Saved in:
Bibliographic Details
Main Authors: Lu, Enzhe, Jiang, Zhejun, Liu, Jingyuan, Du, Yulun, Jiang, Tao, Hong, Chao, Liu, Shaowei, He, Weiran, Yuan, Enming, Wang, Yuzhi, Huang, Zhiqi, Yuan, Huan, Xu, Suting, Xu, Xinran, Lai, Guokun, Chen, Yanru, Zheng, Huabin, Yan, Junjie, Su, Jianlin, Wu, Yuxin, Zhang, Neo Y., Yang, Zhilin, Zhou, Xinyu, Zhang, Mingxing, Qiu, Jiezhong
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2502.13189
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866910835413090304
author Lu, Enzhe
Jiang, Zhejun
Liu, Jingyuan
Du, Yulun
Jiang, Tao
Hong, Chao
Liu, Shaowei
He, Weiran
Yuan, Enming
Wang, Yuzhi
Huang, Zhiqi
Yuan, Huan
Xu, Suting
Xu, Xinran
Lai, Guokun
Chen, Yanru
Zheng, Huabin
Yan, Junjie
Su, Jianlin
Wu, Yuxin
Zhang, Neo Y.
Yang, Zhilin
Zhou, Xinyu
Zhang, Mingxing
Qiu, Jiezhong
author_facet Lu, Enzhe
Jiang, Zhejun
Liu, Jingyuan
Du, Yulun
Jiang, Tao
Hong, Chao
Liu, Shaowei
He, Weiran
Yuan, Enming
Wang, Yuzhi
Huang, Zhiqi
Yuan, Huan
Xu, Suting
Xu, Xinran
Lai, Guokun
Chen, Yanru
Zheng, Huabin
Yan, Junjie
Su, Jianlin
Wu, Yuxin
Zhang, Neo Y.
Yang, Zhilin
Zhou, Xinyu
Zhang, Mingxing
Qiu, Jiezhong
contents Scaling the effective context length is essential for advancing large language models (LLMs) toward artificial general intelligence (AGI). However, the quadratic increase in computational complexity inherent in traditional attention mechanisms presents a prohibitive overhead. Existing approaches either impose strongly biased structures, such as sink or window attention which are task-specific, or radically modify the attention mechanism into linear approximations, whose performance in complex reasoning tasks remains inadequately explored. In this work, we propose a solution that adheres to the ``less structure'' principle, allowing the model to determine where to attend autonomously, rather than introducing predefined biases. We introduce Mixture of Block Attention (MoBA), an innovative approach that applies the principles of Mixture of Experts (MoE) to the attention mechanism. This novel architecture demonstrates superior performance on long-context tasks while offering a key advantage: the ability to seamlessly transition between full and sparse attention, enhancing efficiency without the risk of compromising performance. MoBA has already been deployed to support Kimi's long-context requests and demonstrates significant advancements in efficient attention computation for LLMs. Our code is available at https://github.com/MoonshotAI/MoBA.
format Preprint
id arxiv_https___arxiv_org_abs_2502_13189
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle MoBA: Mixture of Block Attention for Long-Context LLMs
Lu, Enzhe
Jiang, Zhejun
Liu, Jingyuan
Du, Yulun
Jiang, Tao
Hong, Chao
Liu, Shaowei
He, Weiran
Yuan, Enming
Wang, Yuzhi
Huang, Zhiqi
Yuan, Huan
Xu, Suting
Xu, Xinran
Lai, Guokun
Chen, Yanru
Zheng, Huabin
Yan, Junjie
Su, Jianlin
Wu, Yuxin
Zhang, Neo Y.
Yang, Zhilin
Zhou, Xinyu
Zhang, Mingxing
Qiu, Jiezhong
Machine Learning
Artificial Intelligence
Computation and Language
Scaling the effective context length is essential for advancing large language models (LLMs) toward artificial general intelligence (AGI). However, the quadratic increase in computational complexity inherent in traditional attention mechanisms presents a prohibitive overhead. Existing approaches either impose strongly biased structures, such as sink or window attention which are task-specific, or radically modify the attention mechanism into linear approximations, whose performance in complex reasoning tasks remains inadequately explored. In this work, we propose a solution that adheres to the ``less structure'' principle, allowing the model to determine where to attend autonomously, rather than introducing predefined biases. We introduce Mixture of Block Attention (MoBA), an innovative approach that applies the principles of Mixture of Experts (MoE) to the attention mechanism. This novel architecture demonstrates superior performance on long-context tasks while offering a key advantage: the ability to seamlessly transition between full and sparse attention, enhancing efficiency without the risk of compromising performance. MoBA has already been deployed to support Kimi's long-context requests and demonstrates significant advancements in efficient attention computation for LLMs. Our code is available at https://github.com/MoonshotAI/MoBA.
title MoBA: Mixture of Block Attention for Long-Context LLMs
topic Machine Learning
Artificial Intelligence
Computation and Language
url https://arxiv.org/abs/2502.13189