Mono-InternVL-2B / README.md

nielsr HF Staff

Update model card for Mono-InternVL-2B with Mono-InternVL-1.5 paper and comprehensive details

43d1cd7 verified 6 months ago

preview code

raw

history blame

17.8 kB

metadata

base_model:
  - internlm/internlm2-chat-1_8b
language:
  - multilingual
library_name: transformers
license: mit
pipeline_tag: image-text-to-text
tags:
  - internvl
  - vision
  - ocr
  - custom_code
  - moe
base_model_relation: merge

Mono-InternVL-2B

This repository contains the instruction-tuned Mono-InternVL-2B model, which has 1.8B activated parameters (3B in total). It is built upon internlm2-chat-1_8b.

Mono-InternVL-2B is part of the Mono-InternVL-1.5 family, presented in the paper Mono-InternVL-1.5: Towards Cheaper and Faster Monolithic Multimodal Large Language Models. Mono-InternVL-1.5 integrates visual encoding and language decoding into a single model, addressing optimization challenges and catastrophic forgetting common in monolithic MLLMs. It does this by embedding a new visual parameter space into a pre-trained LLM, enabling stable learning of visual knowledge from noisy data via delta tuning. This version features improved Endogenous Visual Pre-training (EViP++) with additional visual attention experts and re-organized pre-training for efficiency. During inference, it includes a fused CUDA kernel to speed up MoE operations, significantly reducing training and inference costs while maintaining competitive performance.

Please refer to our project page and GitHub repository for further introduction and usage.

radar chart

architecture

Introduction

We release Mono-InternVL, a monolithic multimodal large language model (MLLM) that integrates visual encoding and textual decoding into a single LLM. In Mono-InternVL, a set of visual experts is embedded into the pre-trained LLM via a mixture-of-experts (MoE) mechanism. By freezing the LLM, Mono-InternVL ensures that visual capabilities are optimized without compromising the pre-trained language knowledge. Based on this structure, an innovative Endogenous Visual Pretraining (EViP) is introduced to realize coarse-to-fine visual learning.

Mono-InternVL achieves superior performance compared to state-of-the-art MLLM Mini-InternVL-2B-1.5 and significantly outperforms other monolithic MLLMs, as shown in the radar chart above. Meanwhile, it achieves better deployment efficiency, with first token latency reduced by up to 67%.

Performance

Benchmark	Chameleon-7B	EVE-7B (HD)	Emu3	Mini-InternVL-2B-1-5	Mono-InternVL-2B
Type	Monolithic	Monolithic	Monolithic	Modular	Monolithic
#Activated Params	7B	7B	8B	2.2B	1.8B

MMVet	8.3	25.7	37.2	39.3	40.1
MMMU_val	25.4	32.6	31.6	34.6	33.7
MME_sum	170	1628	—	1902	1875
MMBench-EN_test	31.1	52.3	58.5	70.9	65.5
MathVista_testmini	22.3	34.2	—	41.1	45.7
SEED-Image	30.6	64.6	68.2	69.8	67.4
OCRBench	7	398	687	654	767
Hallusion-Bench	17.1	26.4	—	37.5	34.8
CCBench_dev	3.5	16.3	—	63.5	66.3
Avg_multimodal	16.1	38.9	—	54.4	55.2

TextVQA_val	4.8	56.8	64.7	70.5	72.6
SQA-I_test	47.2	64.9	89.2	84.9	93.6
GQA_test	—	62.6	60.3	61.6	59.5
DocVQA_test	1.5	53.0	76.3	85.0	80.0
AI2D_test	46.0	61.0	70.0	69.8	68.6
ChartQA_test	2.9	59.1	68.6	74.8	73.7
InfoVQA_test	5.0	25.0	43.8	55.4	43.0
Avg_VQA	17.9	54.6	67.6	71.7	70.1

Sources of the results include the original papers, our evaluation with VLMEvalKit, and OpenCompass.

Average scores are computed by normalizing each metric to a range between 0 and 100.

Please note that evaluating the same model using different testing toolkits can result in slight differences, which is normal. Updates to code versions and variations in environment and hardware can also cause minor discrepancies in results.

Inference

We provide an example code to run Mono-InternVL-2B inference using transformers.

Please use transformers==4.37.2 to ensure the model works normally.

Inference with Transformers (click to expand)

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 transformers 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


path = 'OpenGVLab/Mono-InternVL-2B'
model = AutoModel.from_pretrained(
    path,
    torch_dtype=torch.bfloat16,
    low_cpu_mem_usage=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('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
generation_config = dict(max_new_tokens=1024, do_sample=True)

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

# single-image single-round conversation
question = '<image>
Please describe the image shortly.'
response = model.chat(tokenizer, pixel_values, question, generation_config)
print(f'User: {question}
Assistant: {response}')

# single-image multi-round conversation
question = '<image>
Please describe the image in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=None, return_history=True)
print(f'User: {question}
Assistant: {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}
Assistant: {response}')

Inference with LMDeploy (click to expand)

Please install lmdeploy>=0.6.3 for Mono-InternVL support.

from lmdeploy import pipeline
from lmdeploy.vl import load_image

image = load_image('./examples/image1.jpg')
pipe = pipeline('OpenGVLab/Mono-InternVL-2B')
response = pipe(('Please describe the image shortly.', image))
print(response.text)

Supervised Finetuning

Currently we provide the supervised finetuning (S2 instruction tuning) code on the LLaVA-v1.5-mix665k dataset. For details on the dataset, please refer to LLaVA-v1.5.

Installation (click to expand)

Clone this repository:

git clone https://github.com/OpenGVLab/Mono-InternVL.git

Create a conda virtual environment and activate it:

conda create -n monointernvl python=3.9 -y
conda activate monointernvl

Install dependencies using requirements.txt:
```
pip install -r requirements.txt
```

Additional: Install flash-attn==2.5.6:

pip install flash-attn==2.5.6 --no-build-isolation

Alternatively you can compile from source:

git clone https://github.com/Dao-AILab/flash-attention.git
cd flash-attention
git checkout v2.5.6
python setup.py install

Dataset Preparation (click to expand)

LLaVA-v1.5-mix665k Dataset

Download the instruction tuning data:

mkdir playground
wget https://huggingface.co/datasets/liuhaotian/LLaVA-Instruct-150K/resolve/main/llava_v1_5_mix665k.json -P playground/

Download image datasets:

COCO: train2017
GQA: images
OCR-VQA: download script
TextVQA: train_val_images
VisualGenome: part1, part2

Organize data as follows:

playground/
├── data/
│   ├── coco/train2017/
│   ├── gqa/images/
│   ├── ocr_vqa/images/
│   ├── textvqa/train_images/
│   └── vg/
│       ├── VG_100K/
│       └── VG_100K_2/
└── llava_v1_5_mix665k.json

Custom Dataset

For custom dataset, format your data in to a JSONL file, where each entry is a dictionary organized in the following format (similar to llava_v1_5_mix665k.json):

{
    "id": "000000120375",
    "image": "coco/train2017/000000120375.jpg",
    "conversations": [
        {
            "from": "human",
            "value": "<image>
What type of vehicle is driving down the street in the image?"
        },
        {
            "from": "gpt",
            "value": "A red sports utility vehicle (SUV) is driving down the street in the image."
        },
        {
            "from": "human",
            "value": "Is the street crowded with people?"
        },
        {
            "from": "gpt",
            "value": "Yes, the street is filled with a considerable number of people, which indicates that the area is busy."
        }
        # (more turns ...)
    ]
}

Then modify the metadata file shell/data_llava_finetune.json:

{
  "name of your dataset": {
    "root": "playground/data/", # combination of "root" and "image" in the JSONL gives the complete image path
    "annotation": "path to your JSONL",
    "data_augment": false,
    "repeat_time": 1,
    "length": 12345 # change to the actual number of samples in your dataset
  }
}

Model Preparation (click to expand)

We provide pretrained models of different stages (S1.1 concept learning, S1.2 semantic learning, S1.3 alignment learning). Choose from the following models and download the weights to workdirs/ folder.

model name	download	size
Mono-InternVL-2B-S1-1	🤗 HF link	6.2 GB
Mono-InternVL-2B-S1-2	🤗 HF link	6.2 GB
Mono-InternVL-2B-S1-3	🤗 HF link	6.2 GB

mkdir workdirs
cd workdirs/
# pip install -U huggingface_hub
huggingface-cli download --resume-download --local-dir-use-symlinks False OpenGVLab/Mono-InternVL-2B-S1-1 --local-dir Mono-InternVL-2B-S1-1

The directory structure is:

workdirs/
├── Mono-InternVL-2B-S1-1/
├── Mono-InternVL-2B-S1-2/
└── Mono-InternVL-2B-S1-3/

Training (click to expand)

Finetuning takes around 12 hours on 8x A100 (80G) GPUs.

Single Node Multi-GPU

MODEL="./workdirs/Mono-InternVL-2B-S1-3" OUTPUT_DIR="./workdirs/mono_internvl_llava_sft" sh shell/mono_internvl_finetune_llava_torchrun.sh

Slurm Cluster

PARTITION="your partition" MODEL="./workdirs/Mono-InternVL-2B-S1-3" OUTPUT_DIR="./workdirs/mono_internvl_llava_sft" sh shell/mono_internvl_finetune_llava_slurm.sh

License

This project is released under the MIT License.

Citation

If you find this work helpful in your research, please consider giving this repo a star ⭐ and citing our paper:

@article{mono_internvl_v1,
  title={Mono-InternVL: Pushing the Boundaries of Monolithic Multimodal Large Language Models with Endogenous Visual Pre-training},
  author={Luo, Gen and Yang, Xue and Dou, Wenhan and Wang, Zhaokai and Liu, Jiawen and Dai, Jifeng and Qiao, Yu and Zhu, Xizhou},
  journal={arXiv preprint arXiv:2410.08202},
  year={2024}
}

@article{mono_internvl_v1.5,
  title={Mono-InternVL-1.5: Towards Cheaper and Faster Monolithic Multimodal Large Language Models},
  author={Luo, Gen and Dou, Wenhan and Li, Wenhao and Wang, Zhaokai and Yang, Xue and Tian, Changyao and Li, Hao and Wang, Weiyun and Wang, Wenhai and Zhu, Xizhou and Qiao, Yu and Dai, Jifeng},
  journal={arXiv preprint arXiv:2507.12566},
  year={2025}
}