--- 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](https://huggingface.co/internlm/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](https://huggingface.co/papers/2507.12566). 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**](https://internvl.github.io/blog/2024-10-10-Mono-InternVL/) and [**GitHub repository**](https://github.com/OpenGVLab/mono-internvl) 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 | | MMMUval | 25.4 | 32.6 | 31.6 | 34.6 | 33.7 | | MMEsum | 170 | 1628 | — | 1902 | 1875 | | MMBench-ENtest | 31.1 | 52.3 | 58.5 | 70.9 | 65.5 | | MathVistatestmini | 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 | | CCBenchdev | 3.5 | 16.3 | — | 63.5 | 66.3 | | Avgmultimodal | 16.1 | 38.9 | — | 54.4 | 55.2 | | | | | | | | | TextVQAval | 4.8 | 56.8 | 64.7 | 70.5 | 72.6 | | SQA-Itest | 47.2 | 64.9 | 89.2 | 84.9 | 93.6 | | GQAtest | — | 62.6 | 60.3 | 61.6 | 59.5 | | DocVQAtest | 1.5 | 53.0 | 76.3 | 85.0 | 80.0 | | AI2Dtest | 46.0 | 61.0 | 70.0 | 69.8 | 68.6 | | ChartQAtest | 2.9 | 59.1 | 68.6 | 74.8 | 73.7 | | InfoVQAtest | 5.0 | 25.0 | 43.8 | 55.4 | 43.0 | | AvgVQA | 17.9 | 54.6 | 67.6 | 71.7 | 70.1 | > * Sources of the results include the original papers, our evaluation with [VLMEvalKit](https://github.com/open-compass/VLMEvalKit), and [OpenCompass](https://rank.opencompass.org.cn/leaderboard-multimodal/?m=REALTIME). > * 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) ```python 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 = ' 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 = ' 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. ```python 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](https://github.com/haotian-liu/LLaVA).
Installation (click to expand) - Clone this repository: ```bash git clone https://github.com/OpenGVLab/Mono-InternVL.git ``` - Create a conda virtual environment and activate it: ```bash conda create -n monointernvl python=3.9 -y conda activate monointernvl ``` - Install dependencies using `requirements.txt`: ```bash pip install -r requirements.txt ``` - Additional: Install `flash-attn==2.5.6`: ```bash pip install flash-attn==2.5.6 --no-build-isolation ``` Alternatively you can compile from source: ```bash 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 1. Download the instruction tuning data: ```sh mkdir playground wget https://huggingface.co/datasets/liuhaotian/LLaVA-Instruct-150K/resolve/main/llava_v1_5_mix665k.json -P playground/ ``` 2. Download image datasets: - COCO: [train2017](http://images.cocodataset.org/zips/train2017.zip) - GQA: [images](https://downloads.cs.stanford.edu/nlp/data/gqa/images.zip) - OCR-VQA: [download script](https://drive.google.com/drive/folders/1_GYPY5UkUy7HIcR0zq3ZCFgeZN7BAfm_?usp=sharing) - TextVQA: [train_val_images](https://dl.fbaipublicfiles.com/textvqa/images/train_val_images.zip) - VisualGenome: [part1](https://cs.stanford.edu/people/rak248/VG_100K_2/images.zip), [part2](https://cs.stanford.edu/people/rak248/VG_100K_2/images2.zip) 3. Organize data as follows: ```none 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`): ```python { "id": "000000120375", "image": "coco/train2017/000000120375.jpg", "conversations": [ { "from": "human", "value": " 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`: ```python { "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](https://huggingface.co/OpenGVLab/Mono-InternVL-2B-S1-1) | 6.2 GB | | Mono-InternVL-2B-S1-2 | 🤗 [HF link](https://huggingface.co/OpenGVLab/Mono-InternVL-2B-S1-2) | 6.2 GB | | Mono-InternVL-2B-S1-3 | 🤗 [HF link](https://huggingface.co/OpenGVLab/Mono-InternVL-2B-S1-3) | 6.2 GB | ```sh 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: ```sh 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 ```sh MODEL="./workdirs/Mono-InternVL-2B-S1-3" OUTPUT_DIR="./workdirs/mono_internvl_llava_sft" sh shell/mono_internvl_finetune_llava_torchrun.sh ``` #### Slurm Cluster ```sh 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](LICENSE). ## Citation If you find this work helpful in your research, please consider giving this repo a star ⭐ and citing our paper: ```bibtex @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} } ```