Adapter merging
Beginner
Folding a small trained add-on back into the model, so it runs as one piece with no extra steps.
Novice
Adding a trained LoRA update into the frozen weights once training is done, giving one ordinary model with no extra cost when it runs. The alternative is keeping adapters separate and loading the right one per request.
Expert
W ← W + (α/r)·BA, e.g. PEFT’s merge_and_unload(). Removes adapter latency but gives up sharing one base across many adapters; merging into a quantized base needs dequantizing first.