QLoRA: Efficient Finetuning of Quantized LLMs

Introduced 4-bit NormalFloat (NF4), Double Quantization (DQ), and Paged Optimizers, enabling fine-tuning of 65B parameter models on a single 48GB GPU with zero performance loss.

Directly guides architectural decisions, alignment strategy, and serving infrastructure for task-text-generation.

Assumptions

• Empirical distribution regularity holds and target domain adheres to pretraining linguistic/visual support

Limitations

• Resource scaling, inference memory requirements, and alignment robustness vary with model size and hardware topology