qllava3_test.py

# coding=utf-8
# Copyright 2024 the HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""PyTorch Llava-NeXT model."""

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

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

from transformers.activations import ACT2FN
from transformers.generation import GenerationMixin
from transformers.image_processing_utils import select_best_resolution
from transformers.modeling_outputs import ModelOutput
from transformers.modeling_utils import PreTrainedModel
from transformers.utils import (
    add_start_docstrings,
    add_start_docstrings_to_model_forward,
    logging,
    replace_return_docstrings,
)
from transformers.models.auto import AutoModel, AutoModelForCausalLM
from transformers.models.llava_next.configuration_llava_next import LlavaNextConfig


logger = logging.get_logger(__name__)

_CONFIG_FOR_DOC = "LlavaNextConfig"
from pathlib import Path

def save_list_to_incremental_file(data_list, save_dir="/common/home/users/w/wzhao/vqclip/llava_next_tensors"):
    """
    将列表保存到指定目录，文件名按数字递增
    
    Args:
        data_list: 要保存的列表数据
        save_dir: 保存目录路径
    
    Returns:
        保存的文件路径
    """
    # 确保目录存在
    save_dir = Path(save_dir)
    save_dir.mkdir(parents=True, exist_ok=True)
    
    # 查找可用的文件名
    index = 1
    while True:
        file_path = save_dir / f"{index}.npy"
        if not file_path.exists():
            break
        index += 1
    
    # 将列表转换为numpy数组并保存
    np_array = np.array(data_list)
    np.save(str(file_path), np_array)
    
    return file_path

def get_anyres_image_grid_shape(image_size, grid_pinpoints, patch_size):
    """
    Calculate the shape of the image patch grid after the preprocessing for images of any resolution.

    Args:
        image_size (`tuple`):
            The size of the input image in the format (width, height).
        grid_pinpoints (`List`):
            A list containing possible resolutions. Each item in the list should be a tuple or list
            of the form `(height, width)`.
        patch_size (`int`):
            The size of each image patch.

    Returns:
        tuple: The shape of the image patch grid in the format (width, height).
    """
    if not isinstance(grid_pinpoints, list):
        raise TypeError("grid_pinpoints should be a list of tuples or lists")

    # ! VERY IMPORTANT if image_size is tensor, must convert to into tuple, otherwise it will cause wrong calculate
    if not isinstance(image_size, (list, tuple)):
        if not isinstance(image_size, (torch.Tensor, np.ndarray)):
            raise TypeError(
                f"image_size invalid type: {type(image_size)} not valid, should be either list, tuple, np.ndarray or tensor"
            )
        image_size = image_size.tolist()

    height, width = select_best_resolution(image_size, grid_pinpoints)
    return height // patch_size, width // patch_size


def image_size_to_num_patches(image_size, grid_pinpoints, patch_size: int):
    """
    Calculate the number of patches after the preprocessing for images of any resolution.

    Args:
        image_size (`torch.LongTensor` or `np.ndarray` or `Tuple[int, int]`):
            The size of the input image in the format (height, width). ?
        grid_pinpoints (`List`):
            A list containing possible resolutions. Each item in the list should be a tuple or list
            of the form `(height, width)`.
        patch_size (`int`):
            The size of each image patch.

    Returns:
        int: the number of patches
    """
    if not isinstance(grid_pinpoints, list):
        raise TypeError("grid_pinpoints should be a list of tuples or lists")

    # ! VERY IMPORTANT if image_size is tensor, must convert to into tuple, otherwise it will cause wrong calculate
    if not isinstance(image_size, (list, tuple)):
        if not isinstance(image_size, (torch.Tensor, np.ndarray)):
            raise TypeError(f"image_size invalid type {type(image_size)} with value {image_size}")
        image_size = image_size.tolist()

    best_resolution = select_best_resolution(image_size, grid_pinpoints)
    height, width = best_resolution
    num_patches = 0
    # consider change to ceil(height/patch_size)*ceil(width/patch_size) + 1
    for i in range(0, height, patch_size):
        for j in range(0, width, patch_size):
            num_patches += 1
    # add the base patch
    num_patches += 1
    return num_patches


def unpad_image(tensor, original_size):
    """
    Unpads a PyTorch tensor of a padded and resized image.

    Args:
        tensor (`torch.Tensor`):
            The image tensor, assumed to be of shape (num_channels, height, width).
        original_size (`tuple`):
            The original size of the image (height, width).

    Returns:
        `torch.Tensor`: The unpadded image tensor.
    """
    if not isinstance(original_size, (list, tuple)):
        if not isinstance(original_size, (torch.Tensor, np.ndarray)):
            raise TypeError(
                f"image_size invalid type: {type(original_size)} not valid, should be either list, tuple, np.ndarray or tensor"
            )
        original_size = original_size.tolist()
    original_height, original_width = original_size
    current_height, current_width = tensor.shape[1:]

    original_aspect_ratio = original_width / original_height
    current_aspect_ratio = current_width / current_height

    if original_aspect_ratio > current_aspect_ratio:
        scale_factor = current_width / original_width
        new_height = int(original_height * scale_factor)
        padding = (current_height - new_height) // 2
        unpadded_tensor = tensor[:, padding : current_height - padding, :]
    else:
        scale_factor = current_height / original_height
        new_width = int(original_width * scale_factor)
        padding = (current_width - new_width) // 2
        unpadded_tensor = tensor[:, :, padding : current_width - padding]

    return unpadded_tensor


@dataclass
class LlavaNextCausalLMOutputWithPast(ModelOutput):
    """
    Base class for LlavaNext causal language model (or autoregressive) outputs.

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

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

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

            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
            heads.
        image_hidden_states (`torch.FloatTensor`, *optional*):
            A `torch.FloatTensor` of size (batch_size * num_patches, num_images, sequence_length, hidden_size)`.
            image_hidden_states of the model produced by the vision encoder and after projecting the last hidden state.
    """

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

class VectorQuantizer(nn.Module):
    def __init__(self, num_embeddings: int, embedding_dim: int, commitment_cost: float = 0.25):
        super().__init__()
        self.num_embeddings = num_embeddings
        self.embedding_dim = embedding_dim
        self.commitment_cost = commitment_cost
        
        # Embedding table
        self.embedding = nn.Embedding(num_embeddings, embedding_dim)
        self.embedding.weight.data.uniform_(-1/num_embeddings, 1/num_embeddings)

    def forward(self, inputs):
        
        self.embedding.weight.data = self.embedding.weight.data.to(dtype=inputs.dtype)
        # Convert inputs from BCHW -> BHWC
        inputs = inputs.permute(0, 2, 1).contiguous()
        input_shape = inputs.shape
        
        # Flatten input
        flat_input = inputs.view(-1, self.embedding_dim)

        # Calculate distances
        distances = (torch.sum(flat_input**2, dim=1, keepdim=True) 
                    + torch.sum(self.embedding.weight**2, dim=1)
                    - 2 * torch.matmul(flat_input, self.embedding.weight.t()))
            
        # Encoding
        encoding_indices = torch.argmin(distances, dim=1).unsqueeze(1)
        encodings = torch.zeros(encoding_indices.shape[0], self.num_embeddings, device=inputs.device).to(inputs.dtype)
        encodings.scatter_(1, encoding_indices, 1)
        #self.embedding.weight = self.embedding.weight.to(input_type)
        # Quantize and unflatten
        #print(inputs.dtype)
        #print(self.embedding.weight.dtype)
        quantized = torch.matmul(encodings, self.embedding.weight).view(input_shape)
        
        # Loss
        e_latent_loss = torch.mean((quantized.detach() - inputs)**2)
        q_latent_loss = torch.mean((quantized - inputs.detach())**2)
        loss = q_latent_loss + self.commitment_cost * e_latent_loss
        print("this is q_latent_loss", q_latent_loss)
        print("This is e_latent_loss", self.commitment_cost * e_latent_loss)
        quantized = inputs + (quantized - inputs).detach()
        avg_probs = torch.mean(encodings, dim=0)
        perplexity = torch.exp(-torch.sum(avg_probs * torch.log(avg_probs + 1e-10)))

        # Convert quantized from BHWC -> BCHW
        return quantized.permute(0, 2, 1).contiguous(), loss, perplexity

# Copied from transformers.models.llava.modeling_llava.LlavaMultiModalProjector with Llava->LlavaNext
class LlavaNextMultiModalProjector(nn.Module):
    def __init__(self, config: LlavaNextConfig):
        super().__init__()

        self.linear_1 = nn.Linear(config.vision_config.hidden_size, config.text_config.hidden_size, bias=True)
        self.act = ACT2FN[config.projector_hidden_act]
        self.linear_2 = nn.Linear(config.text_config.hidden_size, config.text_config.hidden_size, bias=True)
        self.vq = VectorQuantizer(
                    num_embeddings=16000,  # codebook size
                    embedding_dim=config.text_config.hidden_size,  # dimension of each embedding vector
                    commitment_cost=0.5
                )
        self.vq_cls = VectorQuantizerCLS(
            num_embeddings=128,
            embedding_dim=4096,
            commitment_cost=0.25,
            use_cosine=True
        )
    def forward(self, image_features):
        cls_features = image_features[: , :1]
        cls_features = self.linear_1(cls_features)
        cls_features = self.act(cls_features)
        cls_features = self.linear_2(cls_features)
        cls_features = cls_features[:, 0:]
        cls_features = cls_features.mean(dim=0, keepdim=True).squeeze(0)
        #save_list_to_incremental_file(cls_features.cpu().detach().numpy())
        quantized, loss, perplexity, indices = self.vq_cls(cls_features)
        categories = self.vq_cls.get_category_from_index(indices)
        indices = indices.cpu().numpy()
        print(indices)
        print(categories)
        if categories[0] != 0:
            raise ValueError([indices, categories[0]]) 
        #save_list_to_incremental_file(save_list)
        # tensor(54)
        # ['porn']
        image_features = image_features[: , 1:]
        hidden_states = self.linear_1(image_features)
        hidden_states = self.act(hidden_states)
        hidden_states = self.linear_2(hidden_states)
        
        quantized_features, vq_loss, perplexity = self.vq(hidden_states)
        print(quantized_features.shape)
        return quantized_features, vq_loss


LLAVA_NEXT_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 ([`LlavaNextConfig`] or [`LlavaNextVisionConfig`]):
            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.
"""
class VectorQuantizerCLS(nn.Module):
    def __init__(self, num_embeddings: int = 64, embedding_dim: int = 4096, commitment_cost: float = 0.25, 
                 codebook_path: str = None, mapping_path: str = None, use_cosine: bool = True, 
                 randomize_indices: bool = True):
        super().__init__()
        self.num_embeddings = num_embeddings
        self.embedding_dim = embedding_dim
        self.commitment_cost = commitment_cost
        self.use_cosine = use_cosine
        
        # Embedding table
        self.embedding = nn.Embedding(num_embeddings, embedding_dim)
        self.embedding.weight.data.uniform_(-1/num_embeddings, 1/num_embeddings)
        
        # 初始化合适大小的buffer，避免加载时大小不匹配
        self.register_buffer('_category_mapping_indices', torch.zeros(num_embeddings, dtype=torch.long))
        self.register_buffer('_category_mapping_names', torch.zeros(num_embeddings, dtype=torch.long))
        
        # 非持久化属性
        self.center_to_category = None
        
        # 加载预先计算的codebook
        if codebook_path is not None and mapping_path is not None:
            self.load_codebook(codebook_path, mapping_path, randomize_indices)
    
    def load_codebook(self, codebook_path, mapping_path, randomize_indices=True):
        """加载预计算的codebook和类别映射，并可选择随机化索引"""
        try:
            # 加载codebook
            print(f"Loading codebook from {codebook_path}")
            centers = np.load(codebook_path)
            print(f"Loaded codebook with shape: {centers.shape}")
            
            # 加载类别映射
            print(f"Loading category mappings from {mapping_path}")
            with open(mapping_path, 'rb') as f:
                mappings = pickle.load(f)
            
            # 将文本类别映射为数字
            category_mapping_text = mappings['category_mapping']
            classes = {'neutral':0, 'porn':1, 'gun':2, 'cigarette':3, 'alcohol':4, 'knife':5, 'blood':6, 'insulting_gesture':7}
            
            # 转换为数字映射
            center_category_mapping = {}
            for i, category_text in enumerate(category_mapping_text):
                center_category_mapping[i] = classes.get(category_text, 0)  # 默认为neutral(0)
            
            print(f"Loaded {len(center_category_mapping)} category mappings")
            
            # 准备数据
            actual_centers = centers.shape[0]
            print(f"Actual centers: {actual_centers}")
            
            # 更新num_embeddings为实际中心点数量
            self.num_embeddings = actual_centers
            print(f"Setting num_embeddings to {self.num_embeddings}")
            
            # 如果需要随机化索引，创建随机排列
            if randomize_indices:
                print("Randomizing codebook indices to prevent category clustering")
                # 创建随机排列
                permutation = list(range(actual_centers))
                random.shuffle(permutation)
                inverse_permutation = {v: k for k, v in enumerate(permutation)}
                
                # 应用随机排列到中心点和类别映射
                permuted_centers = np.zeros_like(centers)
                permuted_categories = {}
                
                for new_idx, old_idx in enumerate(permutation):
                    permuted_centers[new_idx] = centers[old_idx]
                    if old_idx < len(center_category_mapping):
                        permuted_categories[new_idx] = center_category_mapping[old_idx]
                
                # 使用随机化后的数据
                centers = permuted_centers
                self.center_to_category = permuted_categories
                
                # 打印一些随机化后的映射示例
                print("Sample randomized mappings:")
                for i in range(min(5, len(self.center_to_category))):
                    print(f"  New index {i}: {self.center_to_category[i]}")
            else:
                # 不随机化，直接使用原始映射
                self.center_to_category = {i: center_category_mapping[i] 
                                          for i in range(min(actual_centers, len(center_category_mapping)))}
            
            # 验证类别映射是否完整
            for i in range(self.num_embeddings):
                if i not in self.center_to_category:
                    print(f"Warning: No category mapping for center {i}, setting to 0")
                    self.center_to_category[i] = 0  # 用0代替"unknown"
            
            # 创建embedding数据并更新
            embedding_data = torch.tensor(centers, dtype=torch.float32)
            
            # 重新创建embedding层以匹配实际大小
            self.embedding = nn.Embedding(self.num_embeddings, self.embedding_dim)
            self.embedding.weight.data.copy_(embedding_data)
            
            # 重新注册buffer以匹配新的大小
            self.register_buffer('_category_mapping_indices', torch.zeros(self.num_embeddings, dtype=torch.long))
            self.register_buffer('_category_mapping_names', torch.zeros(self.num_embeddings, dtype=torch.long))
            
            # 将类别映射存储到buffer中（用于state_dict）
            self._store_category_mapping()
            
            print(f"Successfully loaded codebook with {self.num_embeddings} entries")
            
            # 分析类别分布
            category_counts = {}
            for category in self.center_to_category.values():
                if category in category_counts:
                    category_counts[category] += 1
                else:
                    category_counts[category] = 1
            
            print("Category distribution in codebook:")
            for category, count in sorted(category_counts.items()):
                print(f"  {category}: {count} centers")
            
            return True
            
        except Exception as e:
            print(f"Error loading codebook: {e}")
            import traceback
            traceback.print_exc()
            print("Using random initialization instead")
            return False
    
    def _store_category_mapping(self):
        """将类别映射存储到模型的buffer中，以便在state_dict中保存"""
        if not self.center_to_category:
            warnings.warn("No category mapping to store")
            return
        
        # 获取所有类别ID
        all_categories = sorted(set(self.center_to_category.values()))
        
        # 创建索引和对应类别ID的映射
        indices = list(self.center_to_category.keys())
        category_ids = [self.center_to_category[idx] for idx in indices]
        
        # 确保indices数组长度与buffer大小一致
        if len(indices) != self._category_mapping_indices.size(0):
            # 重新注册buffer以匹配大小
            self.register_buffer('_category_mapping_indices', torch.zeros(len(indices), dtype=torch.long))
            self.register_buffer('_category_mapping_names', torch.zeros(len(indices), dtype=torch.long))
        
        # 存储到buffer中
        self._category_mapping_indices.copy_(torch.tensor(indices, dtype=torch.long))
        self._category_mapping_names.copy_(torch.tensor(category_ids, dtype=torch.long))
        
        print(f"Stored category mapping with {len(indices)} entries and {len(all_categories)} unique categories")
    
    def _load_category_mapping(self):
        """从模型的buffer恢复类别映射"""
        if not hasattr(self, '_category_mapping_indices') or self._category_mapping_indices.numel() == 0:
            warnings.warn("No stored category mapping found")
            return {}
        
        # 重建类别映射字典
        indices = self._category_mapping_indices.tolist()
        category_ids = self._category_mapping_names.tolist()
        
        mapping = {}
        for idx, cat_id in zip(indices, category_ids):
            mapping[idx] = cat_id
        
        return mapping
    
    def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs):
        """自定义state_dict加载方法，处理buffer大小不匹配的问题"""
        # 检查并调整buffer大小，以匹配加载的state_dict
        indices_key = prefix + '_category_mapping_indices'
        names_key = prefix + '_category_mapping_names'
        
        if indices_key in state_dict and names_key in state_dict:
            indices_size = state_dict[indices_key].size()
            names_size = state_dict[names_key].size()
            
            # 重新注册buffer以匹配加载的大小
            if hasattr(self, '_category_mapping_indices') and self._category_mapping_indices.size() != indices_size:
                self.register_buffer('_category_mapping_indices', torch.zeros(indices_size, dtype=torch.long))
            
            if hasattr(self, '_category_mapping_names') and self._category_mapping_names.size() != names_size:
                self.register_buffer('_category_mapping_names', torch.zeros(names_size, dtype=torch.long))
        
        # 调用父类方法加载常规参数
        super()._load_from_state_dict(state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs)
        
        # 在加载完成后重建类别映射
        self.center_to_category = self._load_category_mapping()
        
        # 更新num_embeddings以匹配加载的模型
        if hasattr(self, 'embedding') and hasattr(self.embedding, 'weight'):
            self.num_embeddings = self.embedding.weight.size(0)
    
    def forward(self, inputs):
        """
        前向传播，专门处理(1, 4096)形状的输入
        
        Args:
            inputs: 形状为(1, 4096)的特征向量
            
        Returns:
            quantized: 量化后的特征向量
            loss: 承诺损失
            perplexity: 困惑度
            encoding_indices: 编码索引
        """
        # 验证输入形状
        if inputs.shape != (1, 4096):
            raise ValueError(f"Expected input shape (1, 4096), got {inputs.shape}")
        
        # 确保embedding权重与输入使用相同的数据类型
        self.embedding.weight.data = self.embedding.weight.data.to(dtype=inputs.dtype)
        
        # 直接使用输入，不需要形状转换
        flat_input = inputs
        
        # 计算与codebook中各向量的距离
        if self.use_cosine:
            # 归一化向量进行余弦相似度计算
            normalized_input = F.normalize(flat_input, p=2, dim=1)
            normalized_weights = F.normalize(self.embedding.weight, p=2, dim=1)
            
            # 计算余弦相似度
            cosine_sim = torch.matmul(normalized_input, normalized_weights.t())
            
            # 将相似度转换为距离（最大相似度对应最小距离）
            distances = 1 - cosine_sim
        else:
            # 使用欧氏距离
            distances = (torch.sum(flat_input**2, dim=1, keepdim=True) 
                        + torch.sum(self.embedding.weight**2, dim=1)
                        - 2 * torch.matmul(flat_input, self.embedding.weight.t()))
        
        # 找到最近的编码向量索引
        encoding_indices = torch.argmin(distances, dim=1).unsqueeze(1)
        
        # 创建one-hot编码
        encodings = torch.zeros(encoding_indices.shape[0], self.num_embeddings, device=inputs.device).to(inputs.dtype)
        encodings.scatter_(1, encoding_indices, 1)
        
        # 量化
        quantized = torch.matmul(encodings, self.embedding.weight)
        
        # 计算损失
        e_latent_loss = torch.mean((quantized.detach() - flat_input)**2)
        q_latent_loss = torch.mean((quantized - flat_input.detach())**2)
        loss = q_latent_loss + self.commitment_cost * e_latent_loss
        
        print("this is q_latent_loss", q_latent_loss)
        print("This is e_latent_loss", self.commitment_cost * e_latent_loss)
        
        # Straight-through estimator
        quantized = flat_input + (quantized - flat_input).detach()
        
        # 计算perplexity
        avg_probs = torch.mean(encodings, dim=0)
        perplexity = torch.exp(-torch.sum(avg_probs * torch.log(avg_probs + 1e-10)))
        
        # 返回量化后的向量、损失、困惑度和索引
        return quantized, loss, perplexity, encoding_indices.squeeze()
    
    def encode(self, inputs):
        """
        仅执行编码过程，返回索引
        
        Args:
            inputs: 形状为(1, 4096)的特征向量
            
        Returns:
            编码索引
        """
        # 验证输入形状
        if inputs.shape != (1, 4096):
            raise ValueError(f"Expected input shape (1, 4096), got {inputs.shape}")
        
        with torch.no_grad():
            # 计算距离
            if self.use_cosine:
                normalized_input = F.normalize(inputs, p=2, dim=1)
                normalized_weights = F.normalize(self.embedding.weight, p=2, dim=1)
                cosine_sim = torch.matmul(normalized_input, normalized_weights.t())
                distances = 1 - cosine_sim
            else:
                distances = (torch.sum(inputs**2, dim=1, keepdim=True) 
                            + torch.sum(self.embedding.weight**2, dim=1)
                            - 2 * torch.matmul(inputs, self.embedding.weight.t()))
            
            # 找到最近的编码向量索引
            encoding_indices = torch.argmin(distances, dim=1)
            
            return encoding_indices
    
    def get_category_from_index(self, indices):
        """
        根据索引获取对应的类别编号
        
        Args:
            indices: 编码索引
            
        Returns:
            类别编号列表
        """
        # 如果没有类别映射，尝试从buffer恢复
        if self.center_to_category is None:
            self.center_to_category = self._load_category_mapping()
            
        if not self.center_to_category:
            return [0] * indices.numel()  # 使用0(neutral)代替"unknown"
        
        # 将索引张量转为NumPy数组
        indices_np = indices.cpu().numpy().flatten()
        
        # 获取类别
        categories = []
        for idx in indices_np:
            idx_int = int(idx)
            category = self.center_to_category.get(idx_int, 0)  # 默认为0(neutral)
            categories.append(category)
        
        return categories
    
    def classify(self, inputs):
        """
        对输入特征进行分类，返回类别编号和索引
        
        Args:
            inputs: 形状为(1, 4096)的特征向量
            
        Returns:
            categories: 预测的类别编号
            indices: 编码索引
        """
        # 验证输入形状
        if inputs.shape != (1, 4096):
            raise ValueError(f"Expected input shape (1, 4096), got {inputs.shape}")
        
        indices = self.encode(inputs)
        categories = self.get_category_from_index(indices)
        return categories, indices


@add_start_docstrings(
    "The bare LLaMA Model outputting raw hidden-states without any specific head on top.",
    LLAVA_NEXT_START_DOCSTRING,
)
# Copied from transformers.models.llava.modeling_llava.LlavaPreTrainedModel with Llava->LlavaNext,llava->llava_next
class LlavaNextPreTrainedModel(PreTrainedModel):
    config_class = LlavaNextConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["LlavaNextVisionAttention"]
    _skip_keys_device_placement = "past_key_values"
    _supports_cache_class = True
    _supports_flash_attn_2 = True
    _supports_sdpa = True

    def _init_weights(self, module):
        # important: this ported version of LlavaNext isn't meant for training from scratch - only
        # inference and fine-tuning - so the proper init weights code has been removed - the original codebase
        # https://github.com/haotian-liu/LLaVA/tree/main/llava_next should serve for that purpose
        std = (
            self.config.initializer_range
            if hasattr(self.config, "initializer_range")
            else self.config.text_config.initializer_range
        )

        if hasattr(module, "class_embedding"):
            module.class_embedding.data.normal_(mean=0.0, std=std)

        if isinstance(module, (nn.Linear, nn.Conv2d)):
            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_()


LLAVA_NEXT_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)
        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)):
            The tensors corresponding to the input images. Pixel values can be obtained using
            [`AutoImageProcessor`]. See [`LlavaNextImageProcessor.__call__`] for details. [`LlavaProcessor`] uses
            [`LlavaNextImageProcessor`] for processing images.
        image_sizes (`torch.LongTensor` of shape `(batch_size, 2)`, *optional*):
            The sizes of the images in the batch, being (height, width) for each image.
        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:

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

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

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

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

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

            - 1 indicates the head is **not masked**,
            - 0 indicates the head is **masked**.
        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
            config.n_positions - 1]`. [What are position IDs?](../glossary#position-ids)
        past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
            Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
            `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
            `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.

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

            If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
            don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
            `decoder_input_ids` of shape `(batch_size, sequence_length)`.
        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
            model's internal embedding lookup matrix.
        vision_feature_layer (`int`, *optional*, defaults to -2):
            The index of the layer to select the vision feature.
        vision_feature_select_strategy (`str`, *optional*, defaults to `"default"`):
            The feature selection strategy used to select the vision feature from the vision backbone.
            Can be one of `"default"` or `"full"`. If `"default"`, the CLS token is removed from the vision features.
            If `"full"`, the full vision features are used.
        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.
        cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
            Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`,
            this tensor is not affected by padding. It is used to update the cache in the correct position and to infer
            the complete sequence length.
"""


@add_start_docstrings(
    """The LLAVA-NeXT model which consists of a vision backbone and a language model.""",
    LLAVA_NEXT_START_DOCSTRING,
)
class LlavaNextForConditionalGeneration(LlavaNextPreTrainedModel, GenerationMixin):
    def __init__(self, config: LlavaNextConfig):
        super().__init__(config)
        self.vision_tower = AutoModel.from_config(config.vision_config)

        self.multi_modal_projector = LlavaNextMultiModalProjector(config)
        embed_std = 1 / math.sqrt(config.text_config.hidden_size)
        self.image_newline = nn.Parameter(torch.randn(config.text_config.hidden_size, dtype=self.dtype) * embed_std)

        self.vocab_size = config.text_config.vocab_size
        self.language_model = AutoModelForCausalLM.from_config(config.text_config)
        self.pad_token_id = self.config.pad_token_id if self.config.pad_token_id is not None else -1
        self._padding_side = "left"  # set it to left by default, user can use setter to change padding_sides
        self.post_init()

    @property
    def padding_side(self):
        return self._padding_side

    @padding_side.setter
    def padding_side(self, padding_side: str):
        if padding_side not in ["left", "right"]:
            raise ValueError(f"{padding_side} is not `left` or `right`.")
        self._padding_side = padding_side

    # Copied from transformers.models.llava.modeling_llava.LlavaForConditionalGeneration.get_input_embeddings
    def get_input_embeddings(self):
        return self.language_model.get_input_embeddings()

    # Copied from transformers.models.llava.modeling_llava.LlavaForConditionalGeneration.set_input_embeddings
    def set_input_embeddings(self, value):
        self.language_model.set_input_embeddings(value)

    # Copied from transformers.models.llava.modeling_llava.LlavaForConditionalGeneration.get_output_embeddings
    def get_output_embeddings(self):
        return self.language_model.get_output_embeddings()

    # Copied from transformers.models.llava.modeling_llava.LlavaForConditionalGeneration.set_output_embeddings
    def set_output_embeddings(self, new_embeddings):
        self.language_model.set_output_embeddings(new_embeddings)

    # Copied from transformers.models.llava.modeling_llava.LlavaForConditionalGeneration.set_decoder
    def set_decoder(self, decoder):
        self.language_model.set_decoder(decoder)

    # Copied from transformers.models.llava.modeling_llava.LlavaForConditionalGeneration.get_decoder
    def get_decoder(self):
        return self.language_model.get_decoder()

    # Copied from transformers.models.llava.modeling_llava.LlavaForConditionalGeneration.tie_weights
    def tie_weights(self):
        return self.language_model.tie_weights()

    # Copied from transformers.models.llava.modeling_llava.LlavaForConditionalGeneration.resize_token_embeddings
    def resize_token_embeddings(self, new_num_tokens: Optional[int] = None, pad_to_multiple_of=None) -> nn.Embedding:
        model_embeds = self.language_model.resize_token_embeddings(new_num_tokens, pad_to_multiple_of)
        # update vocab size
        self.config.text_config.vocab_size = model_embeds.num_embeddings
        self.vocab_size = model_embeds.num_embeddings
        return model_embeds

    def _merge_input_ids_with_image_features(
        self,
        image_features,
        feature_lens,
        inputs_embeds,
        input_ids,
        attention_mask,
        position_ids=None,
        labels=None,
        image_token_index=None,
        ignore_index=-100,
    ):
        """
        Merge input_ids with with image features into final embeddings

        Args:
            image_features (`torch.Tensor` of shape `(all_feature_lens, embed_dim)`):
                All vision vectors of all images in the batch
            feature_lens (`torch.LongTensor` of shape `(num_images)`):
                The length of visual embeddings of each image as stacked in `image_features`
            inputs_embeds (`torch.Tensor` of shape `(batch_size, sequence_length, embed_dim)`):
                Token embeddings before merging with visual embeddings
            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
                Input_ids of tokens, possibly filled with image token
            attention_mask (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
                Mask to avoid performing attention on padding token indices.
            position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
                Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
                config.n_positions - 1]`.
            labels (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*)
                :abels need to be recalculated to support training (if provided)
            image_token_index (`int`, *optional*)
                Token id used to indicate the special "image" token. Defaults to `config.image_token_index`
            ignore_index (`int`, *optional*)
                Value that is used to pad `labels` and will be ignored when calculated loss. Default: -100.
        Returns:
            final_embedding, final_attention_mask, position_ids, final_labels

        Explanation:
            each image has variable length embeddings, with length specified by feature_lens
            image_features is concatenation of all visual embed vectors
            task: fill each <image> with the correct number of visual embeddings
            Example:
                X (5 patches), Y (3 patches), Z (8)
                X, Y are in the same sequence (in-context learning)
            if right padding
                input_ids: [
                    a b c d e f X g h i j k Y l m
                    o p q r Z s t u v _ _ _ _ _ _
                ]
                input_ids should be: [
                    a b c d e f X X X X X g h i j k Y Y Y l m
                    o p q r Z Z Z Z Z Z Z Z s t u v _ _ _ _ _
                ]
                labels should be: [
                    a b c d e f _ _ _ _ _ g h i j k _ _ _ l m
                    o p q r _ _ _ _ _ _ _ _ s t u v _ _ _ _ _
                ]
            elif left padding
                input_ids: [
                    a b c d e f X g h i j k Y l m
                    _ _ _ _ _ _ o p q r Z s t u v
                ]
                input_ids should be: [
                    a b c d e f X X X X X g h i j k Y Y Y l m
                    _ _ _ _ _ o p q r Z Z Z Z Z Z Z Z s t u v
                ]
                labels should be: [
                    a b c d e f _ _ _ _ _ g h i j k _ _ _ l m
                    _ _ _ _ _ o p q r _ _ _ _ _ _ _ _ s t u v
                ]
            Edge cases:
                * If tokens are same but image token sizes are different, then cannot infer left or right padding
                ```python
                cat_img = Image.open(requests.get("http://images.cocodataset.org/val2017/000000039769.jpg", stream=True).raw)
                chart_img = Image.open(requests.get("https://github.com/haotian-liu/LLaVA/blob/1a91fc274d7c35a9b50b3cb29c4247ae5837ce39/images/llava_v1_5_radar.jpg?raw=true", stream=True).raw)
                prompts = [
                    "[INST] <image>\nWhat is shown in this image? [/INST]",
                    "[INST] <image>\nWhat is shown in this image? [/INST]",
                ]
                inputs = processor(prompts, [chart_img, cat_img], return_tensors='pt', padding=True).to("cuda")
                    chart_img has 2634 tokens, while cat_img has 2340 tokens
                ```

                input_ids: [
                    a b c d X g h
                    i j Y k l m n
                ]
                where X is 3 tokens while Y is 5, this mean after merge
                if left-padding (batched generation)
                    input_ids should be: [
                        _ _ a b c d X X X g h
                        i j Y Y Y Y Y k l m n
                    ]
                elif (right padding) (training)
                    input_ids should be: [
                        a b c d X X X g h _ _
                        i j Y Y Y Y Y k l m n
                    ]
        """
        image_token_index = image_token_index if image_token_index is not None else self.config.image_token_index
        ignore_index = ignore_index if ignore_index is not None else self.config.ignore_index

        if self.training and self.padding_side == "left":
            logger.warning_once(
                "Padding side is set to 'left' but the model is in training mode. For training "
                "it is recommended to set `model.padding_side='right' and `processor.tokenizer.padding_side='right'`. "
                "If that's intended, ignore this warning"
            )
        if not self.training and self.padding_side == "right":
            logger.warning_once(
                "Padding side is set to 'right' but the model is in inference mode. For correct "
                "generation results, please set `model.padding_side='left'` and `processor.tokenizer.padding_side='left'`. "
                "If that's intended, ignore this warning"
            )

        with torch.no_grad():
            # ! in llava 1.6, number of patches is variable
            num_images = feature_lens.size(0)
            num_image_features, embed_dim = image_features.shape
            if feature_lens.sum() != num_image_features:
                raise ValueError(f"{feature_lens=} / {feature_lens.sum()} != {image_features.shape=}")
            batch_size = input_ids.shape[0]
            _left_padding = torch.any(attention_mask[:, 0] == 0)
            _right_padding = torch.any(attention_mask[:, -1] == 0)

            left_padding = self.padding_side == "left"
            if batch_size > 1:
                if _left_padding and _right_padding:
                    raise ValueError(f"both side of attention_mask has zero, invalid. {attention_mask}")
                elif _right_padding and left_padding:
                    left_padding = False
                elif _left_padding and not left_padding:
                    left_padding = True

            # Whether to turn off right padding
            # 1. Create a mask to know where special image tokens are
            special_image_token_mask = input_ids == image_token_index
            # special_image_token_mask: [bsz, seqlen]
            num_special_image_tokens = torch.sum(special_image_token_mask, dim=-1)
            # num_special_image_tokens: [bsz]
            # Reserve for padding of num_images
            total_num_special_image_tokens = torch.sum(special_image_token_mask)
            if total_num_special_image_tokens != num_images:
                raise ValueError(
                    f"Number of image tokens in input_ids ({total_num_special_image_tokens}) different from num_images ({num_images})."
                )
            # Compute the maximum embed dimension
            # max_image_feature_lens is max_feature_lens per batch
            feature_lens = feature_lens.to(input_ids.device)
            feature_lens_batch = feature_lens.split(num_special_image_tokens.tolist(), dim=0)
            feature_lens_batch_sum = torch.tensor([x.sum() for x in feature_lens_batch], device=input_ids.device)
            embed_sequence_lengths = (
                (attention_mask == 1).long().sum(-1) - num_special_image_tokens + feature_lens_batch_sum
            )
            max_embed_dim = embed_sequence_lengths.max()

            batch_indices, non_image_indices = torch.where((input_ids != image_token_index) & (attention_mask == 1))
            # 2. Compute the positions where text should be written
            # Calculate new positions for text tokens in merged image-text sequence.
            # `special_image_token_mask` identifies image tokens. Each image token will be replaced by `nb_text_tokens_per_images` text tokens.
            # `torch.cumsum` computes how each image token shifts subsequent text token positions.
            # - 1 to adjust for zero-based indexing, as `cumsum` inherently increases indices by one.
            # ! instead of special_image_token_mask * (num_image_patches - 1)
            #   special_image_token_mask * (num_feature_len - 1)
            special_image_token_mask = special_image_token_mask.long()
            special_image_token_mask[special_image_token_mask == 1] = feature_lens - 1
            new_token_positions = torch.cumsum((special_image_token_mask + 1), -1) - 1
            if left_padding:
                # shift right token positions so that they are ending at the same number
                # the below here was incorrect? new_token_positions += new_token_positions[:, -1].max() - new_token_positions[:, -1:]
                new_token_positions += max_embed_dim - 1 - new_token_positions[:, -1:]

            text_to_overwrite = new_token_positions[batch_indices, non_image_indices]

        # 3. Create the full embedding, already padded to the maximum position
        final_embedding = torch.zeros(
            batch_size, max_embed_dim, embed_dim, dtype=inputs_embeds.dtype, device=inputs_embeds.device
        )
        final_attention_mask = torch.zeros(
            batch_size, max_embed_dim, dtype=attention_mask.dtype, device=inputs_embeds.device
        )
        final_input_ids = torch.full(
            (batch_size, max_embed_dim), self.pad_token_id, dtype=input_ids.dtype, device=inputs_embeds.device
        )
        # In case the Vision model or the Language model has been offloaded to CPU, we need to manually
        # set the corresponding tensors into their correct target device.
        target_device = inputs_embeds.device
        batch_indices, non_image_indices, text_to_overwrite = (
            batch_indices.to(target_device),
            non_image_indices.to(target_device),
            text_to_overwrite.to(target_device),
        )
        attention_mask = attention_mask.to(target_device)
        input_ids = input_ids.to(target_device)

        # 4. Fill the embeddings based on the mask. If we have ["hey" "<image>", "how", "are"]
        # we need to index copy on [0, 577, 578, 579] for the text and [1:576] for the image features
        final_embedding[batch_indices, text_to_overwrite] = inputs_embeds[batch_indices, non_image_indices]
        final_attention_mask[batch_indices, text_to_overwrite] = attention_mask[batch_indices, non_image_indices]
        final_input_ids[batch_indices, text_to_overwrite] = input_ids[batch_indices, non_image_indices]
        final_labels = None
        if labels is not None:
            labels = labels.to(target_device)
            final_labels = torch.full_like(final_attention_mask, ignore_index).to(torch.long)
            final_labels[batch_indices, text_to_overwrite] = labels[batch_indices, non_image_indices]

        # 5. Fill the embeddings corresponding to the images. Anything that is not `text_positions` needs filling (#29835)
        with torch.no_grad():
            image_to_overwrite = torch.full(
                (batch_size, max_embed_dim), True, dtype=torch.bool, device=inputs_embeds.device
            )
            image_to_overwrite[batch_indices, text_to_overwrite] = False
            embed_indices = torch.arange(max_embed_dim).unsqueeze(0).to(target_device)
            embed_indices = embed_indices.expand(batch_size, max_embed_dim)
            embed_seq_lens = embed_sequence_lengths[:, None].to(target_device)

            if left_padding:
                # exclude padding on the left
                max_embed_dim = max_embed_dim.to(target_device)
                val = (max_embed_dim - embed_indices) <= embed_seq_lens
            else:
                # exclude padding on the right
                val = embed_indices < embed_seq_lens
            image_to_overwrite &= val

            if image_to_overwrite.sum() != num_image_features:
                raise ValueError(
                    f"{image_to_overwrite.sum()=} != {num_image_features=} The input provided to the model are wrong. "
                    f"The number of image tokens is {torch.sum(special_image_token_mask)} while"
                    f" the number of image given to the model is {num_images}. "
                    f"This prevents correct indexing and breaks batch generation."
                )
        final_embedding[image_to_overwrite] = image_features.contiguous().reshape(-1, embed_dim).to(target_device)
        final_attention_mask |= image_to_overwrite
        position_ids = (final_attention_mask.cumsum(-1) - 1).masked_fill_((final_attention_mask == 0), 1)

        return final_embedding, final_attention_mask, position_ids, final_labels, final_input_ids

    def pack_image_features(self, image_features, image_sizes, vision_feature_select_strategy, image_newline=None):
        """
        Reshape, unpad and then pack each image_feature into a single image_features tensor containing all visual vectors.

        Args:
            image_features (`List[torch.Tensor]` of length num_images, each of shape `(num_patches, image_length, embed_dim)`)
                List of image feature tensor, each contains all the visual feature of all patches.
            image_sizes (`torch.Tensor` of shape `(num_images, 2)`)
                Actual image size of each images (H, W).
            vision_feature_select_strategy (`str`)
                The feature selection strategy used to select the vision feature from the vision backbone.
            image_newline (`torch.Tensor` of shape `(embed_dim)`)
                New line embedding vector.
        Returns:
            image_features (`torch.Tensor` of shape `(all_feat_len, embed_dim)`)
            feature_lens (`List[int]`)
                token length of each image in image_features
        """
        new_image_features = []
        feature_lens = []
        for image_idx, image_feature in enumerate(image_features):
            if image_feature.shape[0] > 1:
                base_image_feature = image_feature[0]
                image_feature = image_feature[1:]
                height = width = self.config.vision_config.image_size // self.config.vision_config.patch_size

                if vision_feature_select_strategy == "default":
                    expected_num_patches = height * width
                elif vision_feature_select_strategy == "full":
                    expected_num_patches = height * width + 1
                if expected_num_patches != base_image_feature.shape[0]:
                    raise ValueError("The number of patches is not consistent with the image size.")

                num_patch_height, num_patch_width = get_anyres_image_grid_shape(
                    image_sizes[image_idx],
                    self.config.image_grid_pinpoints,
                    self.config.vision_config.image_size,
                )
                image_feature = image_feature.view(num_patch_height, num_patch_width, height, width, -1)
                image_feature = image_feature.permute(4, 0, 2, 1, 3).contiguous()
                image_feature = image_feature.flatten(1, 2).flatten(2, 3)
                image_feature = unpad_image(image_feature, image_sizes[image_idx])
                if image_newline is not None:
                    image_feature = torch.cat(
                        (
                            image_feature,
                            image_newline[:, None, None].expand(*image_feature.shape[:-1], 1).to(image_feature.dtype),
                        ),
                        dim=-1,
                    )
                image_feature = image_feature.flatten(1, 2).transpose(0, 1)
                image_feature = torch.cat((base_image_feature, image_feature), dim=0)
            else:
                image_feature = image_feature[0]
                if image_newline is not None:
                    image_feature = torch.cat((image_feature, image_newline[None].to(image_feature)), dim=0)
            new_image_features.append(image_feature)
            feature_lens.append(image_feature.size(0))
        image_features = torch.cat(new_image_features, dim=0)
        feature_lens = torch.tensor(feature_lens, dtype=torch.long, device=image_features.device)
        return image_features, feature_lens

    def get_image_features(
        self,
        pixel_values: torch.FloatTensor,
        image_sizes: torch.Tensor,
        vision_feature_layer: int,
        vision_feature_select_strategy: str,
    ):
        """
        Obtains image last hidden states from the vision tower and apply multimodal projection.

        Args:
            pixel_values (`torch.FloatTensor]` of shape `(batch_size, num_patches, channels, height, width)`)
               The tensors corresponding to the input images.
            image_sizes (`torch.Tensor` of shape `(num_images, 2)`)
                Actual image size of each images (H, W).
            vision_feature_layer (`int`):
                The index of the layer to select the vision feature.
            vision_feature_select_strategy (`str`):
                The feature selection strategy used to select the vision feature from the vision backbone.
                Can be one of `"default"` or `"full"`
        Returns:
            image_features (List[`torch.Tensor`]): List of image feature tensor, each contains all the visual feature of all patches
            and are of shape `(num_patches, image_length, embed_dim)`).
        """
        # ! infer image_num_patches from image_sizes
        image_num_patches = [
            image_size_to_num_patches(
                image_size=imsize,
                grid_pinpoints=self.config.image_grid_pinpoints,
                patch_size=self.config.vision_config.image_size,
            )
            for imsize in image_sizes
        ]
        if pixel_values.dim() == 5:
            # stacked if input is (batch_size, num_patches, num_channels, height, width)
            _pixel_values_list = [pix_val[:num_patch] for pix_val, num_patch in zip(pixel_values, image_num_patches)]
            pixel_values = torch.cat(_pixel_values_list, dim=0)
        elif pixel_values.dim() != 4:
            # otherwise has to be stacked from list of (num_patches, num_channels, height, width)
            raise ValueError(f"pixel_values of shape {pixel_values.shape}, expect to be of 4 or 5 dimensions")

        image_features = self.vision_tower(pixel_values, output_hidden_states=True)
        selected_image_feature = image_features.hidden_states[vision_feature_layer]
        if vision_feature_select_strategy == "default":
        #     selected_image_feature = selected_image_feature[:, 1:]
        # elif vision_feature_select_strategy == "full":
            selected_image_feature = selected_image_feature
        image_features, vq_loss = self.multi_modal_projector(selected_image_feature)
        image_features = torch.split(image_features, image_num_patches, dim=0)
        

        return image_features, vq_loss

    @add_start_docstrings_to_model_forward(LLAVA_NEXT_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=LlavaNextCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
    def forward(
        self,
        input_ids: torch.LongTensor = None,
        pixel_values: torch.FloatTensor = None,
        image_sizes: Optional[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,
        vision_feature_layer: Optional[int] = None,
        vision_feature_select_strategy: Optional[str] = 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,
        cache_position: Optional[torch.LongTensor] = None,
        num_logits_to_keep: int = 0,
    ) -> Union[Tuple, LlavaNextCausalLMOutputWithPast]:
        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]`.

            num_logits_to_keep (`int`, *optional*):
                Calculate logits for the last `num_logits_to_keep` tokens. If `0`, calculate logits for all
                `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
                token can save memory, which becomes pretty significant for long sequences or large vocabulary size.

        Returns:

        Example:

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

        >>> model = LlavaNextForConditionalGeneration.from_pretrained("llava-hf/llava-v1.6-mistral-7b-hf")
        >>> processor = AutoProcessor.from_pretrained("llava-hf/llava-v1.6-mistral-7b-hf")

        >>> prompt = "[INST] <image>\nWhat is shown in this image? [/INST]"
        >>> url = "https://www.ilankelman.org/stopsigns/australia.jpg"
        >>> image = Image.open(requests.get(url, stream=True).raw)

        >>> inputs = processor(images=image, text=prompt, return_tensors="pt")

        >>> # Generate
        >>> generate_ids = model.generate(**inputs, max_length=30)
        >>> processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
        "[INST]  \nWhat is shown in this image? [/INST] The image appears to be a radar chart, which is a type of multi-dimensional plot (...)"
        ```"""

        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
        vision_feature_layer = (
            vision_feature_layer if vision_feature_layer is not None else self.config.vision_feature_layer
        )
        vision_feature_select_strategy = (
            vision_feature_select_strategy
            if vision_feature_select_strategy is not None
            else self.config.vision_feature_select_strategy
        )

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

        if pixel_values is not None and inputs_embeds is not None:
            raise ValueError(
                "You cannot specify both pixel_values and inputs_embeds at the same time, and must specify either one"
            )

        legacy_processing = False
        if inputs_embeds is None:
            inputs_embeds = self.get_input_embeddings()(input_ids)

            # if the number of image tokens is more than image embeddings seq length, then prob we expanded it in processing
            # not very reliable, but we don't expect one to actually pass 500+ images for one prompt
            # In case we're in decoding stage, legacy behavior is checked by presence of pixel values even if use_cache=True
            legacy_processing = (
                (input_ids == self.config.image_token_index).sum(1).max() < self.config.image_seq_length
            ) or (input_ids.shape[-1] == 1 and pixel_values is not None)

        image_features = None
        if pixel_values is not None and pixel_values.size(0) > 0:
            image_features, vq_loss = self.get_image_features(
                pixel_values,
                image_sizes,
                vision_feature_layer=vision_feature_layer,
                vision_feature_select_strategy=vision_feature_select_strategy,
            )

            # NOTE we only support multimodal_patch_merge_type == "spatial_unpad"
            image_features, feature_lens = self.pack_image_features(
                image_features,
                image_sizes,
                vision_feature_select_strategy=vision_feature_select_strategy,
                image_newline=self.image_newline,
            )

        if legacy_processing:
            logger.warning_once(
                "Expanding inputs for image tokens in LLaVa-NeXT should be done in processing. "
                "Please add `patch_size` and `vision_feature_select_strategy` to the model's processing config or set directly "
                "with `processor.patch_size = {{patch_size}}` and processor.vision_feature_select_strategy = {{vision_feature_select_strategy}}`. "
                "Using processors without these attributes in the config is deprecated and will throw an error in v4.47."
            )
            if input_ids.shape[1] != 1:
                inputs_embeds = inputs_embeds.to(image_features.dtype)
                inputs_embeds, attention_mask, position_ids, labels, _ = self._merge_input_ids_with_image_features(
                    image_features,
                    feature_lens,
                    inputs_embeds,
                    input_ids,
                    attention_mask,
                    position_ids,
                    labels=labels,
                )
                cache_position = torch.arange(attention_mask.shape[1], device=attention_mask.device)
            else:
                # Retrieve the first layer to inspect the logits and mask out the hidden states
                # that are set to 0
                first_layer_past_key_value = past_key_values[0][0][:, :, :, 0]

                # Sum all dimensions of head_dim (-2) to avoid random errors such as: https://github.com/huggingface/transformers/pull/28032#issuecomment-1863691941
                batch_index, non_attended_tokens = torch.where(first_layer_past_key_value.float().sum(-2) == 0)

                # Get the target length
                target_length = input_ids.shape[1]
                past_length = first_layer_past_key_value.shape[-1]

                extended_attention_mask = torch.ones(
                    (attention_mask.shape[0], past_length),
                    dtype=attention_mask.dtype,
                    device=attention_mask.device,
                )

                # Filter out only the tokens that can be un-attended, this can happen
                # if one uses Llava + Fused modules where the cache on the
                # first iteration is already big enough, or if one passes custom cache
                valid_indices = non_attended_tokens < extended_attention_mask.size(-1)
                new_batch_index = batch_index[valid_indices]
                new_non_attended_tokens = non_attended_tokens[valid_indices]

                # Zero-out the places where we don't need to attend
                extended_attention_mask[new_batch_index, new_non_attended_tokens] = 0
                attention_mask = torch.cat((extended_attention_mask, attention_mask[:, -target_length:]), dim=1)
                position_ids = torch.sum(attention_mask, dim=1).unsqueeze(-1) - 1
                cache_position = torch.arange(attention_mask.shape[1], device=attention_mask.device)[-target_length:]

        # TODO: @raushan retain only the new behavior after v4.47
        elif image_features is not None:
            n_image_tokens = (input_ids == self.config.image_token_index).sum().item()
            n_image_features = image_features.shape[0]
            if n_image_tokens != n_image_features:
                raise ValueError(
                    f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}"
                )
            special_image_mask = (
                (input_ids == self.config.image_token_index)
                .unsqueeze(-1)
                .expand_as(inputs_embeds)
                .to(inputs_embeds.device)
            )
            image_features = image_features.to(inputs_embeds.device, inputs_embeds.dtype)
            inputs_embeds = inputs_embeds.masked_scatter(special_image_mask, image_features)

        outputs = self.language_model(
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            cache_position=cache_position,
            num_logits_to_keep=num_logits_to_keep,
        )

        logits = outputs[0]

        loss = None
        if labels is not None:
            # Shift so that tokens < n predict n
            if attention_mask is not None:
                # we use the input attention mask to shift the logits and labels, because it is 2D.
                # we also crop attn mask in case it is longer, which happens in PrefixTuning with peft
                shift_attention_mask = attention_mask[:, -(logits.shape[1] - 1) :].to(logits.device)
                shift_logits = logits[..., :-1, :][shift_attention_mask.to(logits.device) != 0].contiguous()
                shift_labels = labels[..., 1:][shift_attention_mask.to(labels.device) != 0].contiguous()
            else:
                shift_logits = logits[..., :-1, :].contiguous()
                shift_labels = labels[..., 1:].contiguous()
            # Flatten the tokens
            loss_fct = nn.CrossEntropyLoss()
            loss = loss_fct(
                shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1).to(shift_logits.device)
            )
            print("This is original loss",loss)
            #vq_loss = vq_loss.to(loss.device)
            loss = loss 
        if not return_dict:
            output = (logits,) + outputs[1:]
            return (loss,) + output if loss is not None else output

        return LlavaNextCausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
            image_hidden_states=image_features if pixel_values is not None else None,
        )

    def prepare_inputs_for_generation(
        self,
        input_ids,
        past_key_values=None,
        inputs_embeds=None,
        pixel_values=None,
        image_sizes=None,
        attention_mask=None,
        cache_position=None,
        num_logits_to_keep=None,
        **kwargs,
    ):
        # Overwritten -- in specific circumstances we don't want to forward image inputs to the model

        model_inputs = self.language_model.prepare_inputs_for_generation(
            input_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            attention_mask=attention_mask,
            cache_position=cache_position,
            num_logits_to_keep=num_logits_to_keep,
            **kwargs,
        )

        # If we're in cached decoding stage, pixel values should be None because input ids do not contain special image token anymore
        # Otherwise we need pixel values to be passed to model
        if cache_position[0] == 0:
            model_inputs["pixel_values"] = pixel_values
            model_inputs["image_sizes"] = image_sizes

        return model_inputs