#inference-latency

Question 1: Single choice

After training, a LoRA model can be served with no extra inference latency compared with the original model, while bottleneck adapters (Houlsby et al.) add latency even when the bottleneck is very small. What explains the difference?

Adapters have far more parameters than LoRA, and inference latency grows with parameter count1
LoRA's matrices are so small that computing $BA\mathbf{x}$ as a separate branch costs no time2
LoRA's update is linear and parallel to $W_0$ , so it merges into one matrix $W_0 + BA$ 3
LoRA weights are only needed during training and are thrown away once the model is deployed4

Timed test