SignMuon: Communication-Efficient Distributed Muon Optimization

Sign-Muon is a 1-bit optimizer for distributed training that combines signSGD's majority vote aggregation with Muon’s polar-step framework. It achieves communication efficiency by transmitting only entrywise signs and performs local orthogonalization, leading to faster convergence (O(1/√T)) and significant bandwidth reduction compared to standard float32 methods.

Open the full episode

More from AI Radio