muP scaling

Zero-shot learning-rate transfer using the first-principles muP rule shipped with OKLS.

Transfer calculator

Proxy width

Target width

Base LR

Tune once at proxy scale. OKLS's muP rule keeps the optimal learning rate invariant as width grows - no re-sweeps at frontier scale.

lr_target = lr_proxy × (w_proxy / w_target)

lr_target = 0.02 × (256 / 8192)

Shape scaling covers depth, aspect ratio, and batch - verified against every run on the network leaderboard.