piston

A proof-of-concept library and web playground for training small sequence models such as Transformers and LSTMs directly in the browser using WebGPU. It is for learning and experimentation.

Share on XLicense: MIT

Overview

Piston is a proof-of-concept automatic differentiation library that runs on WebGPU, and Sequence Toy is the web playground built on it. Together they let you train small sequence models such as Transformers, LSTMs, GRUs and vanilla RNNs directly in the browser. Piston is a fork of Ratchet with autodiff added, so it is best seen as a learning and experimentation project.

Key features

  • Trains sequence models in the browser with WebGPU
  • Supports Transformers, LSTMs, GRUs and vanilla RNNs
  • Automatic differentiation library written in Rust
  • Sequence Toy web playground included

Best for

People who want to experiment with training small language models in a browser. The author describes Piston as a proof of concept, so do not expect a production framework.

Upstream
vinhowe/piston
Fork on GitHub
Guo-astro/piston
Upstream stars
36
Category
AI agents and LLM tools
Language
Rust
License
MIT
Forked
2025-11-21
Sync status
In syncLast synced 2026-09-29