4.2 Parameter-Efficient Fine-Tuning (PEFT)
Interview-ready field notes
The 20-second answer
PEFT adapts a large frozen model by training a small number of extra parameters, reducing memory, storage, and deployment cost.
Core ideas to retain
- LoRA learns a low-rank update instead of updating a full weight matrix .
- The base model remains frozen; only adapters receive gradients, so optimizer state and saved checkpoints are much smaller.
- Adapters can be swapped per task. At inference, some LoRA adapters can be merged into base weights to avoid extra latency.
- QLoRA combines a quantized frozen base model with trainable LoRA adapters, making large-model fine-tuning more accessible.
Interview / OA rule
Contrast it clearly: full fine-tuning updates every weight; LoRA assumes the useful update is low-rank and trains a small residual instead.
One good written resource
Hugging Face — PEFT Documentation — read this after the video when you want a clearer mental model, not more pages of notes.
Most asked
Interview questions to practise aloud
Each answer is the level of detail expected for a strong fundamentals round.
What is LoRA?
Low-Rank Adaptation freezes a weight matrix and learns a small low-rank update. It greatly reduces trainable parameters and optimizer memory.
Why is PEFT cheaper than full fine-tuning?
Fewer parameters require gradients, optimizer states, checkpoint storage, and distributed communication. The large base weights stay frozen.
What is QLoRA?
Fine-tune LoRA adapters while the base model is loaded in low-bit quantized form. It saves memory, with some quantization and implementation trade-offs.