Microsoft Open-Sources Dion Optimizer: Integrates Gram-NS, Row Selection, and More, 6x Faster Training

JohnCLangford · x · 2026-08-14

Microsoft releases Dion optimizer, integrating computational and statistical improvements including Gram Newton Schulz, row selection, megabatching, cudagraph capture, symmetric kernels, and joint learning rate+fraction, row norm, fused rounding. Claims 6x faster and more parallelizable than Adam.

Original post →

More from coding & agent

coding & agent channel →