What is Shumai?
Shumai is an experimental, fast differentiable tensor library for TypeScript and JavaScript, developed by Meta (Facebook Research). Built on top of Bun and Flashlight, it provides a network-connected tensor computation framework designed for both software engineers and researchers. Shumai aims to bridge the gap between high-level scripting languages and high-performance numerical computing, enabling seamless integration of machine learning and scientific computing into JavaScript/TypeScript applications.
Key features include automatic differentiation for gradient computation, support for standard array operations (creation, reshaping, slicing, broadcasting), and efficient conversion between JavaScript native arrays and tensors. It leverages the ArrayFire library for backend computation, with GPU acceleration via CUDA on Linux and CPU computation on macOS. The library is designed to be intuitive for developers familiar with NumPy or PyTorch, offering a similar API style.
Benefits include the ability to run complex numerical computations directly in JavaScript/TypeScript environments, reducing the need for Python or C++ dependencies. This is particularly useful for web-based machine learning applications, real-time data processing, and prototyping. Use cases span from training small neural networks to implementing differentiable algorithms in browser or server-side JavaScript (via Bun).
Technical details: Shumai requires Bun and ArrayFire to be installed. Currently, only macOS and Linux are supported, with Linux defaulting to GPU (CUDA) and macOS to CPU. The library is in early development, so performance may vary, and the CPU backend is not fully optimized. Installation is via Bun package manager, and source builds are available for custom setups. Windows users can leverage Docker with WSL2 for Linux compatibility. The project is open-source on GitHub, with documentation and examples provided.
Use Cases
Key Features
automatic differentiation tensor operations GPU acceleration NumPy like API JavaScript integration network connected computation efficient array conversion Flashlight backendWho is it for?
software engineers, machine learning researchers, JavaScript developers, data scientists, scientific computing practitioners, TypeScript developers
