Training
QLoRA
Also called Quantised LoRA.
What is qlora?
QLoRA is LoRA applied on top of a base model quantised to four bits. The frozen base is stored in low precision while the trained adapter stays in higher precision, cutting memory requirements roughly in half again and putting large models within reach of a single GPU.
It is what makes a 70B fine-tune possible on one 80GB card rather than a multi-node cluster. The quantisation applies only to the frozen weights, so the gradients flowing into the adapter are still computed at useful precision and quality loss is small on most tasks.
The cost is speed. Dequantising weights on the fly during the forward pass adds overhead, so a QLoRA run is meaningfully slower per step than the same run in half precision. When both fit, half precision is usually cheaper overall despite the larger card.
The decision rule is simple: use half-precision LoRA if the model fits on the class you want, and QLoRA when it does not. Quantising to save money on a model that would have fitted anyway usually costs more in wall time than it saves in hourly rate.
Related terms
Start with the free tier
A magic link creates your account, your tenant and your first API key. No card until you ask for compute.