Why it matters

Adapters are original PEFT. Understanding shapes history.

Advertisement

The architecture

Adapter: down-projection + activation + up-projection.

Residual connection.

Inserted per block.

Adapter structureDown-projectd -> rActivationGELU / SiLUUp-projectr -> d + residualHoulsby 2019 original; PEFT + Adapter Hub for reuse
Adapter theory.
Advertisement

How it works end to end

Houlsby et al. 2019.

Bottleneck rank r << d.

Adapter Hub for sharing.

Serial + parallel variants.