LLM Customization and Inference

This unit introduces methods to adapt large language models through full and parameter-efficient fine-tuning and examines the cost and performance trade-offs of those approaches. It also covers practical inference techniques—prompt engineering, vector-store caching, and chains for long-document handling—to prepare learners for building efficient, customizable LLM-based solutions in later units.

Learning Objectives

• Analyze trade-offs among full fine-tuning and parameter-efficient fine-tuning (PEFT) methods with respect to model performance, compute requirements, and cost
• Apply parameter-efficient techniques (e.g., adapters, LoRA, prompt tuning) to adapt a pre-trained LLM for a specified downstream task
• Evaluate inference strategies—prompt engineering, caching with vector stores, and chain-based approaches—for latency, accuracy, and cost on long-document or retrieval-augmented tasks
• Demonstrate an end-to-end inference workflow that integrates prompt design, vector retrieval, and chaining to improve relevance and efficiency