qlora-4bit-nf4-lora-paged-optimizer-single-gpu
IN premise — summaries/2026/08/24/wiki-Large_language_model-chunk-5.md
Created 2026-08-24T17:11:17+00:00
QLoRA combines 4-bit NF4 quantization, LoRA adapters, and paged optimizers to enable fine-tuning of 65B+ parameter models on a single GPU (Dettmers et al., 2023)
Summary
By compressing model weights to 4-bit, training only small adapter layers, and paging optimizer state, you can fine-tune a model with 65 billion or more parameters on a single GPU instead of an expensive multi-card rig. In practical terms, this means a single high-end accelerator (or even a 48 GB card) is enough for serious fine-tuning, which removes the need for a cluster and makes that capability accessible to individual researchers and small teams.