model-width

A tweet suggests that scaling the embedding learning rate by model width can replace the need for µP (micro-parameterization), referencing Muon optimizer for hidden layers and Adam for the rest.