Why it matters

RMSNorm is default modern LLM norm. Understanding shapes design.

Advertisement

The architecture

y = x / sqrt(mean(x^2) + eps) * gain.

No bias term.

RMSNorm computationSquare + meanmean(x^2)RMS scale1/sqrt(mean + eps)Multiply gainlearnableSimpler than LayerNorm; skips mean subtraction + bias
RMSNorm.
Advertisement

How it works end to end

Introduced by Zhang and Sennrich 2019.

Simpler than LayerNorm.

Slightly faster; equal quality.