Why it matters

LoRA is default fine-tuning. Understanding shapes strategy.

Advertisement

The architecture

Add A, B matrices (rank r) per target layer.

Train only A, B.

Merge into base at inference.

LoRA FT flowFrozen baseno updatesA, B adaptersrank rTrain + mergeat inferencePEFT library standard; QLoRA combines with 4-bit base for further memory savings
LoRA FT.
Advertisement

How it works end to end

PEFT library.

Rank 8-64 typical.

Target: attention + MLP.

Merge saves inference overhead.