71 / 87

ConceptFine-tuningMultiple choiceEasy

You serve one LoRA fine-tune of a 13B model for a single product. Each attention projection keeps its frozen $d \times k$ weight $W_{0}$ and adds a rank-16 update, computing $W_{0} x + α B A x$ with $B$ of shape $d \times 16$ and $A$ of shape $16 \times k$ .

The adapter runs as a side branch next to each frozen weight, and decoding is measurably slower than the base model.

How would you remove the overhead without changing the model's outputs?