Why Re-Parameterization PEFT Exists

Full fine-tuning updates almost every weight in a giant model. The GPU must hold the weights, the gradients, and the optimizer state at once. For very large models — or for many client-specific copies — that cost grows too fast.

Intuition

A simple picture:

PEFT (parameter-efficient fine-tuning) keeps the base model frozen and changes only a small part. Re-parameterization PEFT goes one step further: it rewrites the update itself into a cheaper form (LoRA and QLoRA are the main examples in this lesson).

How it works

What each method changes

Method What changes Typical footprint
Full fine-tuning Almost all weights Huge weights + gradients + optimizer state
LoRA Small low-rank adapters only A few million trainable parameters instead of billions
QLoRA 4-bit frozen base + LoRA adapters Much smaller training memory; often one GPU

Why this lesson matters

Lesson 3.3 covered adapters and soft prompts. This lesson focuses on the re-parameterization branch of PEFT:

  1. Can we fine-tune large models with limited resources?
  2. Why does a low-rank update often work (intrinsic dimension → LoRA)?
  3. How does storing weights in fewer bits help (QLoRA)?
  4. Can quantization + checkpointing + paged optimizers fit a big model on one GPU?

What goes wrong

One-line summary

Re-parameterization PEFT exists because full fine-tuning is too heavy for many real systems — so we learn a smaller update instead of rewriting the whole model.

Key terms