Efficient Sparse Attention needs Adaptive Token Release

Chaoran Zhang, Lixin Zou, Dan Luo, Xiangyang Luo, Zihao Li, Min Tang, Chenliang Li


Abstract
In recent years, Large Language Models (LLMs) have demonstrated remarkable capabilities across a wide array of text-centric tasks. However, their ‘large’ scale introduces significant computational and storage challenges, particularly in managing the key-value states of the transformer, which limits their wider applicability. Therefore, we propose to adaptively release resources from caches and rebuild the necessary key-value states. Particularly, we accomplish this by a lightweight controller module to approximate an ideal top-K sparse attention. This module retains the tokens with the highest top-K attention weights and simultaneously rebuilds the discarded but necessary tokens, which may become essential for future decoding. Comprehensive experiments in natural language generation and modeling reveal that our method is not only competitive with full attention in terms of performance but also achieves a significant throughput improvement of up to 221.8%. The code for replication is available on the https://github.com/WHUIR/ADORE.
Anthology ID:
2024.findings-acl.837
Volume:
Findings of the Association for Computational Linguistics: ACL 2024
Month:
August
Year:
2024
Address:
Bangkok, Thailand
Editors:
Lun-Wei Ku, Andre Martins, Vivek Srikumar
Venue:
Findings
SIG:
Publisher:
Association for Computational Linguistics
Note:
Pages:
14081–14094
Language:
URL:
https://aclanthology.org/2024.findings-acl.837
DOI:
10.18653/v1/2024.findings-acl.837
Bibkey:
Cite (ACL):
Chaoran Zhang, Lixin Zou, Dan Luo, Xiangyang Luo, Zihao Li, Min Tang, and Chenliang Li. 2024. Efficient Sparse Attention needs Adaptive Token Release. In Findings of the Association for Computational Linguistics: ACL 2024, pages 14081–14094, Bangkok, Thailand. Association for Computational Linguistics.
Cite (Informal):
Efficient Sparse Attention needs Adaptive Token Release (Zhang et al., Findings 2024)
Copy Citation:
PDF:
https://preview.aclanthology.org/add_acl24_videos/2024.findings-acl.837.pdf
Video:
 https://preview.aclanthology.org/add_acl24_videos/2024.findings-acl.837.mp4