Fine-Tuning with LoRA & QLoRA
Full fine-tuning a 7B model requires 56GB of VRAM. You don't have that. Neither do most companies. LoRA lets you fine-tune the same model in 6GB by training less than 1% of the parameters. This isn't a compromise -- it matches full fine-tuning quality on most tasks. The entire open-source fine-tuning ecosystem runs on this one trick. Related: Phase 10 covers the SFT/DPO loops from scratch. This lesson plugs those into the 2026 PEFT toolkits (PEFT, TRL, Unsloth, Axolotl, LLaMA-Factory). Implement LoRA by injecting low-rank adapter matrices (A and B) into a pretrained model's attention layers. Calculate the parameter savings of LoRA vs full fine-tuning: rank r with dmodel dimensions trains 2rd parameters instead of d^2. Fine-tune a model using QLoRA (4-bit quantized base + LoRA adapters) to fit within consumer GPU memory. Merge LoRA weights back into the base model for deployment and compare inference speed with and without adapters. You have a base model. Llama 3 8B. You want it to answer customer support tickets in your company's voice. SFT is the answer. But SFT has a cost problem. Full fine-tuning updates every parameter in the model. Llama 3 8B has 8 billion parameters. In fp16, each parameter takes 2 bytes. That's 16GB just to load the weights. During training, you also need gradients (16GB), optimizer states for…
Fine-Tuning with LoRA & QLoRA: Full fine-tuning a 7B model requires 56GB of VRAM. You don't have that. Neither do most companies. LoRA lets you fine-tune the…
This free lesson is part of the AI Engineering from Scratch curriculum. Read the full explanation, run the lesson code, and verify the result in the interactive reader or from the repository source.
Browse the complete course catalog or open this lesson on GitHub.