LLM Inference Engineering
A growing series of written tutorials that build up LLM inference optimization step by step, from autoregressive generation and the KV cache through prefill-decode disaggregation to production engines such as vLLM and SGLang.
LLM Inference Engineering is a blog-style curriculum on how a large language model actually serves requests efficiently. It starts from the basics of autoregressive generation, the chain rule of probability behind next-token prediction, and why these models need causal masking, then moves into the prefill and decode phases of inference, showing how the KV cache bridges them and where each phase's optimizations differ because one is compute-bound and the other memory-bound.
Later posts cover prefill-decode disaggregation, in which the two phases run on separate machines, along with the throughput and latency metrics used to reason about the trade-offs, such as time to first token and time per output token. The series is maintained by the founder of Outcome School and is stated to keep growing as more posts and videos are added.