How to Make Your Model Fast, first edition

The first edition of How to Make Your Model Fast: A Systems View of Efficient Machine Learning, from Silicon to Agents.

Read it free at ai.usamah.me, or download it below. 👇

Most of what is written about machine learning assumes the model is the interesting part and the machine is a detail. This book is about the boundary where a model meets real hardware under a real budget for latency, memory, power and money: how to predict what that boundary will do to you, and what to change first when it does. Fourteen parts, about 90,000 words, 48 figures, and worked problems with full answers at the end of every chapter.

What is in it

• Foundations: rooflines and budgets; inside an edge accelerator
• The stack between a model and a machine: kernels; compilers, graphs and runtimes
• Making the model smaller: quantisation; pruning, sparsity and distillation
• Applications: vision in the real world; transformers on small machines; perception, VLMs and robots
• Operations: profiling; serving and MLOps; agents
• Conclusion and further reading

Downloads

• how-to-make-your-model-fast.pdf: A4, with every figure, footnote and worked answer
• how-to-make-your-model-fast.epub: reflowable, for e-readers such as Apple Books

Found an error?

There will be errors, and I would genuinely like to hear about them. Open an issue, or email usamahzaheer155 [at] gmail [dot] com.