llava/model/llava_arch.py

#    Copyright 2023 Haotian Liu
#
#    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.

from email.mime import image
import os
from abc import ABC, abstractmethod

import torch
import torch.nn as nn

from .multimodal_encoder.builder import build_adapter_module, build_vision_tower, build_Qformer
from .multimodal_projector.builder import build_vision_projector

from llava.constants import IGNORE_INDEX, MM_TOKEN_INDEX, DEFAULT_IMAGE_PATCH_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN, DEFAULT_VIDEO_PATCH_TOKEN, DEFAULT_VIDEO_START_TOKEN, DEFAULT_VIDEO_END_TOKEN

from llava.mm_utils import get_anyres_image_grid_shape
from llava.utils import master_print
import tensorrt as trt
import pycuda.driver as cuda
import pycuda.autoinit
import subprocess
import torch.onnx

class LlavaMetaModel:

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

        if hasattr(config, "mm_vision_tower"):
            self.vision_tower = build_vision_tower(config, delay_load=True)
            self.mm_projector = build_vision_projector(config)
            if getattr(config, "qformer_model_path", None):
                self.Qformer, self.ln_vision, self.query_tokens = build_Qformer(
                            config.num_query_token, self.vision_tower.hidden_size)
                self.frame_position_encoding = nn.Embedding(
                    config.max_num_segments,
                    self.Qformer.config.hidden_size
                )
            if getattr(config, "adapter_module_name", None):
                self.adapter_module = build_adapter_module(config, self.vision_tower.hidden_size)
            if 'unpad' in getattr(config, 'mm_patch_merge_type', ''):
                self.image_newline = nn.Parameter(
                    torch.empty(config.hidden_size, dtype=self.dtype)
                )

        # Prepare TRT
        # self.trt_logger = trt.Logger(trt.Logger.WARNING)
        # self.trt_runtime = trt.Runtime(self.trt_logger)
        # trt.init_libnvinfer_plugins(None, "")
        
        # nvidia_smi_output = subprocess.check_output(["nvidia-smi", "-L"]).decode()
        # gpu_info = nvidia_smi_output.split(":")[1].split("(")[0].strip()
        # print(gpu_info)
        # if "A10" in gpu_info:
        #     vit_tagging_path = "./a10/vit.trt"
        # elif "A30" in gpu_info:
        #     vit_tagging_path = "./a30/vit.trt"
        # else:
        #     assert False,logging.info("just support in A10,A30")
        #     exit()
        
        # with open(vit_tagging_path, 'rb') as f:
        #     engine_data_vit = f.read()
        # self.vit_tag_trt_engine = self.trt_runtime.deserialize_cuda_engine(engine_data_vit)
        # self.vit_tag_trt_context = self.vit_tag_trt_engine.create_execution_context()
        
        # self.stream =  cuda.Stream()

        # TRT Implementation code stops at self.stream, proceed to the next part

    def get_vision_tower(self):
        vision_tower = getattr(self, 'vision_tower', None)
        if type(vision_tower) is list:
            vision_tower = vision_tower[0]
        return vision_tower
    
    def get_adapter_module(self):
        adapter_module = getattr(self, 'adapter_module', None)
        if type(adapter_module) is list:
            adapter_module = adapter_module[0]
        return adapter_module

    def get_qformer(self):
        qformer = getattr(self, 'Qformer', None)
        if type(qformer) is list:
            qformer = qformer[0]
        return qformer

    def get_ln_vision(self):
        ln_vision = getattr(self, 'ln_vision', None)
        if type(ln_vision) is list:
            ln_vision = ln_vision[0]
        return ln_vision
    
    def get_query_tokens(self):
        query_tokens = getattr(self, 'query_tokens', None)
        if type(query_tokens) is list:
            query_tokens = query_tokens[0]
        return query_tokens

    def get_frame_position_encoding(self):
        frame_position_encoding = getattr(self, 'frame_position_encoding', None)
        if type(frame_position_encoding) is list:
            frame_position_encoding = frame_position_encoding[0]
        return frame_position_encoding    

    def initialize_vision_modules(self, model_args, fsdp=None):
        vision_tower = model_args.vision_tower
        mm_vision_select_layer = model_args.mm_vision_select_layer
        mm_vision_select_feature = model_args.mm_vision_select_feature
        pretrain_mm_mlp_adapter = model_args.pretrain_mm_mlp_adapter
        mm_patch_merge_type = model_args.mm_patch_merge_type
        image_grid_pinpoints = model_args.image_grid_pinpoints
        self.config.mm_vision_tower = vision_tower
        self.config.img_size = model_args.img_size
        self.config.drop_path_rate = model_args.drop_path_rate
        self.config.vit_precision = model_args.vit_precision
        self.config.vit_model_path = model_args.vit_model_path   
        self.config.num_query_token = model_args.num_query_token
        self.config.qformer_model_path = model_args.qformer_model_path
        self.config.adapter_module_name = model_args.adapter_module_name
        self.config.adapter_module_path = model_args.adapter_module_path
        self.config.max_num_segments = model_args.max_num_segments
        self.config.pretrain_mm_mlp_adapter = pretrain_mm_mlp_adapter
        # TODO: FSDP training is not ready
        if self.get_vision_tower() is None:
            vision_tower = build_vision_tower(model_args)

            if fsdp is not None and len(fsdp) > 0:
                self.vision_tower = [vision_tower]
            else:
                self.vision_tower = vision_tower
        else:
            if fsdp is not None and len(fsdp) > 0:
                vision_tower = self.vision_tower[0]
            else:
                vision_tower = self.vision_tower
            vision_tower.load_model()

        self.config.use_mm_proj = True
        self.config.mm_projector_type = getattr(model_args, 'mm_projector_type', 'linear')
        self.config.mm_hidden_size = vision_tower.hidden_size
        self.config.mm_vision_hidden_size = vision_tower.hidden_size
        self.config.mm_vision_select_layer = mm_vision_select_layer
        self.config.mm_vision_select_feature = mm_vision_select_feature
        self.config.mm_patch_merge_type = mm_patch_merge_type
        self.config.image_grid_pinpoints = image_grid_pinpoints

        if getattr(model_args, "qformer_model_path", None):
            if self.get_qformer() is None:
                self.Qformer, self.ln_vision, self.query_tokens = build_Qformer(
                            model_args.num_query_token, self.vision_tower.hidden_size)
                self.frame_position_encoding = nn.Embedding(
                    model_args.max_num_segments,
                    self.Qformer.config.hidden_size
                )
                self.config.mm_hidden_size = self.Qformer.config.hidden_size
            # self.Qformer = self.Qformer.to(torch.bfloat16)
            if model_args.qformer_model_path != 'from_scratch':
                self.load_pretrained_qformer(model_args.qformer_model_path)
            
        if getattr(model_args, 'adapter_module_name', None):
            if self.get_adapter_module() is None:
                self.adapter_module = build_adapter_module(self.config, self.vision_tower.hidden_size)
                self.adapter_module.load_model()
            self.config.mm_hidden_size = self.adapter_module.output_dim

        if getattr(self, 'mm_projector', None) is None:

            self.mm_projector = build_vision_projector(self.config)

            if 'unpad' in mm_patch_merge_type:
                embed_std = 1 / torch.sqrt(torch.tensor(self.config.hidden_size, dtype=self.dtype))
                self.image_newline = nn.Parameter(
                    torch.randn(self.config.hidden_size, dtype=self.dtype) * embed_std
                )
        else:
            # In case it is frozen by LoRA
            for p in self.mm_projector.parameters():
                p.requires_grad = True

        if pretrain_mm_mlp_adapter is not None:
            mm_projector_weights = torch.load(pretrain_mm_mlp_adapter, map_location='cpu')
            def get_w(weights, keyword):
                return {k.split(keyword + '.')[1]: v for k, v in weights.items() if keyword in k}
            
            def get_variable_frame_encoding_w(model_weights, load_weights):
                model_len = model_weights.shape[0]
                load_weights = {'.'.join(k.split('.')[1:]): v for k, v in load_weights.items()}

                load_len = load_weights['frame_position_encoding.weight'].shape[0]
                if model_len == load_len:
                    return get_w(load_weights, 'frame_position_encoding')
                elif model_len < load_len:
                    value = load_weights['frame_position_encoding.weight'][:model_len]
                    return {'weight': value}
                else:
                    value = model_weights.clone().cpu()
                    value[:load_len] = load_weights['frame_position_encoding.weight']
                    return {'weight': value}
            
            self.mm_projector.load_state_dict(get_w(mm_projector_weights, 'mm_projector'))
            if self.get_frame_position_encoding():
                self.frame_position_encoding.load_state_dict(get_variable_frame_encoding_w(self.frame_position_encoding.weight, mm_projector_weights))
            
            master_print(f"Loaded pretrained parameters from {pretrain_mm_mlp_adapter}")


    def load_pretrained_qformer(self, model_path):
        if os.path.isfile(model_path):
            checkpoint = torch.load(model_path, map_location="cpu")
        else:
            raise RuntimeError("checkpoint path is invalid")
        if 'projector.bin' in model_path:
            state_dict = {}
            match_keys = ['Qformer', 'query_tokens']
            for k, v in checkpoint.items():
                flag = False
                for match_key in match_keys:
                    if match_key in k:
                        flag = True
                        break
                if flag:
                    state_dict[k.replace('model.', '')] = v            

        else:
            state_dict = checkpoint["model"]
        msg = self.load_state_dict(state_dict, strict=False)

        master_print(f"Loaded Qformer from {model_path}")
        # master_print(msg)

        # return msg


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 in CxHxW format.
    original_size (tuple): The original size of the image (height, width).

    Returns:
    torch.Tensor: The unpadded image tensor.
    """
    original_width, original_height = 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


class LlavaMetaForCausalLM(ABC):

    @abstractmethod
    def get_model(self):
        pass

    def get_vision_tower(self):
        return self.get_model().get_vision_tower()
    
    def get_adapter_module(self):
        return self.get_model().get_adapter_module()

    def get_ln_vision(self):
        return self.get_model().get_ln_vision()

    def get_qformer(self):
        return self.get_model().get_qformer()

    def get_query_tokens(self):
        return self.get_model().get_query_tokens()

    def get_frame_position_encoding(self):
        return self.get_model().get_frame_position_encoding()

    def encode_images(self, images):
        # Uncomment below to get normal output without tensorrt
        image_features = self.get_vision_tower()(images)

        #return image_features
        #print(image_features.shape)
        #print(images.shape)
        #exit()
        #print(images.shape)
        #exit()
        
        #-------------------- VIT CONVERSION START --------------------------
        # import torch.onnx
        # # Initialize the model, define the input, and export to ONNX
        # model = self.get_model().get_vision_tower().half()
        # device = next(model.parameters()).device

        # # Move all buffers and constants to the correct device
        # model.to(device)

        # # Ensure all buffers are on the same device
        # # for param in model.parameters():
        # #     param.data = param.data.to(device)
        # for buffer in model.buffers():
        #     buffer.data = buffer.data.to(device)

        # # Modify any control flow that uses tensors
        # # For example, in the model's forward method, ensure that any tensor used in control flow is converted to int

        # # Create a dummy input tensor with the same shape as the input tensor you will use in your application
        # dummy_input = torch.randn(10, 3, 224, 224, device=device, dtype=next(model.parameters()).dtype).half()

        # # Export the model
        # onnx_path = "vit.onnx"
        # torch.onnx.export(
        #     model, 
        #     dummy_input, 
        #     onnx_path, 
        #     export_params=True, 
        #     #opset_version=10, 
        #     do_constant_folding=False,  # Disable constant folding, need to do this in order to get onnx file.
        #     input_names=['input'], 
        #     output_names=['output'],
        #     dynamic_axes={'input' : {0 : 'batch_size'}, 'output' : {0 : 'batch_size'}}
        # )

        # print(images.shape)
        # exit()

        #--------------------- VIT CONVERSION ENDS HERE ----------------------

        
        # Get the device of the model's parameters
        # device = torch.device('cuda:0')
        # # Initialize the model, define the input, and export to ONNX
        # model = self.get_model().get_vision_tower()
        # model = model.to(device)
        # # Create a dummy input tensor with the same shape as the input tensor you will use in your application
        # dummy_input = torch.randn(10, 3, 224, 224).to(device)

        
        # # Export the model
        # onnx_path = "simple_model.onnx"
        # torch.onnx.export(
        #                     model, 
        #                     dummy_input, 
        #                     onnx_path, 
        #                     export_params=True, 
        #                     opset_version=10, 
        #                     do_constant_folding=True, 
        #                     input_names=['input'], 
        #                     output_names=['output'],
        #                     dynamic_axes={'input' : {0 : 'batch_size'}, 'output' : {0 : 'batch_size'}})

        
        # #print(images.shape)
        # exit()
        

        if self.get_qformer():
            image_features = self.get_ln_vision()(image_features)
            query_tokens = self.get_query_tokens()
            query_tokens = query_tokens.expand(image_features.shape[0], -1, -1)
            attn_mask = torch.ones(image_features.size()[:-1], dtype=torch.long).to(image_features.device)
            dtype_ = self.get_vision_tower().dtype
            # print(dtype_)
            image_features = self.qformer_fusion(
                query_tokens.to(dtype_),
                image_features.to(dtype_), 
                attn_mask
            ).to(images.dtype)

        # image_features = self.get_model().mm_projector(image_features)
        return image_features

    def qformer_fusion(self, query_tokens, features, attn_mask=None):
        qformer = self.get_qformer()
        query_output = qformer.bert(
            query_embeds=query_tokens,
            encoder_hidden_states=features,
            encoder_attention_mask=attn_mask,
            return_dict=True
        )
        return query_output.last_hidden_state

    def prepare_inputs_labels_for_multimodal(
        self, input_ids, position_ids, attention_mask, past_key_values, labels,
        images, image_sizes=None
    ):  

        vision_tower = self.get_vision_tower()
        if vision_tower is None or images is None or input_ids.shape[1] == 1:
            return input_ids, position_ids, attention_mask, past_key_values, None, labels
        
        # image:  list(B) of tensor[1, 3, 336, 336]
        # video:  list(B) of tensor[N, 3, 336, 336]
        # video_any_res: list(B) of tensor[N, P, 3, 336, 336]
        if type(images) is list or images.ndim == 5:
            if type(images) is list:
                images = [x.unsqueeze(0) if x.ndim == 3 else x for x in images]
            # video any res
            if images[0].ndim == 5:
                concat_images = torch.cat([image.flatten(0, 1) for image in images], dim=0)
                split_sizes = [image.shape[0:2] for image in images]
            else:
                concat_images = torch.cat([image for image in images], dim=0)
                split_sizes = [image.shape[0] for image in images]
            image_features = self.encode_images(concat_images)

            # add frame encoding then projector
            if images[0].ndim == 5:
                frame_ids = []
                for split_size in split_sizes:
                    frame_ids.append(torch.tensor([idx for idx in range(split_size[0]) for _ in range(split_size[1])], \
                                                    dtype=torch.long, device=image_features.device))
            else:
                frame_ids = [torch.arange(split_size, dtype=torch.long, device=image_features.device)
                                for split_size in split_sizes]
            frame_ids = torch.concat(frame_ids)
            frame_position_encoding = self.get_frame_position_encoding()
            if frame_position_encoding:

                frame_embeddings = frame_position_encoding(frame_ids).unsqueeze(-2)
                image_features += frame_embeddings

            # TODO: add fusion model, rewrite this part in the future
            adapter_module = self.get_adapter_module()
            if adapter_module:
                image_features = adapter_module(image_features, frame_ids)
            image_features = self.get_model().mm_projector(image_features)
            if images[0].ndim == 5:
                split_sizes = [split_size[0] * split_size[1] for split_size in split_sizes]
            image_features = torch.split(image_features, split_sizes, dim=0)
            if adapter_module:
                # image_features = [image_features[i].view(images[i].shape[0], images[i].shape[1], -1) for i in range(image_features.shape[0])]
                image_features = [x.view(im.shape[0], -1, x.shape[2]) for x, im in zip(image_features, images)]
                image_features = adapter_module.compress_token_per_img(image_features)

            mm_patch_merge_type = getattr(self.config, 'mm_patch_merge_type', 'flat')
            image_aspect_ratio = getattr(self.config, 'image_aspect_ratio', 'square')
            if mm_patch_merge_type == 'flat':
                image_features = [x.flatten(0, 1) for x in image_features]
            elif mm_patch_merge_type.startswith('spatial'):
                new_image_features = []
                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.get_vision_tower().num_patches_per_side
                        assert height * width == base_image_feature.shape[0]
                        if image_aspect_ratio == 'anyres':
                            num_patch_width, num_patch_height = get_anyres_image_grid_shape(image_sizes[image_idx], self.config.image_grid_pinpoints, self.get_vision_tower().config.image_size)
                            image_feature = image_feature.view(num_patch_height, num_patch_width, height, width, -1)
                        else:
                            raise NotImplementedError
                        if 'unpad' in mm_patch_merge_type:
                            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])
                            image_feature = torch.cat((
                                image_feature,
                                self.model.image_newline[:, None, None].expand(*image_feature.shape[:-1], 1).to(image_feature.device)
                            ), dim=-1)
                            image_feature = image_feature.flatten(1, 2).transpose(0, 1)
                        else:
                            image_feature = image_feature.permute(0, 2, 1, 3, 4).contiguous()
                            image_feature = image_feature.flatten(0, 3)
                        image_feature = torch.cat((base_image_feature, image_feature), dim=0)
                    else:
                        image_feature = image_feature[0]
                        if 'unpad' in mm_patch_merge_type:
                            image_feature = torch.cat((
                                image_feature,
                                self.model.image_newline[None].to(image_feature.device)
                            ), dim=0)
                    new_image_features.append(image_feature)
                image_features = new_image_features
            else:
                raise ValueError(f"Unexpected mm_patch_merge_type: {self.config.mm_patch_merge_type}")

        else:
            image_features = self.encode_images(images)

        # if getattr(self.config, 'tune_mm_mlp_adapter', False) and getattr(self.config, 'mm_use_start_end', False):
        #     raise NotImplementedError


        # TODO: Currently, all the embed_token will bu update when tune_mm_mlp_adapter = True && mm_use_start_end = True

        # Let's just add dummy tensors if they do not exist,
        # it is a headache to deal with None all the time.
        # But it is not ideal, and if you have a better idea,
        # please open an issue / submit a PR, thanks.
        _labels = labels
        _position_ids = position_ids
        _attention_mask = attention_mask
        if attention_mask is None:
            attention_mask = torch.ones_like(input_ids, dtype=torch.bool)
        else:
            attention_mask = attention_mask.bool()
        if position_ids is None:
            position_ids = torch.arange(0, input_ids.shape[1], dtype=torch.long, device=input_ids.device)
        if labels is None:
            labels = torch.full_like(input_ids, IGNORE_INDEX)

        # remove the padding using attention_mask -- FIXME
        _input_ids = input_ids
        input_ids = [cur_input_ids[cur_attention_mask] for cur_input_ids, cur_attention_mask in zip(input_ids, attention_mask)]
        labels = [cur_labels[cur_attention_mask] for cur_labels, cur_attention_mask in zip(labels, attention_mask)]

        new_input_embeds = []
        new_labels = []
        cur_image_idx = 0
        for batch_idx, cur_input_ids in enumerate(input_ids):
            num_images = (cur_input_ids == MM_TOKEN_INDEX).sum()
            if num_images == 0:
                cur_image_features = image_features[cur_image_idx]
                cur_input_embeds_1 = self.get_model().embed_tokens(cur_input_ids)
                cur_input_embeds = torch.cat([cur_input_embeds_1, cur_image_features[0:0]], dim=0)
                new_input_embeds.append(cur_input_embeds)
                new_labels.append(labels[batch_idx])
                cur_image_idx += 1
                continue

            image_token_indices = [-1] + torch.where(cur_input_ids == MM_TOKEN_INDEX)[0].tolist() + [cur_input_ids.shape[0]]
            cur_input_ids_noim = []
            cur_labels = labels[batch_idx]
            cur_labels_noim = []
            for i in range(len(image_token_indices) - 1):
                cur_input_ids_noim.append(cur_input_ids[image_token_indices[i]+1:image_token_indices[i+1]])
                cur_labels_noim.append(cur_labels[image_token_indices[i]+1:image_token_indices[i+1]])
            split_sizes = [x.shape[0] for x in cur_labels_noim]
            cur_input_embeds = self.get_model().embed_tokens(torch.cat(cur_input_ids_noim))
            cur_input_embeds_no_im = torch.split(cur_input_embeds, split_sizes, dim=0)
            cur_new_input_embeds = []
            cur_new_labels = []

            for i in range(num_images + 1):
                cur_new_input_embeds.append(cur_input_embeds_no_im[i])
                cur_new_labels.append(cur_labels_noim[i])
                if i < num_images:
                    cur_image_features = image_features[cur_image_idx]
                    cur_image_idx += 1
                    cur_new_input_embeds.append(cur_image_features)
                    cur_new_labels.append(torch.full((cur_image_features.shape[0],), IGNORE_INDEX, device=cur_labels.device, dtype=cur_labels.dtype))

            cur_new_input_embeds = [x.to(self.device) for x in cur_new_input_embeds]

            cur_new_input_embeds = torch.cat(cur_new_input_embeds)
            cur_new_labels = torch.cat(cur_new_labels)

            new_input_embeds.append(cur_new_input_embeds)
            new_labels.append(cur_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 new_input_embeds]
            new_labels = [x[:tokenizer_model_max_length] for x in new_labels]

        # Combine them
        max_len = max(x.shape[0] for x in new_input_embeds)
        batch_size = len(new_input_embeds)

        new_input_embeds_padded = []
        new_labels_padded = torch.full((batch_size, max_len), IGNORE_INDEX, dtype=new_labels[0].dtype, device=new_labels[0].device)
        attention_mask = torch.zeros((batch_size, max_len), dtype=attention_mask.dtype, device=attention_mask.device)
        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 = cur_new_embed.shape[0]
            if getattr(self.config, 'tokenizer_padding_side', 'right') == "left":
                new_input_embeds_padded.append(torch.cat((
                    torch.zeros((max_len - cur_len, cur_new_embed.shape[1]), dtype=cur_new_embed.dtype, device=cur_new_embed.device),
                    cur_new_embed
                ), dim=0))
                if cur_len > 0:
                    new_labels_padded[i, -cur_len:] = cur_new_labels
                    attention_mask[i, -cur_len:] = True
                    position_ids[i, -cur_len:] = torch.arange(0, cur_len, dtype=position_ids.dtype, device=position_ids.device)
            else:
                new_input_embeds_padded.append(torch.cat((
                    cur_new_embed,
                    torch.zeros((max_len - cur_len, cur_new_embed.shape[1]), dtype=cur_new_embed.dtype, device=cur_new_embed.device)
                ), dim=0))
                if cur_len > 0:
                    new_labels_padded[i, :cur_len] = cur_new_labels
                    attention_mask[i, :cur_len] = True
                    position_ids[i, :cur_len] = torch.arange(0, cur_len, dtype=position_ids.dtype, device=position_ids.device)

        new_input_embeds = torch.stack(new_input_embeds_padded, dim=0)

        if _labels is None:
            new_labels = None
        else:
            new_labels = new_labels_padded

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

        if _position_ids is None:
            position_ids = None

        return None, position_ids, attention_mask, past_key_values, new_input_embeds, new_labels

    def initialize_vision_tokenizer(self, model_args, tokenizer):
        if model_args.mm_use_patch_token:
            tokenizer.add_tokens([DEFAULT_IMAGE_PATCH_TOKEN, DEFAULT_VIDEO_PATCH_TOKEN], special_tokens=True)
            self.resize_token_embeddings(len(tokenizer))

        if model_args.mm_use_start_end:
            num_new_tokens = tokenizer.add_tokens([DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN, DEFAULT_VIDEO_START_TOKEN, DEFAULT_VIDEO_END_TOKEN], special_tokens=True)
            self.resize_token_embeddings(len(tokenizer))

            if num_new_tokens > 0:
                input_embeddings = self.get_input_embeddings().weight.data
                output_embeddings = self.get_output_embeddings().weight.data

                input_embeddings_avg = input_embeddings[:-num_new_tokens].mean(
                    dim=0, keepdim=True)
                output_embeddings_avg = output_embeddings[:-num_new_tokens].mean(
                    dim=0, keepdim=True)

                input_embeddings[-num_new_tokens:] = input_embeddings_avg
                output_embeddings[-num_new_tokens:] = output_embeddings_avg

            if model_args.tune_mm_mlp_adapter:
                for p in self.get_input_embeddings().parameters():
                    p.requires_grad = True
                if 'gemma' in model_args.model_name_or_path:
                    # gemma use the same embedding for input and output
                    pass
                else:
                    for p in self.get_output_embeddings().parameters():
                        p.requires_grad = False

            if model_args.pretrain_mm_mlp_adapter:
                # raise NotImplementedError
                mm_projector_weights = torch.load(model_args.pretrain_mm_mlp_adapter, map_location='cpu')
                mm_projector_weights = {'.'.join(k.split('.')[1:]): v for k, v in mm_projector_weights.items()}
                # embed_tokens_weight = mm_projector_weights['embed_tokens.weight']
                # input_embeddings[:] = embed_tokens_weight
                # if 'gemma' in model_args.model_name_or_path:
                #     output_embeddings[:] = embed_tokens_weight
                assert num_new_tokens == 4
                # if input_embeddings.shape == embed_tokens_weight.shape:
                #     input_embeddings[-num_new_tokens:] = embed_tokens_weight[-num_new_tokens:]
                # elif embed_tokens_weight.shape[0] == num_new_tokens:
                #     input_embeddings[-num_new_tokens:] = embed_tokens_weight
                # else:
                #     raise ValueError(f"Unexpected embed_tokens_weight shape. Pretrained: {embed_tokens_weight.shape}. Current: {input_embeddings.shape}. Numer of new tokens: {num_new_tokens}.")
        elif model_args.mm_use_patch_token:
            if model_args.tune_mm_mlp_adapter:
                for p in self.get_input_embeddings().parameters():
                    p.requires_grad = False
                for p in self.get_output_embeddings().parameters():
                    p.requires_grad = False