Enregistré dans:
Détails bibliographiques
Auteurs principaux: Gao, Yizhao, Wei, Jianyu, Zhang, Qihao, Cheng, Yu, Chen, Shimao, Tang, Zhengju, Jiang, Zihan, Song, Yifan, Zhang, Hailin, Zhao, Liang, Yang, Bo, Wang, Gang, Cao, Shijie, Luo, Fuli
Format: Preprint
Publié: 2026
Sujets:
Accès en ligne:https://arxiv.org/abs/2602.03560
Tags: Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
_version_ 1866910010460602368
author Gao, Yizhao
Wei, Jianyu
Zhang, Qihao
Cheng, Yu
Chen, Shimao
Tang, Zhengju
Jiang, Zihan
Song, Yifan
Zhang, Hailin
Zhao, Liang
Yang, Bo
Wang, Gang
Cao, Shijie
Luo, Fuli
author_facet Gao, Yizhao
Wei, Jianyu
Zhang, Qihao
Cheng, Yu
Chen, Shimao
Tang, Zhengju
Jiang, Zihan
Song, Yifan
Zhang, Hailin
Zhao, Liang
Yang, Bo
Wang, Gang
Cao, Shijie
Luo, Fuli
contents This work introduces Hybrid Sparse Attention (HySparse), a new architecture that interleaves each full attention layer with several sparse attention layers. While conceptually simple, HySparse strategically derives each sparse layer's token selection and KV caches directly from the preceding full attention layer. This architecture resolves two fundamental limitations of prior sparse attention methods. First, conventional approaches typically rely on additional proxies to predict token importance, introducing extra complexity and potentially suboptimal performance. In contrast, HySparse uses the full attention layer as a precise oracle to identify important tokens. Second, existing sparse attention designs often reduce computation without saving KV cache. HySparse enables sparse attention layers to reuse the full attention KV cache, thereby reducing both computation and memory. We evaluate HySparse on both 7B dense and 80B MoE models. Across all settings, HySparse consistently outperforms both full attention and hybrid SWA baselines. Notably, in the 80B MoE model with 49 total layers, only 5 layers employ full attention, yet HySparse achieves substantial performance gains while reducing KV cache storage by nearly 10x.
format Preprint
id arxiv_https___arxiv_org_abs_2602_03560
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle HySparse: A Hybrid Sparse Attention Architecture with Oracle Token Selection and KV Cache Sharing
Gao, Yizhao
Wei, Jianyu
Zhang, Qihao
Cheng, Yu
Chen, Shimao
Tang, Zhengju
Jiang, Zihan
Song, Yifan
Zhang, Hailin
Zhao, Liang
Yang, Bo
Wang, Gang
Cao, Shijie
Luo, Fuli
Computation and Language
Artificial Intelligence
This work introduces Hybrid Sparse Attention (HySparse), a new architecture that interleaves each full attention layer with several sparse attention layers. While conceptually simple, HySparse strategically derives each sparse layer's token selection and KV caches directly from the preceding full attention layer. This architecture resolves two fundamental limitations of prior sparse attention methods. First, conventional approaches typically rely on additional proxies to predict token importance, introducing extra complexity and potentially suboptimal performance. In contrast, HySparse uses the full attention layer as a precise oracle to identify important tokens. Second, existing sparse attention designs often reduce computation without saving KV cache. HySparse enables sparse attention layers to reuse the full attention KV cache, thereby reducing both computation and memory. We evaluate HySparse on both 7B dense and 80B MoE models. Across all settings, HySparse consistently outperforms both full attention and hybrid SWA baselines. Notably, in the 80B MoE model with 49 total layers, only 5 layers employ full attention, yet HySparse achieves substantial performance gains while reducing KV cache storage by nearly 10x.
title HySparse: A Hybrid Sparse Attention Architecture with Oracle Token Selection and KV Cache Sharing
topic Computation and Language
Artificial Intelligence
url https://arxiv.org/abs/2602.03560