Learn — course workspace
Continue the interactive 13-week inference-systems sprint, with foundations and retained learning references nearby.
The active owner route is the 13-week inference-systems sprint, 28 September–27 December 2026. It starts with a transformer forward trace on one Mac and builds toward measurement, runtime changes, serving, and the single-device-to-cluster boundary. Each week asks for a prediction, a bounded check, and an independent explanation. Reading or a passing script alone does not establish mastery.
Start with one session
- Open the interactive inference workspace. Empty browser state starts at Week 1, Day 1; existing progress resumes.
- Learn the Day 1 concepts, record a prediction, do the small CPU exercise, check the result, and explain it in your own words. The full reference-model smoke arrives on Day 6 after attention has been taught.
- Save a browser-local checkpoint and use Continue for the next day. See the manual progress record for historical context; opening a page never records mastery.
Course route
| Stage | Read and do |
|---|---|
| Now — Week 1 | Seven ready sessions: token IDs to logits, causal attention, the reference CPU trace, and a closed-book debug. |
| Next — Weeks 2–3 | Generation, prefill/decode, KV cache, and measured latency and memory. Start only after the preceding immediate gate. |
| Later — Weeks 4–13 | GPU execution, attention performance, vLLM source trace, scheduling, a bounded change, serving, and the Mac-to-cluster boundary. See every week’s contract. |
Repair a prerequisite
The ten-module ground-up curriculum and its ordered lessons teach the smallest missing concept. The session file numbers reflect creation order, not reading order. Use the coverage map to locate a specific subsystem. A prerequisite repair can be selected in the same browser workspace while preserving the sprint draft.
Explore after the current task
- Practical paths and buildable artifacts connect concepts to repository labs and mastery gates.
- Learning library groups modern LLM mechanics, interview maps, Mac-local findings, and session-specific decisions.
- Document library map separates reference, factory contracts, evidence, and historical material across the full corpus.
- Legacy Python/WASM/WebGPU walkthrough remains available as a historical guided tour; it is not the current learning queue.