Technology On the Net

Fine-tuning used to mean one model per use case: a support model, a legal-summary model, a model per big customer. Each one needed its own GPUs, its own deployment, and its own on-call worries. Low-rank adaptation (LoRA) changes the economics. A LoRA fine-tune is a small set of extra weights that sits on top of a shared base model, so one GPU deployment can serve dozens or hundreds of fine-tunes at once. That sounds simple, but multi-adapter serving has its own failure modes: adapters that load too slowly, batches that mix too many adapters, rank mismatches, and tenants who get the wrong fine-tune. This guide covers how multi-LoRA serving works, how to configure it, and the operational rules that keep it safe in production. What a LoRA adapter actually is A full fine-tune updates every weight in the model. LoRA freezes the base weights and learns a small correction for selected layers, usually the attention projections and sometimes the MLP layers. For a weight matrix W , LoRA le...