Back to Search Start Over

SampleAttention: Near-Lossless Acceleration of Long Context LLM Inference with Adaptive Structured Sparse Attention

Authors :
Zhu, Qianchao
Duan, Jiangfei
Chen, Chang
Liu, Siran
Li, Xiuhong
Feng, Guanyu
Lv, Xin
Cao, Huanqi
Chuanfu, Xiao
Zhang, Xingcheng
Lin, Dahua
Yang, Chao
Publication Year :
2024

Abstract

Large language models (LLMs) now support extremely long context windows, but the quadratic complexity of vanilla attention results in significantly long Time-to-First-Token (TTFT) latency. Existing approaches to address this complexity require additional pretraining or finetuning, and often sacrifice model accuracy. In this paper, we first provide both theoretical and empirical foundations for near-lossless sparse attention. We find dynamically capturing head-specific sparse patterns at runtime with low overhead is crucial. To address this, we propose SampleAttention, an adaptive structured and near-lossless sparse attention. Leveraging observed significant sparse patterns, SampleAttention attends to a fixed percentage of adjacent tokens to capture local window patterns, and employs a two-stage query-guided key-value filtering approach, which adaptively select a minimum set of key-values with low overhead, to capture column stripe patterns. Comprehensive evaluations show that SampleAttention can seamlessly replace vanilla attention in off-the-shelf LLMs with nearly no accuracy loss, and reduces TTFT by up to $2.42\times$ compared with FlashAttention.

Details

Database :
arXiv
Publication Type :
Report
Accession number :
edsarx.2406.15486
Document Type :
Working Paper