Reproduce “MACaQ: Fast, Model Agnostic Calibrated Quantization for Low-Precision Weights and Activations” 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

Lorenz K. Muller, Philippe Bich, Chiara Boretti, Hyun-Min Chang, Jiawei Zhuang, Lukas Cavigelli

MACaQ is a post-training method for MXFP4 W4A4 that quantizes weights and activations to four bits. It balances input channels, mixes groups of 32 with paired weight transforms, and refines rounded codes within each layer. On Llama-3-8B, it reaches WikiText-2 perplexity 7.01 in 24.3 seconds versus MR-GPTQ’s 7.11 in 267.9 seconds, with different calibration budgets.

Summary by alphaXiv.

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