Reproduce “INT-FP-QSim: Mixed Precision and Formats For Large Language Models and Vision Transformers” 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
Lakshmi Nair, Mikhail Bernadskiy, Arulselvan Madhavan, Craig Chan, Ayon Basumallik, Darius Bunandar
INT-FP-QSim simulates how neural networks tolerate different low-precision formats for weights and activations. It replaces matrix-multiplication layers with configurable quantizers, including Adaptive Block Floating Point, which scales short vectors by their maximum magnitude. On OPT 125M Wikitext2, W4A4 achieved perplexity 33.14 versus 1130 for static MSE calibration and 25.94 for FP32; this measures accuracy, not hardware efficiency.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.