Customizing and Optimizing LLMs

This unit introduces practical techniques for adapting large language models to domain-specific needs and for optimizing inference behavior. Students explore fine-tuning and parameter-efficient approaches, cost/performance trade-offs, prompt engineering, vector-store caching, and chaining methods for long documents to prepare for deployment-focused topics in later units.

Learning Objectives

• Analyze trade-offs between full-model fine-tuning and parameter-efficient fine-tuning (PEFT) approaches by comparing impacts on model customizability, accuracy, compute requirements, and storage needs for a given use case.
• Apply parameter-efficient fine-tuning methods (e.g., adapters, LoRA, prompt tuning) to design and document a lightweight customization workflow and calculate estimated training time, GPU memory, and storage impacts.
• Demonstrate inference-optimization techniques by crafting effective prompts, implementing basic prompt engineering patterns, and constructing a simple chain that uses vector-store caching for retrieval-augmented generation on long documents.
• Evaluate and justify an end-to-end inference strategy that balances latency, cost, and performance—selecting appropriate PEFT methods, caching policies, and chain architectures for a specified application scenario.