File size: 3,875 Bytes
7d2c4e2 c94e105 7d2c4e2 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 |
---
license: apache-2.0
---
# MoH: Multi-Head Attention as Mixture-of-Head Attention
**Paper or resources for more information:**
[[Paper](https://huggingface.co/papers/2410.11842)] [[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}
}
``` |