BLASST: Dynamic BLocked Attention Sparsity via Softmax Thresholding

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Yuan, Jiayi, Shinn, Cameron, Xu, Kai, Cui, Jingze, Klimiashvili, George, Xiao, Guangxuan, Zheng, Perkz, Li, Bo, Zhou, Yuxin, Ye, Zhouhai, You, Weijie, Zheng, Tian, Brown, Dominic, Wang, Pengbo, Hoehnerbach, Markus, Cai, Richard, Demouth, Julien, Owens, John D., Hu, Xia, Han, Song, Liu, Timmy, Mao, Huizi
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866908996549476352
author Yuan, Jiayi
Shinn, Cameron
Xu, Kai
Cui, Jingze
Klimiashvili, George
Xiao, Guangxuan
Zheng, Perkz
Li, Bo
Zhou, Yuxin
Ye, Zhouhai
You, Weijie
Zheng, Tian
Brown, Dominic
Wang, Pengbo
Hoehnerbach, Markus
Cai, Richard
Demouth, Julien
Owens, John D.
Hu, Xia
Han, Song
Liu, Timmy
Mao, Huizi
author_facet Yuan, Jiayi
Shinn, Cameron
Xu, Kai
Cui, Jingze
Klimiashvili, George
Xiao, Guangxuan
Zheng, Perkz
Li, Bo
Zhou, Yuxin
Ye, Zhouhai
You, Weijie
Zheng, Tian
Brown, Dominic
Wang, Pengbo
Hoehnerbach, Markus
Cai, Richard
Demouth, Julien
Owens, John D.
Hu, Xia
Han, Song
Liu, Timmy
Mao, Huizi
contents The growing demand for long-context inference capabilities in Large Language Models (LLMs) has intensified the computational and memory bottlenecks inherent to the self-attention mechanism. To address this challenge, we introduce BLASST, a drop-in, dynamic sparse attention mechanism that accelerates inference by using only a fixed scalar threshold to skip attention blocks. Our method targets practical inference deployment by removing the barriers to adoption present in existing works. As such, BLASST eliminates training requirements, avoids expensive pre-computation passes, accelerates both prefill and decode across all major attention variants (MHA, GQA, MQA, and MLA), provides optimized support for modern hardware, and easily integrates into existing frameworks. This is achieved by reusing online softmax statistics to identify negligible attention scores, skipping softmax, value block loads, and the subsequent matrix multiplication. We demonstrate the BLASST algorithm by delivering optimized kernels with negligible latency overhead. Our automated threshold calibration procedure reveals a simple inverse relationship between optimal threshold and context length, meaning we require only a single threshold each for prefill and decode per model. Preserving benchmark accuracy, we demonstrate a 1.52x speedup for prefill at 71.9% sparsity and a 1.48x speedup for decode at 73.2% sparsity on modern GPUs.
format Preprint
id arxiv_https___arxiv_org_abs_2512_12087
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle BLASST: Dynamic BLocked Attention Sparsity via Softmax Thresholding
Yuan, Jiayi
Shinn, Cameron
Xu, Kai
Cui, Jingze
Klimiashvili, George
Xiao, Guangxuan
Zheng, Perkz
Li, Bo
Zhou, Yuxin
Ye, Zhouhai
You, Weijie
Zheng, Tian
Brown, Dominic
Wang, Pengbo
Hoehnerbach, Markus
Cai, Richard
Demouth, Julien
Owens, John D.
Hu, Xia
Han, Song
Liu, Timmy
Mao, Huizi
Computation and Language
The growing demand for long-context inference capabilities in Large Language Models (LLMs) has intensified the computational and memory bottlenecks inherent to the self-attention mechanism. To address this challenge, we introduce BLASST, a drop-in, dynamic sparse attention mechanism that accelerates inference by using only a fixed scalar threshold to skip attention blocks. Our method targets practical inference deployment by removing the barriers to adoption present in existing works. As such, BLASST eliminates training requirements, avoids expensive pre-computation passes, accelerates both prefill and decode across all major attention variants (MHA, GQA, MQA, and MLA), provides optimized support for modern hardware, and easily integrates into existing frameworks. This is achieved by reusing online softmax statistics to identify negligible attention scores, skipping softmax, value block loads, and the subsequent matrix multiplication. We demonstrate the BLASST algorithm by delivering optimized kernels with negligible latency overhead. Our automated threshold calibration procedure reveals a simple inverse relationship between optimal threshold and context length, meaning we require only a single threshold each for prefill and decode per model. Preserving benchmark accuracy, we demonstrate a 1.52x speedup for prefill at 71.9% sparsity and a 1.48x speedup for decode at 73.2% sparsity on modern GPUs.
title BLASST: Dynamic BLocked Attention Sparsity via Softmax Thresholding
topic Computation and Language
url https://arxiv.org/abs/2512.12087