Why it matters
muP saves massive sweep cost. Understanding shapes modern LLM training.
Advertisement
The architecture
Rescale init + LR + attention scale by width.
Ensures updates same magnitude.
Tune HPs at small width; use at large.
Advertisement
How it works end to end
Yang + Hu 2021.
Init: fan-in scaling.
LR: per-layer rescaled.
Attention: 1/d instead of 1/sqrt(d) in specific places.