Inference roles learning roadmap

The fifteen handbooks in dependency order: first a solid grounding in the language and in how models work, then GPU programming, then inference systems, a hand-written mini-sglang as the portfolio project, and finally interview preparation.

Four milestones: how to tell where you are

The 17 weeks fall into four phases, each ending with a milestone you can check yourself against. The test is always "can you do it", not "did you read it": if you are not there yet, go through that phase's key chapters and exercises again before moving on. Once you tick chapters off, this shows where you are on the route.

You are here

The full checklist for each milestone (including the performance bar on a real GPU and the mock interview scores) is in the 17-week sprint plan.

The big picture: how the 17 weeks are laid out

Week by week it matches the 17-week sprint plan: about 30 hours a week, read the core chapters closely, and skip what you already know after the self-test at the top of each chapter; portfolio project A (taking the hand-written mini-sglang to a GPU) runs alongside from week 6. Every phase's goal, checks and deliverables come from the sprint plan. If you mainly study on a Mac, start with the setup page to install everything in one go. The suggested times on each handbook's home page assume careful reading from scratch; follow that pace if you have the time. Click a bar to jump to its phase.

Dashed: work that runs alongside the main line or that you look up when needed: filling Python gaps, going back to the LLM and math books while studying CUDA and inference systems, the C++ that runs through the middle and later weeks, and project A.

Chapter by chapter

327 chapters arranged by phase; each phase says what you can do afterwards and what you should produce. Pick the kind of role you are aiming for and its key chapters are highlighted.

Core Recommended Optional, look it up when it comes up Click a circle to mark a chapter done (progress stays in this browser)