DynMoE-StableLM-1.6B

<h2 align="center"> <a href="https://arxiv.org/abs/2405.14297">Dynamic Mixture of Experts: An Auto-Tuning Approach for Efficient Transformer Models</a></h2> <h5 align="center"> If our project helps you, please give us a star ⭐ on <a href="https://github.com/LINs-lab/DynMoE">GitHub</a> and cite our paper!</h2> <h5 align="center">

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!

Why Do We Need DynMoE?

Sparse MoE (SMoE) has an unavoidable drawback: the performance of SMoE heavily relies on the choice of hyper-parameters, such as the number of activated experts per token (top-k) and the number of experts.

Also, identifying the optimal hyper-parameter without a sufficient number of ablation studies is challenging. As the size of the models continues to grow, this limitation could result in a significant waste of computational resources, and in turn, could hinder the efficiency of training MoE-based models in practice.

Now, our DynMoE addresses these challenges through the two components introduced in Dynamic Mixture of Experts (DynMoE).

Dynamic Mixture of Experts (DynMoE)

Top-Any Gating

We first introduce a novel gating method that enables each token to automatically determine the number of experts to activate.

Adaptive Training Process

Our method also includes an adaptive process automatically adjusts the number of experts during training.

Can We Trust DynMoE? Yes!

• On language tasks, DynMoE surpasses the average performance among various MoE settings.
• Effectiveness of DynMoE remains consistent in both Vision and Vision-Language tasks.
• Although sparsity is not enforced in DynMoE, it maintains efficiency by activating even less parameters!

Model Zoo

Directory Specification

Experiment Code

• EMoE/ contains experiments on language and vision tasks, which uses tutel-based DynMoE.
• MoE-LLaVA/ contains experiments on language-vision tasks, which uses deepspeed-0.9.5-based DynMoE.

DynMoE Implementations

• Deepspeed/ provides DynMoE-Deepspeed implementation. (Recommend)
• EMoE/tutel/ provides DynMoE-Tutel implementation.

Environment Setup

Please refer to instructions under EMoE/ and MoE-LLaVA/.

Usage

Tutel Examples

Please refer to EMoE/Language/README.md and EMoE/Language/Vision.md.

DeepSpeed Examples (Recommend)

We give a minimal example to train DynMoE-ViT on ImageNet-1K from scratch at Examples/DeepSpeed-MoE.

• Check Examples/DeepSpeed-MoE/dynmoe_vit.py for how to use DynMoE in model implementation.
• Check Examples/DeepSpeed-MoE/train.py for how to train model with DynMoE.

Acknowledgement

We are grateful for the following awesome projects:

• tutel
• DeepSpeed
• GMoE
• EMoE
• MoE-LLaVA
• GLUE-X

Citation

If you find this project helpful, please consider citing our work:

@article{guo2024dynamic,
  title={Dynamic Mixture of Experts: An Auto-Tuning Approach for Efficient Transformer Models},
  author={Guo, Yongxin and Cheng, Zhenglin and Tang, Xiaoying and Lin, Tao},
  journal={arXiv preprint arXiv:2405.14297},
  year={2024}
}

Star History

Code: https://github.com/LINs-lab/DynMoE