vLLM, Disassembled: Every Screw, Washer, and CUDA Graph Explained

This article is a comprehensive architectural breakdown of vLLM's V1 engine as of August 2025, covering its core components: the LLM engine, scheduler, KV-cache manager, model executor, and worker initialization. It walks through paged attention, continuous batching, CUDA graph capture, and the request lifecycle, then previews advanced features like chunked prefill, prefix caching, speculative decoding, and disaggregated prefill/decode. Later sections promise coverage of multi-GPU scaling, the serving layer, and benchmarking — making it a thorough reference for anyone building on or contributing to high-throughput LLM inference systems.