AllocateWhat is QLoRA?
QLoRA is LoRA fine-tuning run on top of a 4-bit quantized base model. Because the frozen weights take a quarter of their FP16 memory, models that normally need a multi-GPU node can be fine-tuned on a single accelerator, making it the cheapest practical way to train large open models.
The adapter itself still trains at full precision; only the frozen base is compressed. Quality lands close to standard LoRA on most instruction-tuning tasks, which is why QLoRA is the default answer when the base model is large and the budget is not.
The tradeoff is throughput: dequantizing weights on the fly costs compute, so runs take longer per token than LoRA on uncompressed weights. When the model already fits in memory, plain LoRA is usually faster for the same money.
Try it: Fine-tuning cost calculatorRelated terms
Allocate is the cloud inference platform for companies that want to train and run their own models.