Reproduce “Multi-layer Indexer: Overlapping Selection and Attention in DeepSeek Sparse Attention” 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
Serdarcan Dilbaz, Arjun Krishna, Kiran Kamble, Daniel M. Bikel
Multi-layer Indexer (MLI) overlaps long-context decoding work by letting each layer attend to the preceding indexer’s selection while computing its own selection. It does not remove scans or reuse selections across decode steps. Combined with IndexCache, it reaches 1.32× base per-request throughput at 128K on DeepSeek-V3.2, versus 1.28× for IndexCache alone; paired quality intervals span zero for this pair.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.