Inside vLLM: Anatomy of a High-Throughput LLM Inference System - Aleksa Gordić
www.aleksagordic.com
From paged attention, continuous batching, prefix caching, specdec, etc. to multi-GPU, multi-node dynamic serving at scale.