq_parts_value, dim=0)
            
            batch_keys.append(batch_key.unsqueeze(0))  # [1, seq_len, num_heads, head_dim]
            batch_values.append(batch_value.unsqueeze(0))

        # 最终拼装所有batch
        final_key_states = torch.cat(batch_keys, dim=0)  # [bsz, seq_len, num_heads, head_dim]
        final_value = torch.cat(batch_values, dim=0)

        # ========== 6. Flash Attention计算 ==========
        # 重新组合query states
        query_states[:, :, :, mask] = q_pe
        
        # 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).contiguous()
        # final_key_states和final_value已经是目标形状，无需transpose
        key_states = final_key_states.contiguous()
        value_states = final_value.contiguous()

        dropout_rate = 0.0 if not self.training else self.attention_dropout

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

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

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

        current_softmax_scale = self.head_dim ** (-0.5)
        effective_softmax_scale = current_softmax_scale * self.softmax_temperature.to(query_states.dtype)

        attn_output = self._flash_attention_forward(
            query_states,
            key_states,
            value_states,
            attention_mask,
            q_len,
            dropout=dropout_rate,
            softmax_scale=effective_softmax_scale.item(),
        )
        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
        attn_output = self.o_proj(attn_output)

        if not output_attentions:
            attn_weights = None

        return attn_output, attn_weights, past_key_value


    # _flash_attention_forward和_upad_input方法保持不变
    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)
        """
        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 DeepseekV3FlashAttention2 __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),
        )







# Copied from transformers.models.mistral.modeling_mistral.MistralSdpaAttention with Mistral->Qwen2
class Qwen2SdpaAttention(Qwen2Attention):
    """
    Qwen2 attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
    `Qwen2Attention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
    SDPA API.
    """


    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,
        vision_text_mask: Optional[torch.Tensor] = 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.`"
            )

        # 根据处理模式选择不同的forward逻辑
        if self.processing_mode == 'no_compress':
            return self._forward_no_compress(
                hidden_states, attention_mask, position_ids, past_key_value,
                output_attentions, use_cache, vision_text_mask, **kwargs
            )
        elif self.processing_mode == 'compress_all':
            return self._forward_compress_all(
                hidden_states, attention_mask, position_ids, past_key_value,
                output_attentions, use_cache, vision_text_mask, **kwargs
            )
        elif self.processing_mode == 'mixed':
            return self._forward_mixed(
                hidden_states, attention_mask, position_ids, past_key_value,
                output_attentions, use_cache, vision_text_mask, **kwargs
            )
        else:
            raise ValueError(f"Unsupported processing mode: {self.processing_mode}")

    def _forward_no_compress(
        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,
        vision_text_mask: Optional[torch.Tensor] = None,
        **kwargs,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        """不压缩处理，但使用部分RoPE形式"""
        bsz, q_len, _ = hidden_states.size()

        # 计算query states
        query_states = self.q_proj(hidden_states)
        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)

        # 分离query的pe和nope部分
        mask = torch.zeros(self.head_dim, dtype=torch.bool, device=hidden_states.device)
        mask[::self.interval] = True 
        q_pe = query_states[..., mask]

        # 计算key和value states（完整的，不压缩）
        key_states = self.k_proj(hidden_states)
        value_states = self.v_proj(hidden_states)
        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)

        # 从完整的key中分离pe部分（类似query的处理方式）
        k_pe = key_states[..., mask]  # 直接裁剪，不使用k_proj_pe
        k_nope = key_states[..., ~mask]

        # 处理缓存
        kv_seq_len = q_len
        if past_key_value is not None:
            if self.layer_idx is None:
                raise ValueError("layer_idx is required when using cache")
            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)

        # 应用旋转位置编码
        cos, sin = self.rotary_emb(k_pe, position_ids)
        q_pe, k_pe = apply_rotary_pos_emb2(q_pe, k_pe, cos, sin, position_ids)

        # 重新组合key states（pe部分使用RoPE后的结果，nope部分保持原样）
        key_states[..., mask] = k_pe
        key_states[..., ~mask] = k_nope

        # 更新缓存 - 统一使用MixedDynamicCache接口，用None占位不需要的部分
        if past_key_value is not None and use_cache:
            cache_kwargs = {"sin": sin, "cos": cos}
            # no_compress模式：文本用完整缓存，视觉部分用None占位
            _, _, final_key_states, final_value_states = past_key_value.update(
                vision_k_pe=None,  # 不压缩模式不需要
                vision_compressed_kv=None,  # 不压缩模式不需要
                text_key_states=key_states,  # 所有token都作为"文本"处理
                text_value_states=value_states,
                layer_idx=self.layer_idx,
                cache_kwargs=cache_kwargs
            )
        else:
            final_key_states = key_states
            final_value_states = value_states

        # 重复到所有头
        final_key_states = repeat_kv(final_key_states, self.num_key_value_groups)
        final_value_states = repeat_kv(final_value_states, self.num_key_value_groups)

        # ========== SDPA计算 ==========
        # 重新组合query states
        query_states[:, :, :, mask] = q_pe

        if attention_mask is not None:
            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()}"
                )

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

        sdpa_effective_scale = (self.head_dim ** (-0.5)) * self.softmax_temperature.to(query_states.dtype)

        attn_output = torch.nn.functional.scaled_dot_product_attention(
            query_states,
            final_key_states,
            final_value_states,
            attn_mask=attention_mask,
            dropout_p=self.attention_dropout if self.training else 0.0,
            # The q_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case q_len == 1.
            is_causal=self.is_causal and attention_mask is None and q_len > 1,
            scale=sdpa_effective_scale.item(),
        )

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

        attn_output = self.o_proj(attn_output)

        return attn_output, None, past_key_value

    def _forward_compress_all(
        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,
        vision_text_mask: Optional[torch.Tensor] = None,
        **kwargs,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        """统一压缩处理，所有token都使用MLA"""
        bsz, q_len, _ = hidden_states.size()

        # 计算query states
        query_states = self.q_proj(hidden_states)
        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)

        # 分离query的pe和nope部分
        mask = torch.zeros(self.head_dim, dtype=torch.bool, device=hidden_states.device)
        mask[::self.interval] = True 
        q_pe = query_states[..., mask]

        # ========== 统一压缩计算（所有token都使用MLA） ==========
        compressed_kv = self.kv_a_proj_nope(hidden_states)  # [bsz, q_len, num_kv_heads*lora_rank]
        k_pe = self.k_proj_pe(hidden_states).view(bsz, q_len, self.num_key_value_heads, (self.head_dim//self.interval)).transpose(1, 2)
        
        # 为缓存准备数据
        k_pe_for_cache = k_pe
        compressed_kv_for_cache = compressed_kv.view(bsz, q_len, self.num_key_value_heads, self.kv_lora_rank).transpose(1, 2)

        # ========== 处理缓存和位置编码 ==========
        kv_seq_len = q_len
        if past_key_value is not None:
            if self.layer_idx is None:
                raise ValueError("layer_idx is required when using cache")
            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)

        # 应用旋转位置编码
        cos, sin = self.rotary_emb(k_pe_for_cache, position_ids)
        q_pe, k_pe_for_cache = apply_rotary_pos_emb2(q_pe, k_pe_for_cache, cos, sin, position_ids)

        # 更新缓存 - 统一使用MixedDynamicCache接口，用None占位不需要的部分
        if past_key_value is not None and use_cache:
            cache_kwargs = {"sin": sin, "cos": cos}
            # compress_all模式：所有token都用压缩缓存，文本部分用None占位
            cached_k_pe, cached_compressed_kv, _, _ = past_key_value.update(
                vision_k_pe=k_pe_for_cache,  # 所有token都作为"视觉"处理（使用压缩）
                vision_compressed_kv=compressed_kv_for_cache,
                text_key_states=None,  # 统一压缩模式不需要
                text_value_states=None,  # 统一压缩模式不需要
                layer_idx=self.layer_idx,
                cache_kwargs=cache_kwargs
            )
        else:
            # 没有缓存，直接使用当前计算的结果
            cached_k_pe = k_pe_for_cache
            cached_compressed_kv = compressed_kv_for_cache

        # ========== 恢复原始维度 ==========
        seq_len = cached_k_pe.shape[-2]
        compressed_kv_reshaped = cached_compressed_kv.transpose(1, 2).reshape(bsz, seq_len, self.num_key_value_heads * self.kv_lora_rank)
        k_nope = self.k_b_proj_nope(compressed_kv_reshaped).view(bsz, seq_len, self.num_heads, self.head_dim - (self.head_dim//self.interval)).transpose(1, 2)
        value_states = self.v_b_proj(compressed_kv_reshaped).view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2)
        
        # 重复k_pe到所有头（只有k_pe需要repeat）
        k_pe_repeated = repeat_kv(cached_k_pe, self.num_key_value_groups)
        
        # k_nope和value_states已经是所有头的维度，不需要repeat
        k_nope = k_nope * self.k_nope_scale_factor.to(k_nope.dtype)
        
        # 合并key states
        final_key_states = torch.zeros(bsz, self.num_heads, seq_len, self.head_dim, device=hidden_states.device, dtype=hidden_states.dtype)
        final_key_states[:, :, :, mask] = k_pe_repeated
        final_key_states[:, :, :, ~mask] = k_nope
        final_value_states = value_states

        # ========== SDPA计算 ==========
        # 重新组合query states
        query_states[:, :, :, mask] = q_pe

        if attention_mask is not None:
            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()}"
                )

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

        sdpa_effective_scale = (self.head_dim ** (-0.5)) * self.softmax_temperature.to(query_states.dtype)

        attn_output = torch.nn.functional.scaled_dot_product_attention(
            query_states,
            final_key_states,
            final_value_states,
            attn_mask=attention_mask,
            dropout_p=self.attention_dropout if self.training else 0.0,
            # The q_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case q_len == 1.
            is_causal=self.is_causal and attention_mask is None and q_len > 1,
            scale=sdpa_effective_scale.item(),
        )

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

        attn_output = self.o_proj(attn_output)

        return attn_output, None, past_key_value


    def _forward_mixed(
        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,
        vision_text_mask: Optional[torch.Tensor] = None,
        **kwargs,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        """
        混合处理模式的分发函数：根据是否有缓存和视觉token来决定使用prefill还是decode
        判断逻辑：
        1. 如果没有缓存(past_key_value is None)，说明是第一次forward，必然是prefill阶段
        2. 如果有缓存但当前输入包含视觉token，也是prefill阶段（虽然这种情况在实际使用中不太常见）
        3. 如果有缓存且当前输入只有文本token，则是decode阶段
        """
        
        # 检查当前输入是否包含视觉token
        has_vision_tokens = (vision_text_mask is not None and vision_text_mask.any())
        
        # 判断是prefill还是decode阶段
        if past_key_value is None or has_vision_tokens:
            # print("-------------------@prefilling------------------------")
            # prefill阶段：第一次forward或包含视觉token
            return self._forward_mixed_prefill(
                hidden_states=hidden_states,
                attention_mask=attention_mask,
                position_ids=position_ids,
                past_key_value=past_key_value,
                output_attentions=output_attentions,
                use_cache=use_cache,
                vision_text_mask=vision_text_mask,
                **kwargs
            )
        else:
            # decode阶段：有缓存且当前输入只有文本token
            return self._forward_mixed_decode(
                hidden_states=hidden_states,
                attention_mask=attention_mask,
                position_ids=position_ids,
                past_key_value=past_key_value,
                output_attentions=output_attentions,
                use_cache=use_cache,
                vision_text_mask=vision_text_mask,
                **kwargs
            )


    def _forward_mixed_prefill(
        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,
        vision_text_mask: Optional[torch.Tensor] = None,
        **kwargs,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        """
        混合模式的prefill阶段处理：
        - 处理视觉token（使用压缩投影）和文本token（使用完整投影）
        - 分别存储到不同的缓存中
        """
        bsz, q_len, _ = hidden_states.size()
        self.prefill_seq_len = q_len

        # ========== 1. 统一计算query states ==========
        query_states = self.q_proj(hidden_states)
        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
        
        mask = torch.zeros(self.head_dim, dtype=torch.bool, device=hidden_states.device)
        mask[::self.interval] = True 
        q_pe = query_states[..., mask]
        # ========== 2. 分离视觉和文本token ==========
        assert vision_text_mask is not None, "vision_text_mask is required in mixed mode"
        
        vision_mask = vision_text_mask  # [bsz, q_len], True表示视觉token
        text_mask = ~vision_text_mask   # [bsz, q_len], True表示文本token
        
        # 获取视觉token的索引和数据
        vision_indices = vision_text_mask.nonzero(as_tuple=True)  # (batch_idx, seq_idx)
        vision_tokens = hidden_states[vision_indices]  # [num_vision_tokens, hidden_size]
        vision_batch_indices = vision_indices[0]       # 每个视觉token属于哪个batch
        vision_seq_indices = vision_indices[1]         # 每个视觉token在序列中的位置
        num_vision_tokens = vision_tokens.shape[0]
        
        # 获取文本token的索引和数据
        text_indices = (~vision_text_mask).nonzero(as_tuple=True)
        text_tokens = hidden_states[text_indices]      # [num_text_tokens, hidden_size]
        text_batch_indices = text_indices[0]           # 每个文本token属于哪个batch
        text_seq_indices = text_indices[1]             # 每个文本token在序列中的位置
        num_text_tokens = text_tokens.shape[0]
        # 存储到当前层的属性中
        self.prefill_vision_indices = vision_indices
        self.prefill_text_indices = text_indices
        # ========== 3. 分别处理视觉和文本token ==========
        # 视觉token：使用压缩投影
        vision_compressed_kv = self.kv_a_proj_nope(vision_tokens).view(num_vision_tokens, self.num_key_value_heads, self.kv_lora_rank).transpose(0,1)
        vision_k_pe = self.k_proj_pe(vision_tokens).view(num_vision_tokens, self.num_key_value_heads,self.head_dim//self.interval).transpose(0,1)  
        
        # 文本token：使用原始投影
        text_key_states = self.k_proj(text_tokens).view(num_text_tokens, self.num_key_value_heads, self.head_dim).transpose(0,1)  
        text_value_states = self.v_proj(text_tokens).view(num_text_tokens, self.num_key_value_heads, self.head_dim).transpose(0,1)  

        # ========== 4. 处理position encoding ==========
        kv_seq_len = q_len
        if past_key_value is not None:
            if self.layer_idx is None:
                raise ValueError("layer_idx is required when using cache")
            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)

        cos, sin = self.rotary_emb(vision_k_pe, position_ids)
        text_key_pe = text_key_states[..., mask]
        q_pe, vision_k_pe, text_key_pe = apply_rotary_pos_emb2_separate(q_pe, vision_k_pe, text_key_pe, cos, sin,vision_indices,text_indices)
        text_key_states[..., mask] = text_key_pe

        # ========== 5. 更新缓存（分离存储） ==========
        if past_key_value is not None and use_cache:
            cache_kwargs = {"sin": sin, "cos": cos}
            vision_k_pe, vision_compressed_kv, text_key_states, text_value_states = past_key_value.update(
                vision_k_pe=vision_k_pe,
                vision_compressed_kv=vision_compressed_kv,
                text_key_states=text_key_states,
                text_value_states=text_value_states,
                layer_idx=self.layer_idx,
                cache_kwargs=cache_kwargs
            )#(num_key_value_heads,num_text_tokens,  self.head_dim)，可以正常在MixedDynamicCache中工作

        # ========== 6. 根据当前序列的mask按需组装key/value ==========
        final_key_states = torch.zeros(bsz, self.num_heads, kv_seq_len, self.head_dim, device=hidden_states.device, dtype=hidden_states.dtype)
        final_value_states = torch.zeros(bsz, self.num_heads, kv_seq_len, self.head_dim,
                                device=hidden_states.device, dtype=hidden_states.dtype)
        
        # 组装视觉token的K/V（从压缩格式恢复）
        if num_vision_tokens > 0 and vision_k_pe is not None:
            # 从压缩状态恢复nope和value
            # vision_compressed_kv: [num_key_value_heads, num_vision_tokens, lora_rank]
            # 需要重新reshape为 [num_vision_tokens, num_key_value_heads * lora_rank]

            cache_num_vision_tokens = vision_compressed_kv.shape[1]
            vision_compressed_reshaped = vision_compressed_kv.transpose(0, 1).reshape(
                cache_num_vision_tokens, self.num_key_value_heads * self.kv_lora_rank)  # [num_vision_tokens, num_key_value_heads * lora_rank]
            
            # 恢复k_nope和value
            vision_k_nope = self.k_b_proj_nope(vision_compressed_reshaped).view(
                cache_num_vision_tokens, self.num_heads, self.head_dim - self.head_dim // self.interval)
            vision_value = self.v_b_proj(vision_compressed_reshaped).view(
                cache_num_vision_tokens, self.num_heads, self.head_dim)
            vision_k_nope = vision_k_nope * self.k_nope_scale_factor.to(vision_k_nope.dtype)
            
            # 重复k_pe到所有头: [num_key_value_heads, num_vision_tokens, pe_dim] -> [num_heads, num_vision_tokens, pe_dim]
            vision_k_pe_repeated = repeat_kv(vision_k_pe.unsqueeze(0), self.num_key_value_groups).squeeze(0)

            # 使用存储的索引填充到final tensors
            vision_key_full = torch.empty(len(self.prefill_vision_indices[0]), self.num_heads, self.head_dim, device=final_key_states.device, dtype=final_key_states.dtype)
            vision_key_full[..., mask] = vision_k_pe_repeated.transpose(0, 1)
            vision_key_full[..., ~mask] = vision_k_nope
            final_key_states[self.prefill_vision_indices[0], :,self.prefill_vision_indices[1],  :] = vision_key_full
            final_value_states[self.prefill_vision_indices[0], :, self.prefill_vision_indices[1], :] = vision_value

        # 组装文本token的K/V（直接使用完整格式）
        if num_text_tokens > 0 and text_key_states is not None:
            # text_key_states: [num_key_value_heads, num_text_tokens, head_dim]，text_value_states: [num_key_value_heads, num_text_tokens, head_dim]
            # 重复k/v到所有头
            text_key_repeated = repeat_kv(text_key_states.unsqueeze(0), self.num_key_value_groups).squeeze(0)
            text_value_repeated = repeat_kv(text_value_states.unsqueeze(0), self.num_key_value_groups).squeeze(0)

            # 使用存储的索引填充到final tensors
            final_key_states[self.prefill_text_indices[0], :, self.prefill_text_indices[1], :] = text_key_repeated.transpose(0, 1)
            final_value_states[self.prefill_text_indices[0], :, self.prefill_text_indices[1], :] = text_value_repeated.transpose(0, 1)
            final_value_states[self.prefill_text_indices[0], :, self.prefill_text_indices[1], :] = text_value_repeated.transpose(0, 1)


        # ========== SDPA计算 ==========
        # 重新组合query states
        query_states[:, :, :, mask] = q_pe

        if attention_mask is not None:
            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()}"
                )

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

        sdpa_effective_scale = (self.head_dim ** (-0.5)) * self.softmax_temperature.to(query_states.dtype)

        attn_output = torch.nn.functional.scaled_dot_product_attention(
            query_states,
            final_key_states,
            final_value_states,
            attn_mask=attention_mask,
            dropout_p=self.attention_dropout if self.training else 0.0,
            # The q_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case q_len == 1.
            is_causal=self.is_causal and attention_mask is None and q_len > 1,
            scale=sdpa_effective_scale.item(),
        )

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

        attn_output = self.o_proj(attn_output)

        return attn_output, None, past_key_value



    def _forward_mixed_decode(
        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,
        vision_text_mask: Optional[torch.Tensor] = None,
        **kwargs,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        """
        混合模式的decode阶段处理：
        - 当前输入只有文本token（使用完整投影）
        - 从缓存中读取历史的视觉token（压缩格式）和文本token（完整格式）
        - 使用存储的索引正确恢复历史token位置，新文本token追加到末尾
        """
        bsz, q_len, _ = hidden_states.size()
        assert past_key_value is not None, "past_key_value is required in decode stage"
        assert q_len == 1, "Decode stage should have q_len=1"  # decode阶段通常一次只处理一个token
        
        # ========== 1. 统一计算query states ==========
        query_states = self.q_proj(hidden_states)
        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
        
        mask = torch.zeros(self.head_dim, dtype=torch.bool, device=hidden_states.device)
        mask[::self.interval] = True 
        q_pe = query_states[..., mask]

        # ========== 2. 处理当前文本token（使用完整投影） ==========
        # 当前输入全是文本token，使用原始投影
        current_text_key = self.k_proj(hidden_states).view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
        current_text_value = self.v_proj(hidden_states).view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)

        # ========== 3. 处理position encoding ==========
        kv_seq_len = q_len
        if past_key_value is not None:
            if self.layer_idx is None:
                raise ValueError("layer_idx is required when using cache")
            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)

        cos, sin = self.rotary_emb(current_text_key, position_ids)
        q_pe = apply_rotary_pos_emb2_single(q_pe, cos, sin)
        
        # 对当前文本token的key应用位置编码
        current_text_key_pe = current_text_key[..., mask]
        current_text_key_pe = apply_rotary_pos_emb2_single(current_text_key_pe, cos, sin)
        current_text_key[..., mask] = current_text_key_pe

        # ========== 4. 更新缓存（按batch×seqlen格式存储新文本） ==========
        # 直接transpose和view，更简洁的维度变换
        current_text_key_for_cache = current_text_key.transpose(0, 1).contiguous().view(self.num_key_value_heads, bsz * q_len, self.head_dim)
        current_text_value_for_cache = current_text_value.transpose(0, 1).contiguous().view(self.num_key_value_heads, bsz * q_len, self.head_dim)

        if use_cache:
            cache_kwargs = {"sin": sin, "cos": cos}
            cached_vision_k_pe, cached_vision_compressed_kv, cached_text_key, cached_text_value = past_key_value.update(
                vision_k_pe=None,  # decode阶段没有新的视觉token
                vision_compressed_kv=None,
                text_key_states=current_text_key_for_cache,  # 按统一格式添加当前文本token
                text_value_states=current_text_value_for_cache,
                layer_idx=self.layer_idx,
                cache_kwargs=cache_kwargs
            )#(num_key_value_heads , num_tokens , head_dim)
        # ========== 5. 使用存储的索引组装完整的key/value ==========
        vision_token_len = cached_vision_k_pe.shape[-2]  # prefill阶段的vision tokens数量
        prefill_text_token_len = len(self.prefill_text_indices[0])  # prefill阶段的text tokens数量
        total_cached_text_len = cached_text_key.shape[-2]  # 缓存中所有text tokens数量
        decode_text_token_len = total_cached_text_len - prefill_text_token_len  # decode阶段累积的text tokens数量
        prefill_total_token_len = vision_token_len + prefill_text_token_len#prefill阶段的tokens数量
        decode_text_seq_len = decode_text_token_len//bsz #decode阶段的序列长度，不是token数量
        total_seq_len = self.prefill_seq_len + decode_text_seq_len #总共的序列长度，不是token数量
        # 初始化最终的key/value tensors
        final_key_states = torch.zeros(bsz, self.num_heads, total_seq_len, self.head_dim, device=hidden_states.device, dtype=hidden_states.dtype)
        final_value_states = torch.zeros(bsz, self.num_heads, total_seq_len, self.head_dim,device=hidden_states.device, dtype=hidden_states.dtype)

        # 1. 恢复历史vision tokens（使用prefill索引）
        if vision_token_len > 0:
            vision_compressed_reshaped = cached_vision_compressed_kv.transpose(0, 1).reshape(vision_token_len, self.num_key_value_heads * self.kv_lora_rank)
            
            vision_k_nope = self.k_b_proj_nope(vision_compressed_reshaped).view(vision_token_len, self.num_heads, self.head_dim - self.head_dim // self.interval)
            vision_value = self.v_b_proj(vision_compressed_reshaped).view(vision_token_len, self.num_heads, self.head_dim)
            vision_k_nope = vision_k_nope * self.k_nope_scale_factor.to(vision_k_nope.dtype)
            
            vision_k_pe_repeated = repeat_kv(cached_vision_k_pe.unsqueeze(0), self.num_key_value_groups).squeeze(0)#[num_heads, vision_token_len, pe_dim]

            vision_key_full = torch.empty(len(self.prefill_vision_indices[0]), self.num_heads, self.head_dim, device=final_key_states.device, dtype=final_key_states.dtype)
            vision_key_full[..., mask] = vision_k_pe_repeated.transpose(0, 1)
            vision_key_full[..., ~mask] = vision_k_nope
            final_key_states[self.prefill_vision_indices[0], :, self.prefill_vision_indices[1], :] = vision_key_full
            final_value_states[self.prefill_vision_indices[0], :, self.prefill_vision_indices[1], :] = vision_value

        # 2. 恢复历史prefill text tokens（使用prefill索引）
        if prefill_text_token_len > 0:
            # 取缓存中前面的prefill text tokens
            prefill_text_key = cached_text_key[:, :prefill_text_token_len, :]  # [num_key_value_heads, prefill_text_len, head_dim]
            prefill_text_value = cached_text_value[:, :prefill_text_token_len, :]
            
            # 重复到所有头 - 现在都是 [num_heads, prefill_text_len, head_dim]
            prefill_text_key_repeated = repeat_kv(prefill_text_key.unsqueeze(0), self.num_key_value_groups).squeeze(0)
            prefill_text_value_repeated = repeat_kv(prefill_text_value.unsqueeze(0), self.num_key_value_groups).squeeze(0)
            
            final_key_states[self.prefill_text_indices[0], :, self.prefill_text_indices[1], :] = prefill_text_key_repeated.transpose(0, 1)
            final_value_states[self.prefill_text_indices[0], :, self.prefill_text_indices[1], :] = prefill_text_value_repeated.transpose(0, 1)

        # 3. 添加decode阶段的text tokens（按顺序追加在后面）
        if decode_text_token_len > 0:
            # 取缓存中后面的decode text tokens
            decode_text_key = cached_text_key[:, prefill_text_token_len:, :]  # [num_key_value_heads, decode_text_len, head_dim]
            decode_text_value = cached_text_value[:, prefill_text_token_len:, :]
            
            # 重复到所有头并reshape为batch格式
            decode_text_key_repeated = repeat_kv(decode_text_key.unsqueeze(0), self.num_key_value_groups).squeeze(0)  # [num_heads, decode_text_len, head_dim]
            decode_text_value_repeated = repeat_kv(decode_text_value.unsqueeze(0), self.num_key_value_groups).squeeze(0)
            # 将decode tokens reshape回batch格式: [num_heads, bsz*decode_text_seq_len, head_dim] -> [bsz, num_heads, decode_text_seq_len, head_dim]
            decode_text_key_batch = decode_text_key_repeated.view(self.num_heads, bsz, decode_text_seq_len, self.head_dim).transpose(0, 1)
            decode_text_value_batch = decode_text_value_repeated.view(self.num_heads, bsz, decode_text_seq_len, self.head_dim).transpose(0, 1)
            
            final_key_states[:, :, self.prefill_seq_len:, :] = decode_text_key_batch
            final_value_states[:, :, self.prefill_seq_len :, :] = decode_text_value_batch

        # ========== SDPA计算 ==========
        # 重新组合query states
        query_states[:, :, :, mask] = q_pe

        if attention_mask is not None:
            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()}"
                )

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

        sdpa_effective_scale = (self.head_dim ** (-0.5)) * self.softmax_temperature.to(query_states.dtype)

        attn_output = torch.nn.functional.scaled_dot_product_attention(
            query_states,
            final_key_states,
            final_value_states,
            attn_mask=attention_mask,
            dropout_p=self.attention_dropout if self.training else 0.0,
            # The q_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case q_len == 1.
            is_causal=self.is_causal and attention_mask is None and q_len > 1,
            scale=sdpa_effective_scale.item(),
        )

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

        attn_output = self.o_proj(attn_output)

        return attn_output, None, past_key_value





QWEN2_ATTENTION_CLASSES = {
    "eager": Qwen2Attention,
    "flash_attention_2": Qwen2FlashAttention2,
    "sdpa": Qwen2SdpaAttention,
}




class Qwen2DecoderLayer(nn.Module):
    def __init__(self, config: Qwen2Config, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size

        if config.use_sliding_window and config._attn_implementation != "flash_attention_2":
            logger.warning_once(
                f"Sliding Window Attention is enabled but not implemented for `{config._attn_implementation}`; "
                "unexpected results may be encountered."
            )
        self.self_attn = QWEN2_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx)

        self.mlp = Qwen2MLP(config)
        self.input_layernorm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_value: Optional[Tuple[torch.Tensor]] = None,
        output_attentions: Optional[bool] = False,
        use_cache: Optional[bool] = False,
        vision_text_mask: Optional[torch.Tensor] = None,  # 新增参数
        **kwargs,
    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
        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.`"
            )
        """
        Args:
            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
            attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
                `(batch, sequence_length)` where padding elements are indicated by 0.
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
            use_cache (`bool`, *optional*):
                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
                (see `past_key_values`).
            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
            vision_text_mask (`torch.Tensor`, *optional*): mask to distinguish vision and text tokens
        """

        residual = hidden_states

        hidden_states = self.input_layernorm(hidden_states)

        # Self Attention - 传递vision_text_mask
        hidden_states, self_attn_weights, present_key_value = self.self_attn(
            hidden_states=hidden_states,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_value=past_key_value,
            output_attentions=output_attentions,
            use_cache=use_cache,
            vision_text_mask=vision_text_mask,  # 传递mask
        )
        hidden_states = residual + hidden_states

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

        outputs = (hidden_states,)

        if output_attentions:
            outputs += (self_attn_weights,)

        if use_cache:
            outputs += (present_key_value,)

        return outputs


QWEN2_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 ([`Qwen2Config`]):
            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 Qwen2 Model outputting raw hidden-states without any specific head on top.",
    QWEN2_START_DOCSTRING,
)
class Qwen2PreTrainedModel(PreTrainedModel):
    config_class = Qwen2Config
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["Qwen2DecoderLayer"]
    _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_()


QWEN2_INPUTS_DOCSTRING = r"""
    Args:
        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
            it.

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

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

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

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

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

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

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

            - 1 indicates the head is **not masked**,
            - 0 indicates the head is **masked**.
        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
            config.n_positions - 1]`.

            [What are position IDs?](../glossary#position-ids)
        past_key_values (`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 Qwen2 Model outputting raw hidden-states without any specific head on top.",
    QWEN2_START_DOCSTRING,
)
class Qwen2Model_Flash(Qwen2PreTrainedModel):
    """
    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`Qwen2DecoderLayer`]

    Args:
        config: Qwen2Config
    """

    def __init__(self, config: Qwen2Config):
        super().__init__(config)
        self.padding_idx = config.pad_token_id
        self.vocab_size = config.vocab_size

        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
        self.layers = nn.ModuleList(
            [Qwen2DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self._attn_implementation = config._attn_implementation
        self.norm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)

        self.gradient_checkpointing = False


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


    def create_vision_text_mask(self, batch_size, seq_len, device):
        """
        创建区分视觉和文本token的mask
        移到Model类中，避免循环引用
        Returns:
            torch.Tensor: shape=(batch_size, seq_len), True表示视觉token，False表示文本token
            如果没有视觉信息则返回None
        """
        if (hasattr(self, 'first_image_token_position') and 
            hasattr(self, 'num_image_token_lens')):
            vision_mask = torch.zeros(batch_size, seq_len, dtype=torch.bool, device=device)
            for i in range(batch_size):
                if (i < len(self.first_image_token_position) and 
                    i < len(self.num_image_token_lens)):
                    
                    image_start = self.first_image_token_position[i]
                    if image_start != -1:  # 有图像
                        image_length = self.num_image_token_lens[i]
                        end_pos = min(image_start + image_length, seq_len)
                        vision_mask[i, image_start:end_pos] = True
            
            return vision_mask
        return None

    def get_input_embeddings(self):
        return self.embed_tokens

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

    @add_start_docstrings_to_model_forward(QWEN2_INPUTS_DOCSTRING)
    def forward(
        self,
        input_ids: torch.LongTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[List[torch.FloatTensor]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        labels: Optional[torch.Tensor] = None,
    ) -> Union[Tuple, BaseModelOutputWithPast]:
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        use_cache = use_cache if use_cache is not None else self.config.use_cache

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

        # retrieve input_ids and inputs_embeds
        if input_ids is not None and inputs_embeds is not None:
            raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
        elif input_ids is not None:
            batch_size, seq_length = input_ids.shape
        elif inputs_embeds is not None:
            batch_size, seq_length, _ = inputs_embeds.shape
        else:
            raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")

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

        past_key_values_length = 0

        if use_cache:
            if past_key_values is None:
                past_key_values = MixedDynamicCache()
            past_key_values_length = past_key_values.get_usable_length(seq_length)

        if position_ids is None:
            device = input_ids.device if input_ids is not None else inputs_embeds.device
            position_ids = torch.arange(
                past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
            )
            position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
        else:
            position_ids = position_ids.view(-1, seq_length).long()

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

        if attention_mask is not None and self._attn_implementation == "flash_attention_2" and use_cache:
            is_padding_right = attention_mask[:, -1].sum().item() != batch_size
            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 Qwen2. Make sure to "
                    " call `tokenizer.padding_side  = 'left'` before tokenizing the input. "
                )

        if self._attn_implementation == "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._attn_implementation == "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,
                sliding_window=self.config.sliding_window,
            )

        hidden_states = inputs_embeds
        
        # 在这里创建vision_text_mask，避免循环引用
        vision_text_mask = self.create_vision_text_mask(batch_size, seq_length, hidden_states.device)

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

        for layer_idx, decoder_layer in enumerate(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,
                    attention_mask,
                    position_ids,
                    past_key_values,
                    output_attentions,
                    use_cache,
                    vision_text_mask,  # 传递mask
                )
            else:
                layer_outputs = decoder_layer(
                    hidden_states,
                    attention_mask=attention_mask,
                    position_ids=position_ids,
                    past_key_value=past_key_values,
                    output_attentions=output_attentions,
                    use_cache=use_cache,
                    vision_text_mask=vision_text_mask,  # 传递mask
                )

            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],)

            ###### copy from pdrop #########
            # rank & drop after specific layer
            # only drop in prefill stage when inference
            rank_layer = layer_idx+1
            if rank_layer in self.llm_compress_layer_list:
                if hidden_states.shape[1] != 1:  # prefill stage or training
                    stage = self.llm_compress_layer_list.index(rank_layer) # determine current stage
                    (
                        position_ids,
                        attention_mask,
                        hidden_states,
                        labels   # update labels and return
                    ) = self.video_level_compress(    
                        cur_num = stage,
                        rank_layer = rank_layer,
                        features = hidden_states,   
                        position_ids=position_ids,
                        attention_mask=attention_mask,
                        labels = labels
                    )
                    
                    # 重新创建vision_text_mask，因为序列长度可能已经改变
                    vision_text_mask = self.create_vision_text_mask(batch_size, hidden_states.shape[1], hidden_states.device)
                    
                    # process attention_mask again after updating 
                    if self._attn_implementation == "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._attn_implementation == "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, hidden_states.shape[1]),
                            hidden_states,
                            past_key_values_length,
                        )
                    else:
                        # 4d mask is passed through the layers
                        attention_mask = _prepare_4d_causal_attention_mask(
                            attention_mask,
                            (batch_size, hidden_states.shape[1]),
                            hidden_states,
                            past_key_values_length,
                            sliding_window=self.config.sliding_window,
                        )
                        
                else:
                    # update position_ids in decoding stage when inference
                    stage = self.llm_compress_layer_list.index(rank_layer) # determine current stage
                    cur_visual_length = [int(cur_image_token * self.llm_image_token_ratio_list[stage]) for cur_image_token in self.num_image_token_lens]
                    next_visual_length = [int(cur_image_token * self.llm_image_token_ratio_list[stage + 1]) for cur_image_token in self.num_image_token_lens]
                    new_position_ids = []
                    for idx, cur_position_ids in enumerate(position_ids):
                        cur_position_ids = cur_position_ids - (cur_visual_length[idx] - next_visual_length[idx])
                        new_position_ids.append(cur_position_ids)
                    assert idx == 0, idx
                    position_ids = torch.tensor(new_position_ids, dtype=torch.long).unsqueeze(0)

            #################

        hidden_states = self.norm(hidden_states)

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

        next_cache = next_decoder_cache

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


    # implementation of pdrop
    def video_level_compress(
        self, cur_num, rank_layer, features ,
        position_ids, attention_mask, labels
    ):

        if self.llm_compress_type == 'uniform0_attention':
            if cur_num == 0:
                llm_compress_type = 'uniform'
            else:
                llm_compress_type = 'attention'
        else:
            llm_compress_type = self.llm_compress_type
            
        _labels = labels
        _position_ids = position_ids
        _attention_mask = attention_mask

        if position_ids is None:
            position_ids = torch.arange(0, features.shape[1], dtype=torch.long, device=features.device).unsqueeze(0)
        
        if getattr(self.config, 'tokenizer_padding_side', 'right') == "right":
            
            batch_size = features.shape[0]
            image_tokens = [int(cur_image_token * self.llm_image_token_ratio_list[cur_num]) for cur_image_token in self.num_image_token_lens]
            keep_length = [int(cur_image_token * self.llm_image_token_ratio_list[cur_num + 1]) for cur_image_token in self.num_image_token_lens]

            features_list = []
            attention_mask_list = []
            labels_list = []

            if attention_mask is None:
                attention_mask = torch.ones((batch_size,features.shape[1]), dtype=torch.bool, device=features.device)
            else:
                attention_mask = attention_mask.bool()
            if labels is None:
                labels = torch.full((batch_size,features.shape[1]), IGNORE_INDEX, device=features.device)            


            if 'attention' in llm_compress_type:
                # obtain query_states and key_states to calculate attention map
                hidden_states= features.clone().detach()

                self_attn = self.layers[rank_layer].self_attn
                hidden_states = self.layers[rank_layer].input_layernorm(hidden_states)

                num_heads = self_attn.num_heads
                num_key_value_heads = self_attn.num_key_value_heads
                head_dim = self_attn.head_dim

                bsz, q_len, _ = hidden_states.size()

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

                query_states = query_states.view(bsz, q_len, num_heads, head_dim).transpose(1, 2)
                key_states = key_states.view(bsz, q_len, num_key_value_heads, head_dim).transpose(1, 2)
                value_states = value_states.view(bsz, q_len, num_key_value_heads, head_dim).transpose(1, 2)

                kv_seq_len = key_states.shape[-2]
                cos, sin = self_attn.rotary_emb(value_states, seq_len=kv_seq_len)

                query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
                key_states = repeat_kv(key_states, self_attn.num_key_value_groups)

                # attention_mask
                eager_attention_mask = _prepare_4d_causal_attention_mask(
                    attention_mask, (batch_size, q_len), hidden_states, past_key_values_length=0
                ).to(device=query_states.device)
        
            # take valid features
            features = [cur_features[cur_attention_mask] for cur_features, cur_attention_mask in zip(features, attention_mask)]
            labels = [cur_labels[cur_attention_mask] for cur_labels, cur_attention_mask in zip(labels, attention_mask)]
            attention_mask = [cur_attention_mask[cur_attention_mask] for cur_attention_mask, cur_attention_mask in zip(attention_mask, attention_mask)]
            
            # rank & drop
            for i in range(batch_size):
                image_index = self.first_image_token_position[i]
                if image_index == -1:
                    cur_input_embeds = features[i]
                    features_list.append(cur_input_embeds)
                    attention_mask_list.append(attention_mask[i])
                    labels_list.append(labels[i])
                    continue

                if 'attention' in llm_compress_type:
                
                    # obtain current states
                    cur_key_states = key_states[i]
                    cur_query_states = query_states[i] 
                    cur_eager_attention_mask = eager_attention_mask[i] 
                    
                    # choose last instruction token as query
                    if self.training:
                        answer_index = torch.where(labels[i] != -100)[0].tolist()
                        index_before_answer = []
                        for index in answer_index:
                            if labels[i][index-1] == -100:
                                index_before_answer.append(index-1)
                        if index_before_answer == []:
                            cur_input_embeds = features[i]
                            features_list.append(cur_input_embeds)
                            attention_mask_list.append(attention_mask[i])
                            labels_list.append(labels[i])
                            continue

                        index_before_answer=torch.tensor(index_before_answer,device=labels[0].device) 
                        text_query_states = cur_query_states[:,index_before_answer,:]  
                        text_eager_attention_mask = cur_eager_attention_mask[:,index_before_answer,:]

                    else:
                        prompt_total_len = self.text_prompt_lens[i] + image_tokens[i]
                        text_query_states = cur_query_states[:,prompt_total_len-1,:].unsqueeze(1)  
                        text_eager_attention_mask = cur_eager_attention_mask[:,prompt_total_len-1,:].unsqueeze(1)
                    
                    # calculate attention map
                    attn_weights = torch.matmul(text_query_states, cur_key_states.transpose(1, 2)) / math.sqrt(head_dim) #(num_head, text_token,seq_len)
                    attn_weights = attn_weights + text_eager_attention_mask
                    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype) #(num_head, text_token,seq_len)
                    
                    attention_avg_head = torch.mean(attn_weights, dim=0) # ave across heads
                    attention_avg_head = attention_avg_head[:,image_index:image_index+image_tokens[i]] # select image token as keys
                    attention_avg_text = torch.mean(attention_avg_head, dim=0) # (576)

                    if llm_compress_type == 'attention':
                        top_rank_index = attention_avg_text.topk(keep_length[i]).indices
                    else:
                        raise NotImplementedError(llm_compress_type)

                elif llm_compress_type == 'uniform':
                    top_rank_index = torch.linspace(0, image_tokens[i]-1, keep_length[i], dtype=torch.long)
                else:
                    raise NotImplementedError(llm_compress_type)
                    
                top_rank_index = top_rank_index + image_index  
                top_rank_index= top_rank_index.sort().values  

                start_index = image_index + image_tokens[i]
                new_input_embeds = torch.cat([features[i][ :image_index, :] ,features[i][ top_rank_index, :], features[i][start_index:, :]], dim=0)

                new_labels = torch.cat([labels[i][ :image_index],labels[i][ top_rank_index], labels[i][start_index:]], dim=0)
                new_attention_mask = torch.cat([attention_mask[i][:image_index], attention_mask[i][top_rank_index], attention_mask[i][start_index:]], dim=0)

                features_list.append(new_input_embeds)
                attention_mask_list.append(new_attention_mask)
                labels_list.append(new_labels)                

            # Truncate sequences to max length as image embeddings can make the sequence longer
            tokenizer_model_max_length = getattr(self.config, 'tokenizer_model_max_length', None)
            if tokenizer_model_max_length is not None:
                new_input_embeds = [x[:tokenizer_model_max_length] for x in features_list]
                new_attention_mask = [x[:tokenizer_model_max_length] for x in attention_mask_list]
                new_labels = [x[:tokenizer_model_max_length] for x in labels_list]

            max_len = max(x.shape[0] for x in new_input_embeds)

            # padding the sequences to form batch
            embeds_padded=[]
            labels_paded=[]
            attention_mask_padded=[]
            position_ids = torch.zeros((batch_size, max_len), dtype=position_ids.dtype, device=position_ids.device)
            for i, (cur_new_embed, cur_new_labels) in enumerate(zip(new_input_embeds, new_labels)):
                cur_len_emb=cur_new_embed.shape[0]
                dif=max_len - cur_len_emb  # padding to longest seq
                
                cur_new_embed = torch.cat([cur_new_embed,torch.zeros((dif, cur_new_embed.shape[1]), dtype=cur_new_embed.dtype, device=cur_new_embed.device)],dim=0)
                cur_new_labels = torch.cat([cur_new_labels,torch.full((dif,),IGNORE_INDEX,dtype=cur_new_labels.dtype, device=cur_new_labels.device)],dim=0)
                cur_attention_mask = new_attention_mask[i]
                cur_attention_mask = torch.cat([cur_attention_mask,torch.full((dif,),False, dtype=cur_attention_mask.dtype, device=cur_attention_mask.device)],dim=0)
                
                embeds_padded.append(cur_new_embed)
                labels_paded.append(cur_new_labels)
                attention_mask_padded.append(cur_attention_mask)

                cur_len = new_attention_mask[i].sum().item()
                position_ids[i, :cur_len] = torch.arange(0, cur_len, dtype=position_ids.dtype, device=position_ids.device)
            

            new_input_embeds = torch.stack(embeds_padded,dim=0)
            new_input_embeds = new_input_embeds.to(features[0].dtype)

            new_attention_mask = torch.stack(attention_mask_padded,dim=0)
            new_labels = torch.stack(labels_paded,dim=0)
            
            if _position_ids is None:
                position_ids = None
            if _labels is None:
                new_labels = None

            if _attention_mask is None:
                new_attention_mask = None
            else:
                new_attention_mask = new_attention_mask.to(dtype=_attention_mask.dtype)

            return position_ids, new_attention_mask, new_input_embeds, new_labels
        
        else:
            raise ValueError(f"Unexpected tokenizer_padding_side: {self.config.tokenizer_padding_side}")


class Qwen2ForCausalLM_Flash(Qwen2PreTrainedModel):
    _tied_weights_keys = ["lm_head.weight"]

    def __init__(self, config):
        super().__init__(config)
        self.model = Qwen2Model_Flash(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.embed_tokens

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

    def get_output_embeddings(self):
        return self.lm_head

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

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

    def get_decoder(self):
        return self.model

    @add_start_docstrings_to_model_forward(QWEN2_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
    def forward(
        self,
        input_ids: torch.LongTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[List[torch.FloatTensor]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, CausalLMOutputWithPast]:
        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, Qwen2ForCausalLM

        >>> model = Qwen2ForCausalLM.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
        )
        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, labels = self.model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            labels=labels 
        )

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

        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()
            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:]
            return (loss,) + output if loss is not None else output

        return CausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )

    def prepare_inputs_for_generation(
        self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
    ):
        # Omit tokens covered by past_key_values
        if past_key_values is not None:
            cache_length = past_key_values.get_seq_length()
            past_length = past_key_values.seen_tokens
            max_cache_length = past_key_values.get_max_length()

            # 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 exclusively 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] :]

        # 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}
        else:
            model_inputs = {"input_ids": input_ids}

        model_inputs.update(
            {
                "position_ids": position_ids,
                "past_key_values": past_key_values,
                "use_cache": kwargs.get("use_cache"),
                "attention_mask": attention_mask,
            }
        )
        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



# 在Qwen2RMSNorm类之前添加MixedDynamicCache类
class MixedDynamicCache(Cache):
    """
    混合动态缓存，分别处理视觉和文本token的KV缓存
    
    视觉token使用低秩压缩：k_pe + compressed_kv
    文本token保持完整性能：key_states + value_states
    
    缓存结构：
    - vision_k_pe_cache: 视觉token的位置编码部分 [batch_size, num_heads, seq_len, pe_dim]
    - vision_compressed_kv_cache: 视觉token的压缩KV [batch_size, num_heads, seq_len, lora_rank]
    - text_key_cache: 文本token的完整key [batch_size, num_heads, seq_len, head_dim]
    - text_value_cache: 文本token的完整value [batch_size, num_heads, seq_len, head_dim]
    """
    
    def __init__(self, _distributed_cache_data: Iterable = None) -> None:
        super().__init__()
        
        # 视觉token缓存 (低秩压缩) - 对应DynamicCache的key_cache/value_cache结构
        self.vision_k_pe_cache: List[torch.Tensor] = []
        self.vision_compressed_kv_cache: List[torch.Tensor] = []
        
        # 文本token缓存 (完整) - 对应DynamicCache的key_cache/value_cache结构
        self.text_key_cache: List[torch.Tensor] = []
        self.text_value_cache: List[torch.Tensor] = []
        
        # 保持与DynamicCache相同的分布式数据处理逻辑
        if _distributed_cache_data is not None:
            for vision_k_pe, vision_compressed_kv, text_key, text_value in _distributed_cache_data:
                self.vision_k_pe_cache.append(vision_k_pe)
                self.vision_compressed_kv_cache.append(vision_compressed_kv)
                self.text_key_cache.append(text_key)
                self.text_value_cache.append(text_value)

    @property
    def seen_tokens(self) -> int:
        """Returns the number of tokens in the cache."""
        return self.get_seq_length()

    def __getitem__(self, layer_idx: int) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
        """
        支持向后兼容的索引访问，类似DynamicCache的__getitem__
        返回指定层的缓存：(vision_k_pe, vision_compressed_kv, text_key, text_value)
        """
        if layer_idx < len(self):
            return (
                self.vision_k_pe_cache[layer_idx],
                self.vision_compressed_kv_cache[layer_idx], 
                self.text_key_cache[layer_idx],
                self.text_value_cache[layer_idx]
            )
        else:
            raise KeyError(f"Cache only has {len(self)} layers, attempted to access layer with index {layer_idx}")

    def __iter__(self):
        """
        支持向后兼容的迭代，类似DynamicCache的__iter__
        """
        for layer_idx in range(len(self)):
            yield (
                self.vision_k_pe_cache[layer_idx],
                self.vision_compressed_kv_cache[layer_idx],
                self.text_key_cache[layer_idx], 
                self.text_value_cache[layer_idx]
            )

    def __len__(self):
        """
        支持向后兼容的长度查询，类似DynamicCache的__len__
        返回缓存的层数
        """
        return len(self.vision_k_pe_cache)

    def update(
        self,
        vision_k_pe: Optional[torch.Tensor] = None,
        vision_compressed_kv: Optional[torch.Tensor] = None,
        text_key_states: Optional[torch.Tensor] = None,
        text_value_states: Optional[torch.Tensor] = None,
        layer_idx: int = 0,
        cache_kwargs: Optional[Dict[str, Any]] = None,
    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
        """
        更新指定层的缓存，对齐DynamicCache的update逻辑和性能优化
        """
        # 处理跳过的层 - 与DynamicCache保持一致的逻辑
        if len(self.vision_k_pe_cache) <= layer_idx:
            # 填充跳过的层，使用空tensor
            for _ in range(len(self.vision_k_pe_cache), layer_idx):
                self.vision_k_pe_cache.append(torch.tensor([]))
                self.vision_compressed_kv_cache.append(torch.tensor([]))
                self.text_key_cache.append(torch.tensor([]))
                self.text_value_cache.append(torch.tensor([]))
            
            # 添加当前层
            self.vision_k_pe_cache.append(vision_k_pe if vision_k_pe is not None else torch.tensor([]))
            self.vision_compressed_kv_cache.append(vision_compressed_kv if vision_compressed_kv is not None else torch.tensor([]))
            self.text_key_cache.append(text_key_states if text_key_states is not None else torch.tensor([]))
            self.text_value_cache.append(text_value_states if text_value_states is not None else torch.tensor([]))
        else:
            # 更新现有层 - 与DynamicCache相同的concat逻辑
            if vision_k_pe is not None:
                if not self.vision_k_pe_cache[layer_idx].numel():  # 使用numel()检查，与DynamicCache一致
                    self.vision_k_pe_cache[layer_idx] = vision_k_pe
                else:
                    self.vision_k_pe_cache[layer_idx] = torch.cat([self.vision_k_pe_cache[layer_idx], vision_k_pe], dim=-2)
                    
            if vision_compressed_kv is not None:
                if not self.vision_compressed_kv_cache[layer_idx].numel():
                    self.vision_compressed_kv_cache[layer_idx] = vision_compressed_kv
                else:
                    self.vision_compressed_kv_cache[layer_idx] = torch.cat([self.vision_compressed_kv_cache[layer_idx], vision_compressed_kv], dim=-2)

            if text_key_states is not None:
                if not self.text_key_cache[layer_idx].numel():
                    self.text_key_cache[layer_idx] = text_key_states
                else:
                    self.text_key_cache[layer_idx] = torch.cat([self.text_key_cache[layer_idx], text_key_states], dim=-2)
                    
            if text_value_states is not None:
                if not self.text_value_cache[layer_idx].numel():
                    self.text_value_cache[layer_idx] = text_value_states
                else:
                    self.text_value_cache[layer_idx] = torch.cat([self.text_value_cache[layer_idx], text_value_states], dim=-2)

        return (
            self.vision_k_pe_cache[layer_idx],
            self.vision_compressed_kv_cache[layer_idx],
            self.text_key_cache[layer_idx],
            self.text_value_cache[layer_idx]
        )

    def get_seq_length(self, layer_idx: Optional[int] = 0) -> int:
        """返回缓存序列长度，与DynamicCache保持一致的检查逻辑"""
        is_empty_layer = (
            len(self.vision_k_pe_cache) == 0  # no cache in any layer
            or len(self.vision_k_pe_cache) <= layer_idx  # skipped `layer_idx` and hasn't run a layer with cache after it
            or (not self.vision_k_pe_cache[layer_idx].numel() and not self.text_key_cache[layer_idx].numel())  # the layer has no cache
        )
        if is_empty_layer:
            return 0
            
        vision_len = self.vision_k_pe_cache[layer_idx].shape[-2] if self.vision_k_pe_cache[layer_idx].numel() else 0
        text_len = self.text_key_cache[layer_idx].shape[-2] if self.text_key_cache[layer_idx].numel() else 0
        
        return vision_len + text_len

    def get_max_length(self) -> Optional[int]:
        """返回最大序列长度，MixedDynamicCache没有最大限制，与DynamicCache保持一致"""
        return None

    def get_max_cache_shape(self) -> Optional[int]:
        """返回最大缓存长度，与DynamicCache保持一致"""
        return None

    def get_usable_length(self, new_seq_length: int, layer_idx: Optional[int] = 0) -> int:
        """
        获取可用长度，与DynamicCache保持一致的实现
        由于MixedDynamicCache没有大小限制，所有缓存都是可用的
        """
        # Cache without size limit -> all cache is usable
        # Cache with size limit -> if the length cache plus the length of the new inputs is larger the maximum cache
        #   length, we will need to evict part of the cache (and thus not all cache is usable)
        max_length = self.get_max_length()
        previous_seq_length = self.get_seq_length(layer_idx)
        if max_length is not None and previous_seq_length + new_seq_length > max_length:
            return max_length - new_seq_length
        return previous_seq_length

    def to_legacy_cache(self) -> Tuple[Tuple[torch.Tensor], ...]:
        """转换为传统缓存格式 - 已废弃，不再使用"""
        raise NotImplementedError("Legacy cache format is no longer supported. Use MixedDynamicCache directly.")

    @classmethod
    def from_legacy_cache(cls, past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None) -> "MixedDynamicCache":
        """从传统缓存格式创建MixedDynamicCache - 已废弃，不再使用"""
        raise NotImplementedError("Legacy cache format is no longer supported. Use MixedDynamicCache directly.")

    def crop(self, max_length: int):
        """裁剪缓存到指定长度，与DynamicCache保持一致的逻辑"""
        # 处理负数长度
        if max_length < 0:
            max_length = self.get_seq_length() - abs(max_length)

        if self.get_seq_length() <= max_length:
            return

        self._seen_tokens = max_length
        for idx in range(len(self.vision_k_pe_cache)):
            # 只处理非空tensor，与DynamicCache保持一致
            if self.vision_k_pe_cache[idx].numel():
                self.vision_k_pe_cache[idx] = self.vision_k_pe_cache[idx][..., :max_length, :]
            if self.vision_compressed_kv_cache[idx].numel():
                self.vision_compressed_kv_cache[idx] = self.vision_compressed_kv_cache[idx][..., :max_length, :]
            if self.text_key_cache[idx].numel():
                self.text_key_cache[idx] = self.text_key_cache[idx][..., :max_length, :]
            if self.text_value_cache[idx].numel():
                self.text_value_cache[idx] = self.text_value_cache[idx][..., :max_length, :]

    def batch_split(self, full_batch_size: int, split_size: int) -> List["MixedDynamicCache"]:
        """按批次大小分割当前实例，与DynamicCache保持一致的逻辑"""
        out = []
        for i in range(0, full_batch_size, split_size):
            current_split = MixedDynamicCache()
            current_split._seen_tokens = self._seen_tokens
            current_split.vision_k_pe_cache = [tensor[i : i + split_size] for tensor in self.vision_k_pe_cache]
            current_split.vision_compressed_kv_cache = [tensor[i : i + split_size] for tensor in self.vision_compressed_kv_cache]
            current_split.text_key_cache = [tensor[i : i + split_size] for tensor in self.text_key_cache]
            current_split.text_value_cache = [tensor[i : i + split_size] for tensor in self.text_value_cache]
            out.append(current_split)
        return out

    @classmethod
    def from_batch_splits(cls, splits: List["MixedDynamicCache"]) -> "MixedDynamicCache":
        """与batch_split相反的操作，与DynamicCache保持一致的逻辑"""
        cache = cls()
        for idx in range(len(splits[0])):
            vision_k_pe_cache = [current.vision_k_pe_cache[idx] for current in splits if current.vision_k_pe_cache[idx].numel()]
            vision_compressed_cache = [current.vision_compressed_kv_cache[idx] for current in splits if current.vision_compressed_kv_cache[idx].numel()]
            text_key_cache = [current.text_key_cache[idx] for current in splits if current.text_key_cache[idx].numel()]
            text_value_cache = [current.text_value_cache[idx] for current in splits if current.text_value_cache[idx].numel()]
            
            # 合并非空缓存
            vision_k_pe = torch.cat(vision_k_pe_cache, dim=0) if vision_k_pe_cache else None
            vision_compressed = torch.cat(vision_compressed_cache, dim=0) if vision_compressed_cache else None
            text_key = torch.cat(text_key_cache, dim=0) if text_key_cache else None
            text_value = torch.cat(text_value_cache, dim=0) if text_value_cache else None
            
            if any([vision_k_pe, vision_compressed, text_key, text_value]):
                cache.update(vision_k_pe, vision_compressed, text_key, text_value, idx)
        return cache

    def batch_repeat_interleave(self, repeats: int):
        """在批次维度重复缓存，与DynamicCache保持一致"""
        for layer_idx in range(len(self)):
            if self.vision_k_pe_cache[layer_idx].numel():
                self.vision_k_pe_cache[layer_idx] = self.vision_k_pe_cache[layer_idx].repeat_interleave(repeats, dim=0)
            if self.vision_compressed_kv_cache[layer_idx].numel():
                self.vision_compressed_kv_cache[layer_idx] = self.vision_compressed_kv_cache[layer_idx].repeat_interleave(repeats, dim=0)
            if self.text_key_cache[layer_idx].numel():
                self.text_key_cache[layer_idx] = self.text_key_cache[layer_idx].repeat_interleave(repeats, dim=0)
            if self.text_value_cache[layer_idx].numel():
                self.text_value_cache[layer_idx] = self.text_value_cache[layer_idx].repeat_interleave(repeats, dim=0)

    def batch_select_indices(self, indices: torch.Tensor):
        """只保留批次维度中的指定索引，与DynamicCache保持一致"""
        for layer_idx in range(len(self)):
            if self.vision_k_pe_cache[layer_idx].numel():
                self.vision_k_pe_cache[layer_idx] = self.vision_k_pe_cache[layer_idx][indices, ...]
            if self.vision_compressed_kv_cache[layer_idx].numel():
                self.vision_compressed_kv_cache[layer_idx] = self.vision_compressed_kv_cache[layer_idx][indices, ...]
            if self.text_key_cache[layer_idx].numel():
                self.text_key_cache[layer_idx] = self.text_key_cache[layer_idx][indices, ...]
            if self.text_value_cache[layer_idx].numel():
                self.text_value_cache[layer_idx] = self.text_value_cache[layer_idx][indices, ...]

    # 保留一些MixedDynamicCache特有的辅助方法，但简化实现
    def get_vision_seq_length(self, layer_idx: Optional[int] = 0) -> int:
        """获取视觉token的序列长度"""
        if (len(self.vision_k_pe_cache) <= layer_idx or 
            not self.vision_k_pe_cache[layer_idx].numel()):
            return 0
        return self.vision_k_pe_cache[layer_idx].shape[-2]

    def get_text_seq_length(self, layer_idx: Optional[int] = 0) -> int:
        """获取文本token的序列长度"""
        if (len(self.text_key_cache) <= layer_idx or 
            not self.text_key_cache[layer_idx].numel()):
            return 0
        return self.text_key_cache[layer_idx].shape[-2]

    def has_vision_cache(self, layer_idx: Optional[int] = 0) -> bool:
        """检查是否有视觉缓存"""
        return (len(self.vision_k_pe_cache) > layer_idx and 
                self.vision_k_pe_cache[layer_idx].numel() > 0)

    def has_text_cache(self, layer_idx: Optional[int] = 0) -> bool:
        """检查是否有文本缓存"""
        return (len(self.text_key_cache) > layer_idx and 
                self.text_key_cache[layer_idx].numel() > 0)