Why it matters

DoRA improves LoRA quality. Understanding shapes advanced PEFT.

Advertisement

The architecture

W = ||W|| * (W / ||W||).

LoRA on direction W / ||W||.

Separately train magnitude vector.

DoRA theoryDecompose Wmag + directionLoRA on directionrank rTrain magnitudecolumn-wise vectorLiu et al. 2024; small overhead vs LoRA; consistent quality gain
DoRA theory.
Advertisement

How it works end to end

Liu et al. 2024.

Small overhead per column magnitude.

PEFT library supports.