Download
Shumai is an experimental differentiable tensor library for TypeScript and JavaScript, developed by Facebook Research. It provides a high-performance framework for numerical computing and machine learning within modern JavaScript runtimes. Built on Bun and Flashlight, with ArrayFire as its numerical backend, Shumai brings GPU-accelerated tensor operations, automatic differentiation, and scientific computing tools directly to JavaScript developers. It allows seamless integration of machine learning, deep learning, and custom differentiable programs into web-based or server-side environments without relying on Python frameworks. The library supports matrix operations, gradient computation, and tensor conversions with intuitive APIs and near-native speed, thanks to Bun’s low-overhead FFI bindings. It can automatically leverage GPU acceleration on Linux (via CUDA) and CPU computation on macOS.
Features
- Fast GPU-accelerated tensor operations powered by ArrayFire and Flashlight
- Automatic differentiation with flexible gradient control and detachment
- Supports matrix multiplication, elementwise operations, and data conversion between JS arrays and tensors
- Built with Bun for high-speed JIT and minimal FFI overhead
- Cross-platform support for macOS (CPU) and Linux (CUDA GPU)
- Integrated statistics logging, profiling, and memory tuning APIs
Project Samples
