Most ML curricula stop at the model: architectures, loss functions, a training loop that fits in a notebook. The harder problem — making that model run efficiently and reliably on real hardware, from a single GPU to a datacenter — is exactly where this textbook lives, treating ML systems as an engineering discipline with its own underlying science.
What Sets It Apart
- It teaches principles, not a tool stack. Where MLOps guides chain together vendors and deep-learning texts dwell on architectures, this explains why systems behave as they do, so the reasoning transfers to any infrastructure.
- Two volumes mirror how scale actually breaks things: Volume I covers single-machine fundamentals (optimization, INT8 quantization, KV-cache memory limits, model-to-silicon mapping); Volume II covers distributed production systems (GPU scheduling, fault tolerance, governance).
- It ships a full learning ecosystem, not just prose: interactive Marimo labs, TinyTorch (a build-your-own ML framework), hardware kits for Arduino/Raspberry Pi, and lecture slides.
- It is genuinely free and open (CC-BY-NC-SA), with 100+ contributors and 25,000+ GitHub stars; a print edition with MIT Press is planned for 2026.
Who It's For
A strong fit if you can already train a model but freeze when asked how to serve it under latency and memory budgets, or if you're an educator assembling an ML systems course from scratch. Look elsewhere if you want a deep-learning theory text or a copy-paste MLOps recipe book — this trades quick wins for durable understanding, and rewards readers willing to think about hardware.
