“Deploying models using LoRa adapters incurs a 20% to 40% latency penalty, which requires merging the adapter into the base model for some latency-sensitive use cases.”