Why it matters

Activation scaling shapes stability. Understanding shapes deep training.

Advertisement

The architecture

LayerNorm / RMSNorm control layer output.

Residual: identity + delta.

Init controls delta scale.

Activation scaling flowInput variancecontrolled by normDelta scaleinit + res gainOutput variancesimilar to inputPost-LN vs Pre-LN affect variance flow; modern uses Pre-LN or RMSNorm
Activation scaling.
Advertisement

How it works end to end

Pre-LN vs Post-LN.

Residual gain init.

Careful across depth.

Signal propagation theory.