time-to-first-token

Why we included this project

Most LLM inference learning material is a pile of disconnected tutorials. This roadmap is the opposite: all fifty sessions feed one OpenAI-compatible inference service that you deploy on a rented GPU, instrument with Grafana, load test past a thousand concurrent requests, and optimize with quantization and speculative decoding. The roofline model comes before any tuning knob, so you understand why each optimization matters before you touch it. Engineers who already know Python and transformers but have never run a serving stack get a coherent path from first principles to a reproducible benchmark, and the skim-versus-build session labels make it easy to compress material you already know. That also makes it work as a refresher for people with some serving experience, not just a beginner course.