SparseBalance: Load-Balanced Long Context Training with Dynamic Sparse Attention

Hongtao Xu 1,2 , Jianchao Tan 2 , Yuxuan Hu 2 , Pengju Lu 1 , Hongyu Wang 1 , Pingwei Sun 2 , Yerui Sun 2 , Yuchen Xie 2 , Xunliang Cai 2 , Mingzhen Li 3,* , and Weile Jia 3,* · 2026

Read SparseBalance: Load-Balanced Long Context Training with Dynamic Sparse Attention by Hongtao Xu 1,2 , Jianchao Tan 2 , Yuxuan Hu 2 , Pengju Lu 1 , Hongyu Wang 1 , Pingwei Sun 2 , Yerui Sun 2 , Yuchen Xie 2 , Xunliang Cai 2 , Mingzhen Li 3,* , and Weile Jia 3,* (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-13847.txt.

Description

Xu and ten colleagues present SparseBalance for long-context training with dynamic sparse attention. Sparse attention eases long-context compute, but distributed training faces wild variation in sequence length and sparsity response, causing imbalance and weaker results. Their co-design tunes sparsity with two-way adjustment, removing stragglers and using idle bubbles for free gains. Sparsity-aware batching adds coarse balance. Results show up to a 1.33-fold end-to-end speedup while LongBench capability rises by 0.46 percent.

Catalog record

CREATORHongtao Xu 1,2 , Jianchao Tan 2 , Yuxuan Hu 2 , Pengju Lu 1 , Hongyu Wang 1 , Pingwei Sun 2 , Yerui Sun 2 , Yuchen Xie 2 , Xunliang Cai 2 , Mingzhen Li 3,* , and Weile Jia 3,*
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE52102
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.13847

Browse the full library · About the Tower of Babel