Welcome to this course on efficient inference with SGLang, where you learn about both text and image generation. This course is built in partnership with LMSys and RadixArk. LLM inference is one of the most expensive parts of running AI applications. It generates text one token at a time, and each new token depends on everything that came before it. So, if we were to use a naive implementation, then for every token the model generates, it has to reload all that prior context from memory. SGLang is an open-source inference engine to solve that memory bottleneck by caching and reusing computation that the model has already done. Our instructor for this course, Richard Chen, is a member of technical staff at RadixArk, where he's focused on making AI infrastructure open and accessible to everyone. Thanks, Andrew. The memory bottleneck you mentioned is exactly what we are addressing in this course. The first major optimization you will apply is the KV cache. Instead of recomputing all previous context for every new token, you store and reuse those intermediate values. This alone can give you a significant speed up. SGLang takes that further with RadixAttention, a method that extends caching across requests using a data structure called a radix tree. that identifies shared prefixes, so repeated context is only processed once. When multiple users share the same system prompts, or when a RAG application sends the same document as a context with different questions, RadixAttention recognizes the shared prefix and skips the redundant computation. This speed up grows quickly, especially when there's a lot of shared context across different requests. Exactly. And SGLang doesn't just work on text. It also accelerates image generation using similar caching strategies and multi-GPU parallelism. In this course, you will focus on three things. First, you will learn how LLM inference works under the hood to process text token by token. and how the KV cache stores intermediate computations for reuse. Next, you will see how SGLang's RadixAttention extends that caching across users and requests, and how to implement and measure those speedups yourself. And then, you'll apply similar techniques to diffusion models for faster image generation. Each lesson also includes hands-on practice so you can implement and measure these speed-ups yourself. Many people have worked to create this course. From RadixArk, I'd like to thank Liangsheng Yin, Mick Qian, Banghua Zhu, and Ying Sheng. And from DeepLearning.AI, Summer Rae also contributed to this course. In lesson one, we will walk through how a model processes a single request. from input token to the final generated response. By the end, you will understand exactly why inference gets expensive and you will be ready to dive into optimization that makes large scale LLM serving viable. Yep, that sounds wonderful. Let's head to the next video and get started.