QLoRA: Quantized LoRA
QLoRA makes LoRA cheaper by loading the base model in low precision while training small adapters.
The idea
text
quantized frozen base model + trainable LoRA adapters
The base model uses much less memory. The adapters learn the task.
Why it matters
QLoRA made it practical for more teams to fine-tune larger open models without full enterprise-scale hardware.
Key concepts
| Concept | Meaning |
|---|---|
| 4-bit quantization | store base weights compactly |
| NF4 | quantization format used for neural weights |
| LoRA adapter | small trainable update |
| paged optimizer | reduces memory spikes |
Risks
- data quality still matters
- quantization can hurt hard tasks
- adapter overfitting is possible
- evals must compare against baseline
Knowledge check
Q1: What is quantized in QLoRA?
The frozen base model weights.
Q2: What is trained?
The LoRA adapter parameters.