LLMs from scratch - Rust & Burn

Burn: A next generation Deep Learning Framework.

Many thanks to LLMs from scratch - Rust (candle).

Backends

Burn & Wgpu should be compatible with any operating system and will use the GPU.

Usage

Clone this repo, and use cargo to run examples ch02 - ch05:

cargo run --example ch02

By default, Wgpu will be used. In some cases (rare though), Wgpu backend is known to generate wrong results. If that happens, you can use the ndarray feature to recheck the results:

cargo run -F ndarray --example ch02

Note

For ch05, gpt2-small-124M.safetensors needs to be downloaded from Hugging Face

You will need to put it under the assets folder.