QLoRA
Also: quantised lora
LoRA fine-tuning performed on a quantised base model, so a large model can be adapted on a single consumer GPU. It combines quantisation with low-rank adapters.
QLoRA makes fine-tuning large models affordable. It holds the base model in a quantised 4-bit form to keep its memory footprint small, then trains small LoRA adapters on top. The base weights stay frozen; only the lightweight adapters are learned.
The result is that a model which would need server-class hardware to fine-tune in full precision can be adapted on a single consumer GPU, with little loss in quality for most tasks. It is the default approach for local fine-tuning, and what tools like Unsloth and Axolotl make straightforward.