MLOps and LLMOps

Throughout the book, we’ve already used machine learning operations (MLOps) components and principles such as a model registry to share and version our fined-tuned large language models (LLMs), a logical feature store for our fine-tuning and RAG data, and an orchestrator to glue all our ML pipelines together. But MLOps is not just about these components; it takes an ML application to the next level by automating data collection, training, testing, and deployment. Thus, the end goal of MLOps is to automate as much as possible and let users focus on the most critical decisions, such as when a change in distribution is detected and a decision must be taken on whether it is essential to retrain the model or not. But what about LLM operations (LLMOps)? How does it differ from MLOps?

The term LLMOps is a product of the widespread adoption of LLMs. It is built on top of MLOps, which is built on top of development operations (DevOps). Thus, to fully understand...