Grouped-query attention (GQA) has been widely adopted in LLMs to mitigate the complexity of multi-head attention (MHA). To transform an MHA to a GQA, neighbour queries in MHA are evenly split into groups where each group shares the value and key layers. In this work, we propose AsymGQA, an activation-informed approach to asymmetrically grouping an MHA to a GQA for better model performance. Our AsymGQA outperforms the GQA within the same model size budget. For example, AsymGQA LLaMA-2-7B has an accuracy increase of 7.5% on MMLU compared to neighbour grouping. Our approach addresses the GQA's trade-off problem between model performance and hardware efficiency.
@article{arxiv.2406.14963,
title = {Optimised Grouped-Query Attention Mechanism for Transformers},
author = {Yuang Chen and Cheng Zhang and Xitong Gao and Robert D. Mullins and George A. Constantinides and Yiren Zhao},
journal= {arXiv preprint arXiv:2406.14963},
year = {2024}
}