Reproduce “SPARTA: A Sparse Ternary LLM Accelerator with Adaptive Attention and Unified Datapath for Long-Context Edge Inference” 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

Yanzhen Zhou, Zhengyu Han, Qunyou Liu, Zheyu Yan, Weihong Xu, Cheng Zhuo

SPARTA combines sparse computation in ternary model layers and attention on one compute fabric, while selecting key pages and value tokens to limit long-context cache traffic. At its default setting, it reports 4.1× less KV traffic and 5.9× fewer operations than full attention for BitCPM-3B at 8K context. Top-p operates only within Top-k-selected pages, and quality losses vary by task.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.