Customizing Open-Source LLMs

This comprehensive course is designed to guide learners through the process of fine-tuning open-source Large Language Models (LLMs) such as Meta LLaMA 3 using PyTorch, with a particular emphasis on cloud-native training and deployment. The course covers everything from the fundamentals to advanced concepts, ensuring students acquire both theoretical knowledge and practical skills.\nThe journey begins with an introduction to LLMs, focusing on their architecture, capabilities, and the specific features of Meta LLaMA 3. Next, the course dives into PyTorch fundamentals, teaching students how to perform basic operations with tensors and build simple neural networks. This foundation is crucial for understanding the mechanics behind LLMs. Data preparation is a crucial aspect of training models. The course covers comprehensive data collection and preprocessing techniques, such as tokenization and text normalisation. These steps are essential for preparing datasets suitable for fine-tuning LLMs like Meta LLaMA 3. Through practical exercises, students learn how to handle and preprocess various types of text data, ensuring they can prepare their datasets for optimal model performance.\nFine-tuning Meta LLaMA 3.2 with PyTorch forms a significant part of the course. Students will delve into the architecture of Meta LLaMA 3, learn how to load pre-trained models, and apply fine-tuning techniques. The course covers advanced topics such as regularisation and optimization strategies to enhance model performance. Practical sessions guide students through the entire fine-tuning process on custom datasets, emphasising best practices and troubleshooting techniques.\nA critical aspect of this course is its focus on cloud-native training and deployment using Nvidia NIM. Furthermore, students learn how to deploy models using Docker and Kubernetes, set up monitoring and maintenance tools, and ensure their models are scalable and efficient. To round off the learning experience, the course includes an in-depth segment on exporting models for inference and building robust inference pipelines. Students will deploy models on cloud platforms, focusing on practical aspects of setting up monitoring tools to maintain model performance and reliability.\nThe course culminates in a capstone project, where students apply all the skills they have learned to fine-tune and deploy Meta LLaMA 3 on a chosen platform. This project allows students to demonstrate their understanding and proficiency in the entire process, from data preparation to cloud-native deployment.