Skip to main content
All terms
Architectures

RMSNorm

A lightweight step that keeps the numbers flowing through a network at a steady scale.

Definition

RMSNorm, short for Root Mean Square Normalization, rescales the numbers passing through a layer so they stay a consistent size, which keeps training stable. It is a simpler, faster cousin of layer normalization and is used in many modern language models.