LINs-lab/DynMoE-Qwen-1.8B

Dynamic Mixture of Experts: An Auto-Tuning Approach for Efficient Transformer Models

If our project helps you, please give us a star ⭐ on GitHub and cite our paper!

📰 News

[2025.01.23]: 🎉 Our paper is accepted to ICLR 2025!
[2024.05.25] 🔥 Our checkpoints are available now!
[2024.05.23] 🔥 Our paper is released!

😎 What's Interesting?

Dynamic Mixture of Experts (DynMoE) incorporates (1) a novel gating method that enables each token to automatically determine the number of experts to activate. (2) An adaptive process automatically adjusts the number of experts during training.

Top-Any Gating

Adaptive Training Process

💡 Model Details

🤔 DynMoE-Qwen is a MoE model with dynamic top-k gating, finetuned on LanguageBind/MoE-LLaVA-Qwen-Stage2.
🚀 Our DynMoE-Qwen-1.8B has totally 3.1B parameters, but only 2.2B are activated! (average top-k = 1.86)
⌛ With the DynMoE tuning stage, we can complete training on 8 A100 GPUs within 40 hours.

👍 Acknowledgement

We are grateful for the following awesome projects:

🔒 License

This project is released under the Apache-2.0 license as found in the LICENSE file.

✏️ Citation

@misc{guo2024dynamic,
      title={Dynamic Mixture of Experts: An Auto-Tuning Approach for Efficient Transformer Models}, 
      author={Yongxin Guo and Zhenglin Cheng and Xiaoying Tang and Tao Lin},
      year={2024},
      eprint={2405.14297},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

LINs-lab
/

DynMoE-Qwen-1.8B