ML-Rust

A family of machine learning libraries written from scratch in Rust, with zero ML dependencies.

Libraries

Library	Focus	Size
mlrust-core	Shared tensor & autograd foundation	~1,200 lines
mlrust	Transformers & NLP	~10,000 lines
MLRustFoundation	CNNs & computer vision	~2,600 lines

See ECOSYSTEM.md for the full architecture, feature comparison, and integration examples across the three libraries.

Quick start

[dependencies]
mlrust-core = { path = "../mlrust-core" }

use mlrust_core::Tensor;

let a = Tensor::ones((2, 2), true);
let b = Tensor::ones((2, 2), true);

let c = a.add(&b);
let mut loss = c.sum();
loss.backward();

println!("gradient of a: {:?}", a.grad());

Each crate has its own README with install instructions, quick-start examples, and a rundown of what is (and isn't) implemented yet:

• mlrust-core/README.md
• mlrust/README.md
• mlrust-transformers/README.md
• MLRustFoundation/README.md

Building

Each library is its own Cargo project:

cd mlrust-core && cargo test
cd MLRustFoundation && cargo test
cd mlrust-transformers && cargo test

Philosophy

Clarity over performance. These are learning projects — every component is written to be readable and easy to follow rather than optimized for production workloads.

License

MIT - see LICENSE.