Shumai

Shumai is a fast differentiable tensor library written for TypeScript and JavaScript, designed to run with Bun and the Flashlight stack. It targets software engineers and researchers who need tensor operations, autograd, and neural network utilities within a JavaScript/TypeScript environment. The project is experimental software and invites issues if problems arise. It emphasizes networked, differentiable tensors with GPU-backed computation when available.

The documentation suggests that different platforms have varying backend requirements and that issues are welcome to improve the project. The tool aims to be accessible for developers who want to experiment with differentiable tensors in JS/TS ecosystems.

The core capabilities include creating tensors, performing matrix operations, converting between JS arrays and tensor data, handling scalar tensors, and computing gradients with backpropagation. The project provides examples for random tensors, matrix multiplications, and gradient computation, along with notes on detaching gradients and basic arithmetic operations. It also signals that more examples and operator references are available within the project.

Shumai lists platform availability and installation notes, including macOS and Linux guidance, Homebrew-based ArrayFire installation, and Bun-based installation steps. It notes that Windows support exists as a non-official path using Docker + WSL2 + Linux, and it describes that the installation process is a work in progress. The documentation also covers building from source and profiling, with specific mentions of backend options and CUDA/OpenCL considerations.

Use cases

• Experiment with differentiable tensors and autograd in JS/TS