MoE-LLaVA: Mixture of Experts for Large Vision-Language Models

If you like our project, please give us a star ⭐ on GitHub for latest update.

💡 I also have other vision-language projects that may interest you ✨.

Open-Sora Plan: Open-Source Large Video Generation Model
Bin Lin and Yunyang Ge and Xinhua Cheng and Zongjian Li and Bin Zhu and Shaodong Wang and Xianyi He and Yang Ye and Shenghai Yuan and Liuhan Chen and Tanghui Jia and Junwu Zhang and Zhenyu Tang and Yatian Pang and Bin She and Cen Yan and Zhiheng Hu and Xiaoyi Dong and Lin Chen and Zhang Pan and Xing Zhou and Shaoling Dong and Yonghong Tian and Li Yuan

Video-LLaVA: Learning United Visual Representation by Alignment Before Projection
Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, Li Yuan

LanguageBind: Extending Video-Language Pretraining to N-modality by Language-based Semantic Alignment
Bin Zhu, Bin Lin, Munan Ning, Yang Yan, Jiaxi Cui, HongFa Wang, Yatian Pang, Wenhao Jiang, Junwu Zhang, Zongwei Li, Wancai Zhang, Zhifeng Li, Wei Liu, Li Yuan

📣 News

• [2025.07.15] 🔥🔥🔥 Our MoE-LLaVA has been accepted at IEEE Transactions on Multimedia (TMM)!
• [2024.03.16] 🎉 We release all stage2 models, cheching our model zoo.
• [2024.02.03] 🎉 We release a stronger MoE-LLaVA-StableLM. The average performance is close to LLaVA-1.5-7B by using 2.0B sparse activated parameters, checking our model zoo.
• [2024.02.02] 🤝 Enjoying the and , created by @camenduru, who generously supports our research!
• [2024.02.01] 🔥 People who cannot access HF can now download the model through the model scope, checking our model zoo.
• [2024.01.30] 🔥 We release a stronger MoE-LLaVA-Phi2. The average performance surpasses LLaVA-1.5-7B by using 3.6B sparse activated parameters, checking our model zoo.
• [2024.01.27] 🤗 Hugging Face demo and all codes & datasets are available now! Welcome to watch 👀 this repository for the latest updates.

😮 Highlights

MoE-LLaVA shows excellent performance in multi-modal learning.

🔥 High performance, but with fewer parameters

• with just 3B sparsely activated parameters, MoE-LLaVA demonstrates performance comparable to the LLaVA-1.5-7B on various visual understanding datasets and even surpasses the LLaVA-1.5-13B in object hallucination benchmarks.

🚀 Simple baseline, learning multi-modal interactions with sparse pathways.

• With the addition of a simple MoE tuning stage, we can complete the training of MoE-LLaVA on 8 A100 GPUs within 1 days.

🤗 Demo

Gradio Web UI

Highly recommend trying out our web demo by the following command, which incorporates all features currently supported by MoE-LLaVA. We also provide online demo in Huggingface Spaces.

# use phi2
deepspeed --include localhost:0 moellava/serve/gradio_web_server.py --model-path "LanguageBind/MoE-LLaVA-Phi2-2.7B-4e" 
# use qwen
deepspeed --include localhost:0 moellava/serve/gradio_web_server.py --model-path "LanguageBind/MoE-LLaVA-Qwen-1.8B-4e" 
# use stablelm
deepspeed --include localhost:0 moellava/serve/gradio_web_server.py --model-path "LanguageBind/MoE-LLaVA-StableLM-1.6B-4e" 

https://github.com/PKU-YuanGroup/MoE-LLaVA/assets/62638829/8541aac6-9ef6-4fde-aa94-80d0375b9bdb

CLI Inference

# use phi2
deepspeed --include localhost:0 moellava/serve/cli.py --model-path "LanguageBind/MoE-LLaVA-Phi2-2.7B-4e"  --image-file "image.jpg"
# use qwen
deepspeed --include localhost:0 moellava/serve/cli.py --model-path "LanguageBind/MoE-LLaVA-Qwen-1.8B-4e"  --image-file "image.jpg"
# use stablelm
deepspeed --include localhost:0 moellava/serve/cli.py --model-path "LanguageBind/MoE-LLaVA-StableLM-1.6B-4e"  --image-file "image.jpg"

🐳 Model Zoo

🚨 Please know https://github.com/PKU-YuanGroup/MoE-LLaVA/issues/27.

Stage2 Model

Model	Checkpoint
MoE-LLaVA-1.6B×4-Top2	LanguageBind/MoE-LLaVA-StableLM-Stage2
MoE-LLaVA-1.6B×4-Top2-384	LanguageBind/MoE-LLaVA-StableLM-Stage2-384
MoE-LLaVA-1.8B×4-Top2	LanguageBind/MoE-LLaVA-Qwen-Stage2
MoE-LLaVA-2.7B×4-Top2	LanguageBind/MoE-LLaVA-Phi2-Stage2
MoE-LLaVA-2.7B×4-Top2-384	LanguageBind/MoE-LLaVA-Phi2-Stage2-384

Pretrain Model

Model	Checkpoint
MoE-LLaVA-1.6B×4-Top2	LanguageBind/MoE-LLaVA-StableLM-Pretrain
MoE-LLaVA-1.6B×4-Top2-384	LanguageBind/MoE-LLaVA-StableLM-384-Pretrain
MoE-LLaVA-1.8B×4-Top2	LanguageBind/MoE-LLaVA-Qwen-Pretrain
MoE-LLaVA-2.7B×4-Top2	LanguageBind/MoE-LLaVA-Phi2-Pretrain
MoE-LLaVA-2.7B×4-Top2-384	LanguageBind/MoE-LLaVA-Phi2-384-Pretrain

⚙️ Requirements and Installation

We recommend the requirements as follows.

• Python == 3.10
• Pytorch == 2.0.1
• CUDA Version >= 11.7
• Transformers == 4.37.0
• Tokenizers==0.15.1
• Install required packages:

git clone https://github.com/PKU-YuanGroup/MoE-LLaVA
cd MoE-LLaVA
conda create -n moellava python=3.10 -y
conda activate moellava
pip install --upgrade pip  # enable PEP 660 support
pip install -e .
pip install -e ".[train]"
pip install flash-attn --no-build-isolation

# Below are optional. For Qwen model.
git clone https://github.com/Dao-AILab/flash-attention
cd flash-attention && pip install .
# Below are optional. Installing them might be slow.
# pip install csrc/layer_norm
# If the version of flash-attn is higher than 2.1.1, the following is not needed.
# pip install csrc/rotary

[!Warning]

🚨 We find that using flash attention2 makes performance degradation.

🗝️ Training & Validating

The training & validating instruction is in TRAIN.md & EVAL.md.

💡 Customizing your MoE-LLaVA

The instruction is in CUSTOM.md.

😍 Visualization

The instruction is in VISUALIZATION.md.

🤖 API

We open source all codes. If you want to load the model (e.g. LanguageBind/MoE-LLaVA-Phi2-2.7B-4e) on local, you can use the following code snippets.

Using the following command to run the code.

deepspeed --include localhost:0 predict.py

import torch
from PIL import Image
from moellava.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN
from moellava.conversation import conv_templates, SeparatorStyle
from moellava.model.builder import load_pretrained_model
from moellava.utils import disable_torch_init
from moellava.mm_utils import tokenizer_image_token, get_model_name_from_path, KeywordsStoppingCriteria

def main():
    disable_torch_init()
    image = 'moellava/serve/examples/extreme_ironing.jpg'
    inp = 'What is unusual about this image?'
    model_path = 'LanguageBind/MoE-LLaVA-Phi2-2.7B-4e'  # LanguageBind/MoE-LLaVA-Qwen-1.8B-4e or LanguageBind/MoE-LLaVA-StableLM-1.6B-4e
    device = 'cuda'
    load_4bit, load_8bit = False, False  # FIXME: Deepspeed support 4bit or 8bit?
    model_name = get_model_name_from_path(model_path)
    tokenizer, model, processor, context_len = load_pretrained_model(model_path, None, model_name, load_8bit, load_4bit, device=device)
    image_processor = processor['image']
    conv_mode = "phi"  # qwen or stablelm
    conv = conv_templates[conv_mode].copy()
    roles = conv.roles
    image_tensor = image_processor.preprocess(Image.open(image).convert('RGB'), return_tensors='pt')['pixel_values'].to(model.device, dtype=torch.float16)

    print(f"{roles[1]}: {inp}")
    inp = DEFAULT_IMAGE_TOKEN + '\n' + inp
    conv.append_message(conv.roles[0], inp)
    conv.append_message(conv.roles[1], None)
    prompt = conv.get_prompt()
    input_ids = tokenizer_image_token(prompt, tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt').unsqueeze(0).cuda()
    stop_str = conv.sep if conv.sep_style != SeparatorStyle.TWO else conv.sep2
    keywords = [stop_str]
    stopping_criteria = KeywordsStoppingCriteria(keywords, tokenizer, input_ids)

    with torch.inference_mode():
        output_ids = model.generate(
            input_ids,
            images=image_tensor,
            do_sample=True,
            temperature=0.2,
            max_new_tokens=1024,
            use_cache=True,
            stopping_criteria=[stopping_criteria])

    outputs = tokenizer.decode(output_ids[0, input_ids.shape[1]:], skip_special_tokens=True).strip()
    print(outputs)

if __name__ == '__main__':
    main()

🙌 Related Projects

• Video-LLaVA This framework empowers the model to efficiently utilize the united visual tokens.
• LanguageBind An open source five modalities language-based retrieval framework.

👍 Acknowledgement

• LLaVA The codebase we built upon and it is an efficient large language and vision assistant.

🔒 License

• The majority of this project is released under the Apache 2.0 license as found in the LICENSE file.
• The service is a research preview intended for non-commercial use only, subject to the model License of LLaMA, Terms of Use of the data generated by OpenAI, and Privacy Practices of ShareGPT. Please contact us if you find any potential violation.

✏️ Citation

If you find our paper and code useful in your research, please consider giving a star :star: and citation :pencil:.

@article{lin2024moe,
  title={MoE-LLaVA: Mixture of Experts for Large Vision-Language Models},
  author={Lin, Bin and Tang, Zhenyu and Ye, Yang and Cui, Jiaxi and Zhu, Bin and Jin, Peng and Zhang, Junwu and Ning, Munan and Yuan, Li},
  journal={arXiv preprint arXiv:2401.15947},
  year={2024}
}

@article{lin2023video,
  title={Video-LLaVA: Learning United Visual Representation by Alignment Before Projection},
  author={Lin, Bin and Zhu, Bin and Ye, Yang and Ning, Munan and Jin, Peng and Yuan, Li},
  journal={arXiv preprint arXiv:2311.10122},
  year={2023}
}

✨ Star History

🤝 Contributors