Full fine-tuning updates every parameter in a Large Language Model, demanding substantial computational resources and memory. This often makes adapting the largest models impractical without specialized hardware. This chapter presents Parameter-Efficient Fine-tuning (PEFT) methods as a resource-conscious alternative.

PEFT techniques modify only a small fraction of the model's parameters, drastically reducing computational overhead and storage requirements while often achieving performance comparable to full fine-tuning on specific tasks.

You will study the principles and implementations of several prominent PEFT approaches:

• Low-Rank Adaptation (LoRA): Injecting trainable low-rank matrices into transformer layers.
• Quantized LoRA (QLoRA): Combining LoRA with quantization for even greater memory savings.
• Adapter Modules: Inserting small, trainable bottleneck layers within the existing architecture.
• Prompt Tuning & Prefix Tuning: Learning continuous embeddings prepended to the input or hidden states, keeping the base model frozen.

We will analyze the mechanics behind each technique, compare their respective advantages and trade-offs, and walk through practical implementations using common libraries. By the end of this chapter, you will understand how to select and apply appropriate PEFT methods to adapt LLMs efficiently.