harvard-edge/cs249r_book

I consider myself a student of my own learning.

Whenever I want to understand an emerging or complex systems discipline, I learn it by trying to teach and build it from first principles. Writing forces me to slow down: to identify invariants, formalize abstractions, derive quantitative limits, and separate durable engineering foundations from passing industry hype.

This repository—the books, software frameworks, and hardware kits—is my open research and teaching instrument. It is how I clarify my own thinking for research, benchmarking programs (like MLPerf), and classroom instruction. If sharing this open process helps other engineers and students along the way, you are more than welcome to follow along.

This is for anyone who wants to engineer intelligent systems, not only train models: students, working engineers moving into ML infrastructure, and educators building a course. We assume you can program in Python and have met basic machine learning ideas, but the book builds the systems concepts from the ground up. You do not need a background in computer architecture, distributed systems, or datacenter operations. Volume I starts at the foundations, and the rest of the curriculum (TinyTorch, labs, hardware kits, and the simulator) lets you learn by building rather than only by reading.

The two volumes differ in scope, not depth. Both are equally rigorous. Volume I is the single-machine world: how an ML system works on one node with a handful of accelerators, from data and a single neuron's computation up through training, optimization, and deployment. Volume II is the at-scale world: many machines across a network, distributed training, fault tolerance, fleet orchestration, inference at scale, and governance. Volume II does not assume you have read Volume I, so you can start there if you already have the foundations. The natural path, though, is Volume I to build the mental models and Volume II to apply them across the fleet. The analogy we follow is Hennessy and Patterson: Computer Organization and Design first, then Computer Architecture: A Quantitative Approach.

You can just read the book. Each volume stands on its own. The rest of the curriculum (TinyTorch, the labs, the hardware kits, the simulator, and the interview practice) exists to deepen what the book teaches by making you build and measure it, but none of it is required to follow the text. Start with the book, and reach for the hands-on pieces when you want a concept to become muscle memory.

Deep learning books (Goodfellow et al.'s Deep Learning, Bishop, d2l.ai, fast.ai) teach you to design and train models: architectures, optimization, and the mathematics of learning. They mostly stop at the model. This book starts where they leave off. It treats the model as one component inside a system that has to ingest data, run on real silicon under power and latency budgets, serve predictions reliably, and keep working as the world drifts. You can finish a deep learning course knowing how a transformer learns and still not know why it stalls on a 4,000-accelerator training run, what the KV cache does to your serving memory, or why your accelerator sits idle. That gap is what we teach. Learn the model from a deep learning text, then learn the system here.

This is the most common mix-up, because "ML systems" and "MLOps" sound interchangeable and several good practitioner books share the words. MLOps books are operations guides: how to wire up a feature store, a pipeline, and a deployment with today's tools. They are valuable, and they age with the tooling. This book teaches the layer underneath: the physics and quantitative reasoning that explain why those tools exist and what they cost. We ask which questions matter, why a design is the way it is, and what it cannot escape (bandwidth, latency, power, failure rates).

Think of the difference between following a recipe and understanding how cooking works. A recipe gives you exact steps for one dish: this temperature, this pan, this many minutes. It works beautifully until the oven, the ingredients, or the kitchen changes. Understanding why heat, salt, acid, and time transform food is different. It lets you cook in any kitchen, rescue a dish that is going wrong, and invent one that no recipe covers.

An MLOps book hands you the recipe for the stack you have today. This book teaches the underlying science, so you can reason about any stack, debug the one that is failing, and design the one that does not exist yet.

References like Barroso, Hölzle, and Clidaras's The Datacenter as a Warehouse-Scale Computer are excellent. They distill how one organization engineered one canonical system, written by the people who built it. This project is a different kind of artifact, and the two are complementary rather than competing.

1. Curriculum, not reference. A synthesis lecture documents a finished design for practitioners who already know the field. This book teaches the discipline from the ground up, with lear