Reproduce “Deciding Scale Placement in MXFP4 Dot Products with Integer Arithmetic” 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
Shreyas Vinayakumar
The authors derive a compiler guard for moving shared power-of-two scales after MXFP4 dot-product reduction. They model finite values, infinities, and NaNs with integer coefficients and rounding conditions, then give scale-code bounds for a fixed binary reduction tree. In a 32-lane example, a left-fold result overflows while hoisting stays finite; the guarantee assumes separate binary32 operations, round-to-nearest ties-to-even, and gradual underflow.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.