Reproduce “Accelerating LLM Decoding with In-DRAM Parallelism” 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
Kevin Xu, Tatsuya Kubo, Daichi Tokuda, Shinya Takamaeda-Yamazaki, Issei Sato
PUD-LLM restructures batch-one LLM decoding by computing a provisional next-projection result from residual state, then correcting it when an update arrives. It retains partial results in DRAM and schedules work to use spare columns or independent banks. In simulation, Gemma 12B on modified PUD achieves up to 1.41× lower decoding latency than the PUD baseline, when resources and overlap permit.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.