surfgrad

surfgrad is a high-performance, WebGPU-powered AutoGrad library that enables browser-based tensor operations with GPU acceleration.

Key Features:

• 🚀 Blazing-fast tensor operations leveraging WebGPU
• 🧠 Automatic differentiation for deep learning in the browser
• 🌐 Zero backend dependencies - runs entirely client-side
• 📦 Lightweight and easy to integrate into existing web projects

Perfect for running tensor operations and (in the future) machine learning models in the browser!

It's heavily inspired by micrograd, tinygrad, and PyTorch and aims to leverage the power of WebGPU/WGSL for in-browser machine learning.

Usage

surfgrad supports basic tensor operations such as matmul, mul, add, exp, and log.

To use surfgrad,

import { Tensor } from "surfgrad";

const tensorA = new Tensor(new Float32Array([1, 2, 3, 4]), [2, 2], true);
const tensorB = new Tensor(new Float32Array([5, 6, 7, 8]), [2, 2], true);

const [result, executionTime] = await tensorA.matmul(tensorB);

console.log(result);

await result.backward();

Testing

surfgrad has unit tests and integration tests. To run the unit tests, run the following command:

npm run unit

and to run the integration tests, run the following command:

npm run integration

Benchmarks

We also have benchmarks that can be helpful to demonstrate the performance of the matmul kernels. To run the benchmarks, run the following command:

npm run benchmark

and open a browser to localhost:9000.

This will run the benchmarks for the library and display the results.

Contributing

Contributions to surfgrad are welcome! If you'd like to contribute, please fork the repository and submit a pull request.

License

SurfGrad is licensed under the Apache 2.0 License.