Harvard CS249r Machine Learning Systems (Open Textbook, 4 Volumes)

New
License: CC BY-NC-ND 4.0 (book text); CC BY-NC-SA 4.0 (labs/slides) · Updated 2026-11-12
My rating

What it is

An open, MIT-Press-aligned textbook series from Harvard's CS249r course that explains how machine learning systems are built and why they are built that way. It ships with companion Python labs, hardware kits, and teaching tools.

Volume structure

  • Volume I: Foundations - single-machine execution, memory hierarchy, kernel efficiency, KV cache, data pipelines, production monitoring.
  • Volume II: Scaling - distributed training across accelerator clusters, fault tolerance, serving.
  • Volume III: Agentic AI - agent loops, tool sandboxes, context memory.
  • Volume IV: Physical AI - real-time sensor-actuator loops, vision and robotics model inference.

Why it fits here

Volume IV grounds physical AI in concrete embedded and robotics workflows, while the earlier volumes build the efficient-ML and edge-deployment foundations (TinyML, quantization, memory/latency/power constraints) that physical-AI deployments rely on.

Who it is for

Students and engineers who want first-principles understanding of ML systems and edge/physical-AI deployment, not just model training. Each volume includes runnable code (TinyTorch, Marimo Labs) in Python.

Tags: