Read the book online: https://arafatkatze.github.io/inference-math/ (PDF and ePub downloads are in the site's sidebar)
A Quarto book covering the math of LLM inference and the actual dollars: loading the model, serving it (prefill/decode/KV cache), batching, load testing, sizing the GPU fleet, the monthly bill, and working with inference providers.
The book renders to three formats from a single source:
- HTML — the website (published on GitHub Pages)
- PDF — the printable book
- ePub — the e-reader version
Install Quarto, then:
# live-reloading preview in the browser
quarto preview
# render all formats (HTML + PDF + ePub) into _book/
quarto renderPDF output additionally requires a LaTeX distribution and the DejaVu fonts:
quarto install tinytex
sudo apt-get install fonts-dejavu # Linux; macOS ships DejaVu-compatible fontsTo render only one format: quarto render --to html (or pdf, epub).
The book is served at https://arafatkatze.github.io/inference-math/. Every push to main triggers .github/workflows/publish.yml, which renders the book (HTML + PDF + ePub) and deploys it to GitHub Pages.
One-time setup (requires repo admin): open Settings → Pages and set Source to GitHub Actions, then re-run the latest "Publish Quarto Book" workflow (or push any commit to main).
├── _quarto.yml # one config for all formats
├── index.qmd # preface
├── 01-architecture.qmd # model architecture
├── 02-picking-the-gpu.qmd # picking the GPU + node cost
├── 03-loading-the-model.qmd # GPUs needed to load the model
├── 04-serving.qmd # prefill / decode / KV cache
├── 05-batching.qmd # batching regions + load test math
├── 06-sizing-the-fleet.qmd # sizing the GPU fleet
├── 07-the-bill.qmd # the monthly bill
├── 08-load-testing.qmd # how to actually run a load test
├── 09-working-with-providers.qmd # working with inference providers
├── 10-closing.qmd # closing note + credits
├── references.qmd
└── images/