Author here. I wanted to understand what vLLM and llama.cpp are actually doing under the hood, but the codebases are massive. So I wrote a stripped down version from scratch to see the core ideas without the production complexity.
The blog walks through why your first token is always the slowest, why output tokens cost 5x more, and how stuff like speculative decoding and chunked prefill actually work, from the perspective of a systems engineer!
Author here. A bit more context: By day I'm a systems engineer building AI networking infrastructure. So I kept ending up in conversations where I'm not exactly able to wrap my head on the latest inference magic trick.
Like when someone mentioned vLLM's paged attention, I knew virtual memory paging, but had no idea someone had applied the same idea to KV cache allocation on GPUs.
Code: https://github.com/Anirudh171202/WhiteLotus