xnn

A lightweight ML framework built from scratch in Rust with GPU-first architecture.

Features

• GPU acceleration via wgpu (Vulkan, Metal, DX12, WebGPU)
• Element types: f32, i32, u32, bool
• Cross-platform: Linux, macOS, Windows, Web/WASM
• Automatic compute pipeline caching
• No unsafe code

Tensor

N-dimensional array with GPU-accelerated operations and automatic broadcasting.

Elements

Type	Numeric	Signed	Integer	Float	Logical
f32	✓	✓		✓
i32	✓	✓	✓
u32	✓		✓
bool					✓

Examples

MNIST Training

Trains a fully-connected network on the MNIST dataset.

cd examples/mnist-train
cargo run --release

MNIST Web

Interactive digit recognition in the browser using WebGPU and WASM.

cd examples/mnist-web
wasm-pack build --target web
python3 -m http.server

Linear regression

Trains a simple linear model to fit y = wx + b.

cargo run --release --example linreg

XOR

Trains a 2-layer neural network to solve the XOR problem.

cargo run --release --example xor

License

MIT — see LICENSE for details.