MoH: Multi-Head Attention as Mixture-of-Head Attention
Paper or resources for more information:
[Paper] [Code]
⚡ 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.
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}
}
