LLM Training Calculator

Estimate pre-training wall-clock for any transformer architecture, or reverse-engineer the MFU you're actually achieving from a measured step time.

Supports MLA / MHA / GQA / MQA / Sliding Window attention, dense or MoE FFN, with optional Multi-Token Prediction.

npm install
npm run dev

Open http://localhost:5173 — done.

npm run build
npm run preview   # preview the build locally

The static site goes to dist/.

1. Create the repo on GitHub (e.g. llm-training-calculator).
2. Edit vite.config.js — change base to match your repo name:

   base: '/<your-repo-name>/',

   If you're deploying to a user/organization root site (<user>.github.io), set base: '/' instead.
3. Enable Pages: in the repo, go to Settings → Pages → Source → "GitHub Actions".

git init
git add .
git commit -m "init"
git branch -M main
git remote add origin git@github.com:<you>/<repo>.git
git push -u origin main

The workflow at .github/workflows/deploy.yml will build and publish automatically on every push to main. Site goes live at https://<you>.github.io/<repo>/ within ~1 minute.

Add a CNAME file in the repo root with your domain name, then set base: '/' in vite.config.js.

.
├── index.html              # Vite entry (loads fonts, mounts root)
├── src/
│   ├── main.jsx            # React mount point
│   └── Calculator.jsx      # The full component (single-file)
├── vite.config.js          # Build config; set `base` to your repo name
├── package.json
└── .github/workflows/deploy.yml

MIT