gpu-utils
Tiny libraries for doing things in the browser with WebGPU.
Each package is a small model, trained from scratch, that runs on the user’s GPU in the browser: no server, no API key, no network round trip, and a bundle measured in tens of kilobytes. The approach follows gpu-lexer and gpu-time: a mechanical tokenizer, hashed sparse features, a ~30K-parameter tagger in hand-written WGSL, and a deterministic compiler that turns tags into typed output.
Packages
| Package | Task | Size | Status |
|---|---|---|---|
| gpu-cite | Parse citation and reference strings into structured bibliographic fields | 54.1 KiB | beta |
| gpu-email | Split plain-text emails into reply, quoted history, and signature, and extract contact details | 92.9 KiB | experimental |
| gpu-log | Universal log line parser: timestamps, levels, sources, key=value pairs, messages, stack frames | 97.5 KiB | experimental |
| gpu-paste | Understand pasted text: detect what it is and extract structured fields | 51.0 KiB | experimental |
| gpu-tailwind | Natural language to Tailwind CSS utility classes | 52.3 KiB | experimental |
| gpu-view | Natural language to table view specs: filter, sort, group, aggregate, limit, chart | 29.6 KiB | experimental |
Development
pnpm install
pnpm lint && pnpm typecheck && pnpm test
pnpm build && pnpm size && pnpm check:package
pnpm new gpu-thing "Natural language to thing" # scaffold a package
cd packages/gpu-thing/training && uv sync && uv run python -m gpu_thing.train
bash video/render.sh gpu-thing draft # explainer video
Node 24, pnpm 11, Python 3.12 via uv. See AGENTS.md for the full recipe and definition of done.
Releasing
Add a changeset (pnpm changeset). Merging to main opens a “Version Packages” pull
request; merging that publishes to npm with provenance via npm Trusted Publishing (OIDC), so no
npm token is stored in the repo.
License
MIT