Reproduce “MF-QAT: Multi-Format Quantization-Aware Training for Elastic Inference” 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
Zifei Xu, Sayeh Sharify, Hesham Mostafa
MF-QAT trains one model across multiple numerical formats so it can serve deployments that require different precisions. Slice-and-Scale converts an 8-bit anchor checkpoint into lower-precision MXINT or MXFP weights. In evaluations, MF-QAT was often competitive with format-specific QAT, including at some untrained intermediate widths, while conversion closely matched direct quantization; MXINT2 was an exception, and runtime speed was not measured.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.