Reproduce “DexAttention: Distributed Sparse Attention through Rapid Online Planning and Persistent GPU Execution” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Sirui Chen, Xuehai Qian
DexAttention reduces data movement in sparse distributed attention by assigning query tiles to GPUs that share key/value inputs and scheduling transfers and computation with a persistent kernel that dispatches work as dependencies arrive. Across 11 configurations on four nodes and 16 GPUs, it achieves a 1.39× geometric-mean end-to-end training speedup over the fastest applicable baseline; reassignment does not always help.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.