Reproduce “Polar Switch: Fast Muon Orthogonalization with Polynomial Circuits” 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

Zhehang Du, Weijie Su

Polar Switch speeds Muon's polar-factor orthogonalization with short polynomial circuits of symmetric matrix products and reusable buffers, replacing repeated quintic updates. On one NVIDIA B200, PS2-55 and PS3-55 reduced latency 15.5–18.7% versus Polar Express on square matrices and 8.6–16.2% versus Gram Newton–Schulz on rectangular matrices. Across three setups, final validation losses were similar; these are orthogonalization-call results, not full-training speedups.

Summary by alphaXiv.

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