The Ultra-Scale Playbook

The Ultra-Scale Playbook

Teaches how to train large language models on big GPU clusters, as a web guide you read in the browser with a PDF copy in its files. For engineers and researchers who run or plan large training jobs.

Made by nanotron on Hugging Face. These notes are Mindket's, written from its Space page and README; the app's own files and license are on Hugging Face.

About

What it is: The Ultra-Scale Playbook is a guide to training large language models on large GPU clusters. It is a web page you read in the browser, built from source files in the repository, and a PDF copy of the guide sits alongside them.

Who it suits: It suits people who train or plan to train big models on many GPUs, and people who want to learn the field from one long guide.

Industry | Who uses it | How they use it | Example
AI research | Researchers | Read the guide before planning a large training run | Open the page and read it from the top
Machine learning engineering | ML engineers | Keep it as a reference while setting up GPU clusters | Open the PDF copy from the repository
Education | Students | Learn how large model training is organised | Work through the page a section at a time
Web publishing | Front-end developers | Study how the page is assembled from fragments | Read the files in src and the webpack config

Skip it if you want a software tool to run, since this is reading material and the code behind its page.

Where to use it:

• A web browser, on the Space's page at Hugging Face
• A local copy of the files on Windows, macOS or Linux, opened from dist/index.html
• A PDF copy that comes with the repository

Use case | Who | What they get back
Read the guide online | Engineers | The full page with nothing to install
Keep an offline copy | Students | The files on your computer, including the PDF
Change the page | Web developers | A build you can edit in src and rebuild

How to set it up: Reading online needs a browser only. A local copy needs Python to fetch the files. Editing the page also needs Node with npm and Git LFS.

• A web browser
• Python and pip, to fetch the files
• Node with npm, and Git LFS, only if you want to change the page

1. Install the Hugging Face hub library.
2. Download the Space's files into a folder named ultrascale-playbook.
3. Move into that folder.
4. Load dist/index.html in a web browser.

pip install -U huggingface_hub
hf download nanotron/ultrascale-playbook --repo-type space --local-dir ultrascale-playbook
cd ultrascale-playbook

To edit the page, work in the project folder with these commands.

# fetch LFS-tracked files (large assets)
git lfs pull

npm install
npm run build
npm run dev

How to use it: Read the guide in the browser. If you edit it, change files under src, run the build, and commit the dist folder.

Tool | What it does
git lfs pull | Fetches the large files the repository tracks with LFS
npm install | Installs what the page's build needs
npm run dev | Runs the page for development and lists the available fragments in the console
npm run build | Builds the page from src into dist, which you then commit

Example: You want a banner on the page. You add a file named banner.html under src/fragments, put its placeholder in src/index.html, run the dev command, and check that the console lists the new fragment.

For newcomers and experts: Open the page and read it from the top. You need no setup to do that, and the PDF copy in the repository holds the same guide for offline reading.

The page is made from HTML fragments, which can load in two ways. Compile-time fragments are baked into the built HTML, so every visitor sees them right after load. Run-time fragments are fetched and inserted later, which keeps the final HTML small, ideally under 1MB. The trade-off is that a run-time fragment appears a moment after the page. The repository also holds convert scripts and a markdown source for the guide.

Edit files in src
Run npm run dev and check the console
Run npm run build
Commit the dist folder
Open dist/index.html

Good to know:

• Size: The repository holds 383 files, and large assets are tracked with Git LFS.
• Gaps: The source describes the build and not the guide's chapters.
• License: apache-2.
• Cost to run: none named.
• Data it reaches: Local files only. The page runs in the browser, and the source names no account or key.

On Hugging Face, 13 November 2026

• Likes: 4,131
• Built with: static
• License: apache-2.0

Links

• Space: https://huggingface.co/spaces/nanotron/ultrascale-playbook
• Files: https://huggingface.co/spaces/nanotron/ultrascale-playbook/tree/main

Try it

Open the Space in your browser, or copy it to your own Hugging Face account from its page. Nothing in this kit runs by itself.