Show HN: jax-js – JAX in pure JavaScript (github.com)

A new machine learning framework called jax-js has been launched, designed to operate in web browsers and deliver JAX-style performance using JavaScript. By translating array operations into a compiler-friendly format and synthesizing kernels in WebAssembly and WebGPU, jax-js enables developers to run high-performance numerical applications entirely client-side, making it one of the most portable GPU ML frameworks available. It closely mirrors the API of JAX and NumPy, providing familiar syntax for JavaScript developers. The significance of jax-js lies in its potential to democratize machine learning by making powerful tools accessible within web environments without requiring extensive backend infrastructure. Its unique memory management system uses reference counting to help developers manage resources efficiently, which is crucial for performance in JavaScript's garbage-collected environment. With support for similar composable transformations as JAX and optimized operation fusions through JIT compilation, jax-js offers a valuable alternative for in-browser ML that could rival existing libraries like TensorFlow.js and ONNX Runtime Web.