LoRA and PEFT · Deep Learning
LoRA = Low-Rank Adaptation — fine-tune 1% of parameters, get 99% of the performance
QLoRA: 4-bit quantization plus LoRA — fine-tune a 65B model on a single consumer GPU
F
Freeze the base model — no changes to original weights
LoRA keeps the entire original pretrained weight matrix W completely frozen, unchanged throughout fine-tuning.
Example: a massive pretrained language model's billions of original parameters remain completely untouched during LoRA fine-tuning.
L
Low-rank update — training a tiny fraction of parameters
Instead of updating the full weight matrix, LoRA adds a low-rank update ΔW = B × A, where B is d×r and A is r×k, with r chosen to be much smaller than d — meaning only A and B need to be trained, representing 100 to 10,000 times fewer parameters than the full matrix.
Example: instead of training billions of parameters in the full weight matrix, LoRA might only need to train a few million parameters in the much smaller low-rank matrices.
M
Merge back — zero added latency at inference
At inference time, the trained low-rank update can be merged directly back into the original weight matrix W, meaning there's no additional computational latency penalty compared to using the original model.
Example: after fine-tuning, the small B×A update is mathematically merged back into W, so the fine-tuned model runs exactly as fast as the original.
Q
QLoRA — combining 4-bit quantization with LoRA
QLoRA combines 4-bit quantization of the frozen base model with LoRA's low-rank fine-tuning, dramatically reducing memory requirements — enough to fine-tune a 65-billion-parameter model on a single consumer GPU, democratizing LLM fine-tuning.
Example: QLoRA making it possible for an individual researcher with a single consumer-grade GPU to fine-tune a massive 65B parameter model that would otherwise require enterprise-grade hardware.
1
A researcher wants to fine-tune a massive 65-billion-parameter language model, but only has access to a single consumer-grade GPU.
2
Fully fine-tuning all 65 billion parameters directly would require far more memory and compute than a single consumer GPU can provide.
3
Using QLoRA — combining 4-bit quantization of the frozen base model with LoRA's low-rank update training — dramatically reduces the memory footprint required.
4
This combination makes it feasible to fine-tune this massive model on a single consumer GPU, something that would have been completely impractical with traditional full fine-tuning.

Exams test whether you understand LoRA's core mechanism (freeze the base model, train only a small low-rank update, merge back for zero added latency) and whether you know QLoRA specifically adds 4-bit quantization on top of LoRA to further reduce memory requirements, enabling fine-tuning of very large models on modest hardware.

The most common trap is assuming LoRA fine-tuning adds extra inference latency, since it introduces new trainable parameters. In fact, the trained low-rank update gets merged directly back into the original weight matrix at inference time, resulting in zero added latency compared to the original model.

1. Does LoRA update the original pretrained weight matrix directly?
No — the original weight matrix stays completely frozen; LoRA trains a separate, much smaller low-rank update instead.
Tap to reveal / hide
2. Roughly how many fewer parameters does LoRA typically train compared to full fine-tuning?
100 to 10,000 times fewer parameters.
Tap to reveal / hide
3. Does LoRA fine-tuning add extra latency at inference time?
No — the trained low-rank update is merged back into the original weight matrix, resulting in zero added latency.
Tap to reveal / hide
4. What does QLoRA add on top of standard LoRA?
4-bit quantization of the frozen base model, further reducing memory requirements.
Tap to reveal / hide
5. What did QLoRA specifically enable, in terms of hardware accessibility?
Fine-tuning a 65-billion-parameter model on a single consumer GPU, democratizing LLM fine-tuning.
Tap to reveal / hide