---
license: apache-2.0
---
# MoH: Multi-Head Attention as Mixture-of-Head Attention

**Paper or resources for more information:**
[[Paper]()] [[Code](https://github.com/SkyworkAI/MoH)]

## ⚡ Overview
We propose Mixture-of-Head attention (MoH), a new architecture that treats attention heads as experts in the Mixture-of-Experts (MoE) mechanism. MoH has two significant advantages:
* First, MoH enables each token to select the appropriate attention heads, enhancing inference efficiency without compromising accuracy or increasing the number of parameters. 
* Second, MoH replaces the standard summation in multi-head attention with a weighted summation, introducing flexibility to the attention mechanism and unlocking extra performance potential.



## 😮 Highlights
### 💡 General Framework
We evaluate our proposed MoH across various popular model frameworks, including Vision Transformers (ViT) for image classification, Diffusion models with Transformers (DiT) for class-conditional image generation, and Large Language Models (LLMs) for language tasks.

<div align=center>

|                   Code                    |                                                                                                                         HuggingFace Model                                                                                                                         |  
|:-----------------------------------------:|:-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------:|
|     **[MoH-ViT](https://github.com/SkyworkAI/MoH/tree/main/MoH-ViT)**      | 🤗 [MoH-ViT-B-75](https://huggingface.co/Chat-UniVi/MoH-ViT-B-75), [MoH-ViT-B-50](https://huggingface.co/Chat-UniVi/MoH-ViT-B-50), [MoH-ViT-S-80](https://huggingface.co/Chat-UniVi/MoH-ViT-S-80), [MoH-ViT-S-75](https://huggingface.co/Chat-UniVi/MoH-ViT-S-75) |
|     **[MoH-DiT](https://github.com/SkyworkAI/MoH/tree/main/MoH-DiT)**      |                                                                                                 😊 [MoH-DiT-90](https://huggingface.co/Chat-UniVi/MoH-DiT-XL-90)                                                                                                  | 
| **[MoH-LLaMA3-8B](https://github.com/SkyworkAI/MoH/tree/main/MoH-LLaMA3)** |                                                                                                                        😊 [MoH-LLaMA3-8B](https://huggingface.co/Chat-UniVi/MoH-LLaMA3-8B)                                                                                                                         | 

</div>

### 🔥 High Performance
Extensive experiments on ViT, DiT, and LLMs demonstrate that MoH outperforms multi-head attention by using only **50%~90%** of the attention heads.

### 🤗 Support Continue-Tuning Starting from the Multi-Head Attention Models
we demonstrate that pre-trained multi-head attention models, such as LLaMA3-8B, can be further continue-tuned into our MoH models. Notably, MoH-LLaMA3-8B achieves an average accuracy of 64.0% across 14 benchmarks, outperforming LLaMA3-8B by 2.4% by utilizing only 75% of the attention heads.


The MoH model quickly recovers to over **95%** of the performance of the original model within a training budget of 10B tokens. Then, the performance gradually improves with the increase of the training tokens.


## ✏️ Citation
If you find this paper useful, please consider staring 🌟 this repo and citing 📑 our paper:
```
@article{jin2024moh,
  title={MoH: Multi-Head Attention as Mixture-of-Head Attention}, 
  author={Peng Jin and Bo Zhu and Li Yuan and Shuicheng Yan},
  year={2024}
}
```