01引言

近期Internvl2.5发布，性能与GPT-4o和Claude-3.5-sonnet等领先的商业模型相媲美，成为首个在MMMU上超过70%的开源模型，通过链式思考（CoT）推理实现了3.7个百分点的提升，展示了强大的测试时间可扩展性潜力。InternVL 2.5是基于InternVL 2.0发展而来，通过增强训练和测试策略以及提高数据质量来进一步提升性能。该模型在多个方面进行了优化，包括视觉编码器、语言模型、数据集大小和测试时间配置等方面的研究，旨在探索模型规模与性能之间的关系。InternVL 2.5经过广泛的评估，在多项基准测试中展现了竞争性的性能，特别是在多学科推理、文档理解、多图像/视频理解、现实世界理解、多模态幻觉检测、视觉地面化、多语言能力以及纯语言处理等领域。

📕 课代表划重点：

1. InternVL 2.5的开源多模态大型语言模型发布，通过数据和测试时间扩展提高了性能边界

2. 实验表明，InternVL 2.5在多种基准测试中表现出竞争力，并超越了商业模型GPT-4o和Claude-3.5-Sonnet

3. 该模型采用了新的训练和测试策略以及高质量的数据集，能够处理多种模态的信息，包括文本、图像和视频等

4. 通过链式思维推理等方式，该模型在MMMU基准测试中实现了超过70%的准确率，展现了强大的测试时间扩展潜力

5. 该研究为开放源代码社区提供了一个新标准，用于开发和应用多模态AI系统

InternVL 2.5保留了其前身的相同模型架构：InternVL 1.5 和 InternVL 2.0 ，遵循各种 MLLM 研究中广泛采用的“ViT-MLP-LLM”范式。InternVL 2.5实现将一种新的增量预训练的InternViT-6B或InternViT-300M与各种不同大小和类型的预先训练的LLMs集成在一起，包括InternLM2.5和Qwen 2.5，使用随机初始化的两层MLP投影器。正如之前的版本一样，为了增强高分辨率处理的可扩展性，研究团队简单地应用了一个像素无序操作，将视觉令牌的数量减少到原始数量的一半。因此，在InternVL 2.5中，一个448×448图像块由256个视觉令牌表示。

模型链接：

https://www.modelscope.cn/collections/InternVL-25-fbde6e47302942

 

02模型下载

命令行下载：

modelscope download --model OpenGVLab/InternVL2_5-4B

Python SDK下载：

#模型下载
from modelscope import snapshot_download
model_dir = snapshot_download('OpenGVLab/InternVL2_5-4B')

03模型推理

transformers推理

import numpy as np
import torch
import torchvision.transforms as T
from decord import VideoReader, cpu
from PIL import Image
from torchvision.transforms.functional import InterpolationMode
from modelscope import AutoModel, AutoTokenizer

IMAGENET_MEAN = (0.485, 0.456, 0.406)
IMAGENET_STD = (0.229, 0.224, 0.225)

def build_transform(input_size):
    MEAN, STD = IMAGENET_MEAN, IMAGENET_STD
    transform = T.Compose([
        T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img),
        T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC),
        T.ToTensor(),
        T.Normalize(mean=MEAN, std=STD)
    ])
    return transform

def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
    best_ratio_diff = float('inf')
    best_ratio = (1, 1)
    area = width * height
    for ratio in target_ratios:
        target_aspect_ratio = ratio[0] / ratio[1]
        ratio_diff = abs(aspect_ratio - target_aspect_ratio)
        if ratio_diff < best_ratio_diff:
            best_ratio_diff = ratio_diff
            best_ratio = ratio
        elif ratio_diff == best_ratio_diff:
            if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
                best_ratio = ratio
    return best_ratio

def dynamic_preprocess(image, min_num=1, max_num=12, image_size=448, use_thumbnail=False):
    orig_width, orig_height = image.size
    aspect_ratio = orig_width / orig_height

    # calculate the existing image aspect ratio
    target_ratios = set(
        (i, j) for n in range(min_num, max_num + 1) for i in range(1, n + 1) for j in range(1, n + 1) if
        i * j <= max_num and i * j >= min_num)
    target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])

    # find the closest aspect ratio to the target
    target_aspect_ratio = find_closest_aspect_ratio(
        aspect_ratio, target_ratios, orig_width, orig_height, image_size)

    # calculate the target width and height
    target_width = image_size * target_aspect_ratio[0]
    target_height = image_size * target_aspect_ratio[1]
    blocks = target_aspect_ratio[0] * target_aspect_ratio[1]

    # resize the image
    resized_img = image.resize((target_width, target_height))
    processed_images = []
    for i in range(blocks):
        box = (
            (i % (target_width // image_size)) * image_size,
            (i // (target_width // image_size)) * image_size,
            ((i % (target_width // image_size)) + 1) * image_size,
            ((i // (target_width // image_size)) + 1) * image_size
        )
        # split the image
        split_img = resized_img.crop(box)
        processed_images.append(split_img)
    assert len(processed_images) == blocks
    if use_thumbnail and len(processed_images) != 1:
        thumbnail_img = image.resize((image_size, image_size))
        processed_images.append(thumbnail_img)
    return processed_images

def load_image(image_file, input_size=448, max_num=12):
    image = Image.open(image_file).convert('RGB')
    transform = build_transform(input_size=input_size)
    images = dynamic_preprocess(image, image_size=input_size, use_thumbnail=True, max_num=max_num)
    pixel_values = [transform(image) for image in images]
    pixel_values = torch.stack(pixel_values)
    return pixel_values

# If you want to load a model using multiple GPUs, please refer to the `Multiple GPUs` section.
path = 'OpenGVLab/InternVL2_5-4B'
model = AutoModel.from_pretrained(
    path,
    torch_dtype=torch.bfloat16,
    low_cpu_mem_usage=True,
    use_flash_attn=True,
    trust_remote_code=True).eval().cuda()
tokenizer = AutoTokenizer.from_pretrained(path, trust_remote_code=True, use_fast=False)

# set the max number of tiles in `max_num`
pixel_values = load_image('./awesome.png', max_num=12).to(torch.bfloat16).cuda()
generation_config = dict(max_new_tokens=1024, do_sample=True)

# pure-text conversation (纯文本对话)
question = 'Hello, who are you?'
response, history = model.chat(tokenizer, None, question, generation_config, history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')

question = 'Can you tell me a story?'
response, history = model.chat(tokenizer, None, question, generation_config, history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')

# single-image single-round conversation (单图单轮对话)
question = '<image>\nPlease describe the image shortly.'
response = model.chat(tokenizer, pixel_values, question, generation_config)
print(f'User: {question}\nAssistant: {response}')

# single-image multi-round conversation (单图多轮对话)
question = '<image>\nPlease describe the image in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')

question = 'Please write a poem according to the image.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')

# multi-image multi-round conversation, combined images (多图多轮对话，拼接图像)
pixel_values1 = load_image('./awesome.png', max_num=12).to(torch.bfloat16).cuda()
pixel_values2 = load_image('./noword.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0)

question = '<image>\nDescribe the two images in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
                               history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')

question = 'What are the similarities and differences between these two images.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
                               history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')

# multi-image multi-round conversation, separate images (多图多轮对话，独立图像)
pixel_values1 = load_image('./awesome.png', max_num=12).to(torch.bfloat16).cuda()
pixel_values2 = load_image('./noword.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0)
num_patches_list = [pixel_values1.size(0), pixel_values2.size(0)]

question = 'Image-1: <image>\nImage-2: <image>\nDescribe the two images in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
                               num_patches_list=num_patches_list,
                               history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')

question = 'What are the similarities and differences between these two images.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
                               num_patches_list=num_patches_list,
                               history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')

# batch inference, single image per sample (单图批处理)
pixel_values1 = load_image('./awesome.png', max_num=12).to(torch.bfloat16).cuda()
pixel_values2 = load_image('./noword.jpg', max_num=12).to(torch.bfloat16).cuda()
num_patches_list = [pixel_values1.size(0), pixel_values2.size(0)]
pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0)

questions = ['<image>\nDescribe the image in detail.'] * len(num_patches_list)
responses = model.batch_chat(tokenizer, pixel_values,
                             num_patches_list=num_patches_list,
                             questions=questions,
                             generation_config=generation_config)
for question, response in zip(questions, responses):
    print(f'User: {question}\nAssistant: {response}')

# video multi-round conversation (视频多轮对话)
def get_index(bound, fps, max_frame, first_idx=0, num_segments=32):
    if bound:
        start, end = bound[0], bound[1]
    else:
        start, end = -100000, 100000
    start_idx = max(first_idx, round(start * fps))
    end_idx = min(round(end * fps), max_frame)
    seg_size = float(end_idx - start_idx) / num_segments
    frame_indices = np.array([
        int(start_idx + (seg_size / 2) + np.round(seg_size * idx))
        for idx in range(num_segments)
    ])
    return frame_indices

def load_video(video_path, bound=None, input_size=448, max_num=1, num_segments=32):
    vr = VideoReader(video_path, ctx=cpu(0), num_threads=1)
    max_frame = len(vr) - 1
    fps = float(vr.get_avg_fps())

    pixel_values_list, num_patches_list = [], []
    transform = build_transform(input_size=input_size)
    frame_indices = get_index(bound, fps, max_frame, first_idx=0, num_segments=num_segments)
    for frame_index in frame_indices:
        img = Image.fromarray(vr[frame_index].asnumpy()).convert('RGB')
        img = dynamic_preprocess(img, image_size=input_size, use_thumbnail=True, max_num=max_num)
        pixel_values = [transform(tile) for tile in img]
        pixel_values = torch.stack(pixel_values)
        num_patches_list.append(pixel_values.shape[0])
        pixel_values_list.append(pixel_values)
    pixel_values = torch.cat(pixel_values_list)
    return pixel_values, num_patches_list

video_path = './showcase.mp4'
pixel_values, num_patches_list = load_video(video_path, num_segments=8, max_num=1)
pixel_values = pixel_values.to(torch.bfloat16).cuda()
video_prefix = ''.join([f'Frame{i+1}: <image>\n' for i in range(len(num_patches_list))])
question = video_prefix + 'What is the red panda doing?'
# Frame1: <image>\nFrame2: <image>\n...\nFrame8: <image>\n{question}
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
                               num_patches_list=num_patches_list, history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')

question = 'Describe this video in detail. Don\'t repeat.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
                               num_patches_list=num_patches_list, history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')

 

流式输出：

from transformers import TextIteratorStreamer
from threading import Thread

# Initialize the streamer
streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True, timeout=10)
# Define the generation configuration
generation_config = dict(max_new_tokens=1024, do_sample=False, streamer=streamer)
# Start the model chat in a separate thread
thread = Thread(target=model.chat, kwargs=dict(
    tokenizer=tokenizer, pixel_values=pixel_values, question=question,
    history=None, return_history=False, generation_config=generation_config,
))
thread.start()

# Initialize an empty string to store the generated text
generated_text = ''
# Loop through the streamer to get the new text as it is generated
for new_text in streamer:
    if new_text == model.conv_template.sep:
        break
    generated_text += new_text
    print(new_text, end='', flush=True)  # Print each new chunk of generated text on the same line

显存占用：

 

lmdeploy推理

安装依赖：

pip install lmdeploy -U

示例代码：

from lmdeploy import pipeline, TurbomindEngineConfig
from lmdeploy.vl import load_image
from modelscope import snapshot_download

model = snapshot_download('OpenGVLab/InternVL2_5-4B')
image = load_image('https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/tests/data/tiger.jpeg')
pipe = pipeline(model, backend_config=TurbomindEngineConfig(session_len=8192))
response = pipe(('describe this image', image))
print(response.text)

 

lmdeploy部署本地服务：

lmdeploy serve api_server ./InternVL2_5-4B/ --backend turbomind --server-port 23333
推理服务：
from openai import OpenAI

client = OpenAI(api_key='YOUR_API_KEY', base_url='http://0.0.0.0:23333/v1')
model_name = client.models.list().data[0].id
response = client.chat.completions.create(
    model=model_name,
    messages=[{
        'role':
        'user',
        'content': [{
            'type': 'text',
            'text': 'describe this image',
        }, {
            'type': 'image_url',
            'image_url': {
                'url':
                'https://modelscope.oss-cn-beijing.aliyuncs.com/resource/tiger.jpeg',
            },
        }],
    }],
    temperature=0.8,
    top_p=0.8)
print(response)

04 模型训练

我们使用ms-swift 3.0对InternVL2.5-2B进行微调. ms-swift是魔搭社区官方提供的LLM与多模态LLM微调部署框架，支持400+LLM和100+多模态LLM。

这里我们使用python的方式对InternVL2.5-2B进行Latex-OCR的微调。借此我们可以快速了解微调中的一些细节，这对自定义训练过程很有帮助。

如果出现兼容问题，请关注：

https://github.com/modelscope/ms-swift/tree/main/examples/train/notebook

首先我们需要从源代码安装ms-swift3.0

git clone https://github.com/modelscope/ms-swift.git
cd ms-swift
pip install -e '.[llm]'

 

首先我们需要导入一些包：

import os
os.environ['CUDA_VISIBLE_DEVICES'] = '0'

from swift.llm import (
    get_model_tokenizer, load_dataset, get_template, EncodePreprocessor, get_model_arch,
    get_multimodal_target_regex, LazyLLMDataset
)
from swift.utils import get_logger, get_model_parameter_info, plot_images, seed_everything
from swift.tuners import Swift, LoraConfig
from swift.trainers import Seq2SeqTrainer, Seq2SeqTrainingArguments
from functools import partial

logger = get_logger()
seed_everything(42)

 

设置一些训练的超参数：

# model
model_id_or_path = 'OpenGVLab/InternVL2_5-2B'
system = None  # 使用template中定义的默认system
output_dir = 'output/InternVL2_5-2B'

# dataset
dataset = ['AI-ModelScope/LaTeX_OCR#20000']  # dataset_id或者dataset_path。这里我们采样20000条数据样本
data_seed = 42
max_length = 8192
split_dataset_ratio = 0.01  # 切分验证集的比例
num_proc = 4  # 数据处理的进程数
strict = False

# lora
lora_rank = 8
lora_alpha = 32
freeze_llm = False
freeze_vit = True
freeze_aligner = True

# training_args
training_args = Seq2SeqTrainingArguments(
    output_dir=output_dir,
    learning_rate=1e-4,
    per_device_train_batch_size=1,
    per_device_eval_batch_size=1,
    gradient_checkpointing=True,
    weight_decay=0.1,
    lr_scheduler_type='cosine',
    warmup_ratio=0.05,
    report_to=['tensorboard'],
    logging_first_step=True,
    save_strategy='steps',
    save_steps=100,
    eval_strategy='steps',
    eval_steps=100,
    gradient_accumulation_steps=16,
    # 为了更快查看训练效果，这里设置为1。通常情况下，需要设置为更大的数。
    num_train_epochs=1,
    metric_for_best_model='loss',
    save_total_limit=2,
    logging_steps=5,
    dataloader_num_workers=4,
    data_seed=data_seed,
    remove_unused_columns=False
)

output_dir = os.path.abspath(os.path.expanduser(output_dir))
logger.info(f'output_dir: {output_dir}')

准备模型和对话模板：

# 获取model和template
model, processor = get_model_tokenizer(model_id_or_path)
logger.info(f'model_info: {model.model_info}')
template = get_template(model.model_meta.template, processor, default_system=system, max_length=max_length)
template.set_mode('train')

# 获取target_modules并在模型中加入可训练的LoRA模块
model_arch = get_model_arch(model.model_meta.model_arch)
target_modules = get_multimodal_target_regex(model_arch, freeze_llm=freeze_llm, freeze_vit=freeze_vit, 
                            freeze_aligner=freeze_aligner)
lora_config = LoraConfig(task_type='CAUSAL_LM', r=lora_rank, lora_alpha=lora_alpha,
                         target_modules=target_modules)
model = Swift.prepare_model(model, lora_config)
logger.info(f'lora_config: {lora_config}')

# 打印模型结构和可训练参数
logger.info(f'model: {model}')
model_parameter_info = get_model_parameter_info(model)
logger.info(f'model_parameter_info: {model_parameter_info}')

 

准备训练和验证数据集：

# 下载并载入数据集，切分成训练集和验证集
train_dataset, val_dataset = load_dataset(dataset, split_dataset_ratio=split_dataset_ratio, num_proc=num_proc,
                                          strict=strict, seed=data_seed)

logger.info(f'train_dataset: {train_dataset}')
logger.info(f'val_dataset: {val_dataset}')
logger.info(f'train_dataset[0]: {train_dataset[0]}')

# 将文本encode成tokens
train_dataset = LazyLLMDataset(
            train_dataset, template.encode, strict=strict, random_state=data_seed)
val_dataset = LazyLLMDataset(
            val_dataset, template.encode, strict=strict, random_state=data_seed)
data = train_dataset[0]
logger.info(f'encoded_train_dataset[0]: {data}')

template.print_inputs(data)

 

使用trainer开启训练：

model.enable_input_require_grads()  # 兼容gradient checkpointing
template.register_post_encode_hook([model])  # 将post_encode注册到forward_pre_hook中
trainer = Seq2SeqTrainer(
    model=model,
    args=training_args,
    data_collator=template.data_collator,
    train_dataset=train_dataset,
    eval_dataset=val_dataset,
    template=template,
)
trainer.model_accepts_loss_kwargs = True  # 兼容transformers>=4.46
trainer.train()

last_model_checkpoint = trainer.state.last_model_checkpoint
logger.info(f'last_model_checkpoint: {last_model_checkpoint}')

 

可视化训练loss: （这里我们只训练了400个steps）

你也可以使用tensorboard在训练过程中可视化训练loss，输入以下命令：`tensorboard --logdir '{output_dir}/runs'`

images_dir = os.path.join(output_dir, 'images')
logger.info(f'images_dir: {images_dir}')
plot_images(images_dir, training_args.logging_dir, ['train/loss'], 0.9)  # 保存训练loss图

训练后推理

导入一些包：

import os
os.environ['CUDA_VISIBLE_DEVICES'] = '0'

from swift.llm import (
    InferEngine, InferRequest, PtEngine, RequestConfig, get_template, load_dataset, load_image
)
from swift.tuners import Swift
from swift.utils import get_model_parameter_info, get_logger, seed_everything
logger = get_logger()
seed_everything(42)

 

推理超参数设置：

last_model_checkpoint = 'output/InternVL2_5-2B/vx-xxx/checkpoint-xxx'

# model
model_id_or_path = 'OpenGVLab/InternVL2_5-2B'  # model_id or model_path

# dataset
dataset = ['AI-ModelScope/LaTeX_OCR#20000']
data_seed = 42
split_dataset_ratio = 0.01
num_proc = 4
strict = False

# generation_config
max_new_tokens = 512
temperature = 0

 

我们使用infer_backend 'pt'来对训练后的模型进行推理，如果要使用vllm/lmdeploy进行加速，可以参考：https://github.com/modelscope/ms-swift/blob/main/examples/infer/demo_mllm.py

engine = PtEngine(model)
engine.model = Swift.from_pretrained(engine.model, last_model_checkpoint)
engine.model.requires_grad_()  # 修复peft将embedding层的requires_grad设置为True
template = get_template(engine.model.model_meta.template, engine.tokenizer)

model_parameter_info = get_model_parameter_info(engine.model)
logger.info(f'model_parameter_info: {model_parameter_info}')

获取验证集：

# 由于设置了data_seed，这里的验证集即为训练时的验证集
_, val_dataset = load_dataset(dataset, split_dataset_ratio=split_dataset_ratio, num_proc=num_proc,
                                          strict=strict, seed=data_seed)
val_dataset = val_dataset.select(range(10))  # 取前10条

流式推理，并保存验证集中的图片：

def infer_stream(engine: InferEngine, infer_request: InferRequest):
    request_config = RequestConfig(max_tokens=max_new_tokens, temperature=temperature, stream=True)
    gen = engine.infer([infer_request], request_config)
    query = infer_request.messages[0]['content']
    print(f'query: {query}\nresponse: ', end='')
    for resp_list in gen:
        print(resp_list[0].choices[0].delta.content, end='', flush=True)
    print()

os.makedirs('images', exist_ok=True)
for i, data in enumerate(val_dataset):
    image = load_image(data['images'][0]['bytes'])
    image.save(f'images/{i}.png')
    infer_stream(engine, InferRequest(**data))
    print('-' * 50)

推理效果：

 

 

 

 

 


点击链接查看原文：ModelScope 魔搭社区

 

 

Logo
欢迎加入ModelScope魔搭中文开源社区

ModelScope旨在打造下一代开源的模型即服务共享平台，为泛AI开发者提供灵活、易用、低成本的一站式模型服务产品，让模型应用更简单！

更多推荐

AI赋能大学计划·大模型技术与应用实战学生训练营——电子科技大学站圆满结营

12月05日，由中国软件行业校园招聘与实习公共服务平台携手阿里魔搭社区共同举办的AI赋能大学计划·大模型技术与产业趋势高校行AIGC项目实战营·电子科技大学站圆满结营。

avatar ModelScope魔搭社区
cover

Llama 3.3开源！70B媲美405B性能，支持128K上下文

avatar ModelScope魔搭社区
cover

魔搭社区每周速递（12.01-12.07）

avatar ModelScope魔搭社区