What this book covers

Chapter 1, Understanding the LLM Twin Concept and Architecture, introduces the LLM Twin project, which is used throughout the book as an end-to-end example of a production-level LLM application, and defines the FTI architecture for building scalable ML systems and applies it to the LLM Twin use case.

Chapter 2, Tooling and Installation, presents Python, MLOps, and cloud tools used to build real-world LLM applications, such as an orchestrator, experiment tracker, prompt monitoring and LLM evaluation tool. It shows how to use and install them locally for testing and development.

Chapter 3, Data Engineering, shows the implementation of a data collection pipeline that scrapes multiple sites, such as Medium, GitHub and Substack and stores the raw data in a data warehouse. It emphasizes collecting raw data from dynamic sources over static datasets for real-world ML applications.

Chapter 4, RAG Feature Pipeline, introduces RAG fundamental concepts, such as embeddings, the vanilla RAG framework, vector databases, and how to optimize RAG applications. It applies the RAG theory by architecting and implementing LLM Twin’s RAG feature pipeline using software best practices.

Chapter 5, Supervised Fine-Tuning, explores the process of refining pre-trained language models for specific tasks using instruction-answer pairs. It covers creating high-quality datasets, implementing fine-tuning techniques like full fine-tuning, LoRA, and QLoRA, and provides a practical demonstration of fine-tuning a Llama 3.1 8B model on a custom dataset.

Chapter 6, Fine-Tuning with Preference Alignment, introduces techniques for aligning language models with human preferences, focusing on Direct Preference Optimization (DPO). It covers creating custom preference datasets, implementing DPO, and provides a practical demonstration of aligning the TwinLlama-3.1-8B model using the Unsloth library.

Chapter 7, Evaluating LLMs, details various methods for assessing the performance of language models and LLM systems. It introduces general-purpose and domain-specific evaluations and discusses popular benchmarks. The chapter includes a practical evaluation of the TwinLlama-3.1-8B model using multiple criteria.

Chapter 8, Inference Optimization, covers key optimization strategies such as speculative decoding, model parallelism, and weight quantization. It discusses how to improve inference speed, reduce latency, and minimize memory usage, introducing popular inference engines and comparing their features.

Chapter 9, RAG Inference Pipeline, explores advanced RAG techniques by implementing methods such as self-query, reranking, and filtered vector search from scratch. It covers designing and implementing the LLM Twin’s RAG inference pipeline and a custom retrieval module similar to what you see in popular frameworks such as LangChain.

Chapter 10, Inference Pipeline Deployment, introduces ML deployment strategies, such as online, asynchronous and batch inference, which will help in architecting and deploying the LLM Twin fine-tuned model to AWS SageMaker and building a FastAPI microservice to expose the RAG inference pipeline as a RESTful API.

Chapter 11, MLOps and LLMOps, presents what LLMOps is, starting with its roots in DevOps and MLOps. This chapter explains how to deploy the LLM Twin project to the cloud, such as the ML pipelines to AWS and shows how to containerize the code using Docker and build a CI/CD/CT pipeline. It also adds a prompt monitoring layer on top of LLM Twin’s inference pipeline.

Appendix, MLOps Principles, covers the six MLOps principles used to build scalable, reproducible, and robust ML applications.