LoRA
Fine-tuning by training a small pair of matrices alongside the frozen model instead of changing the model itself.
how it works · the vocabulary
adapterQLoRAPEFT
A LoRA adapter is a few megabytes where the model is hundreds of gigabytes, so one GPU can hold one base model and serve a thousand customers' adapters. That is why fine-tuning is cheap to buy where it once was not.