Shumai: Fast Differentiable Tensor Library for JS/TS

Shumai is an open-source tensor library built for developers who want the power of machine learning tooling without leaving the JavaScript ecosystem. Powered by Bun and the Flashlight C++ engine, it delivers fast, differentiable tensor operations that support gradient-based optimization, making it possible to train models and run advanced inference logic directly in TypeScript or JavaScript. Network connectivity is baked in, so tensors and data can move seamlessly across distributed setups. Designed by Meta for software engineers, researchers, and data scientists, Shumai focuses on three goals: simplifying dataset creation, speeding up training for smaller models, and making fine-grained training and inference logic more expressive. Whether you're prototyping a new model architecture or building an application that needs on-the-fly tensor math, Shumai aims to keep the developer experience enjoyable rather than bogged down in boilerplate. Under the hood, Shumai leverages CUDA for GPU acceleration on Linux and falls back to CPU computation on macOS, giving it flexibility across common development environments. Because it's built on Bun, a fast all-in-one JavaScript runtime, and Flashlight, a performant ML backend, Shumai bridges the gap between native-speed tensor computation and the accessibility of the JS/TS language ecosystem.

1. Install Bun, the JavaScript runtime Shumai depends on. 2. Install ArrayFire, a required backend library for tensor operations. 3. Run 'bun install @shumai/shumai' to add the Shumai package to your project. 4. Use Shumai's array utilities to create tensors, converting to and from native JavaScript arrays as needed. 5. Leverage differentiable tensors and gradients to build and train your models or inference pipelines.