Reproduce “Training with Fewer Bits: Unlocking Edge LLMs Training with Stochastic Rounding” 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
Taowen Liu, Marta Andronic, Deniz Gündüz, George A. Constantinides
The paper studies low-precision neural-network training using stochastic rounding and larger mini-batches to reduce per-sample rounding noise. Under independent elementwise rounding and i.i.d. samples, the authors show that this variance component declines as 1/batch size, while shared quantized weights can retain bias. On GSM8K, Llama-3.2-3B E4M1 accuracy rose from 0.290 to 0.390 as batch size increased from 8 to 32.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.