Add 🦀 Rust port with safety + CUDA to notable forks

[phildav]

Motivations

Rust is focused on performance and reliability (cf Rust), so I found it particularly interesting to see how these principles apply to an LLM training loop.

Although there are 2 Rust projects listed in notable forks, their focus is not on safety or CUDA support. This was my motivation starting this new port.

I tried to make the code as idiomatic and safe as possible, while keeping the CPU/GPU path close to the llm.c structure for easier comparison and educational value.

Features

• 100% safe CPU training loop
• safe (minimal unsafe blocks for the GPU kernel launchers) and idiomatic Rust
• CUDA support
• training results are bit-exact with the reference implementation
• devcontainer and modal.com support for the GPU poor 🙋

Link

llm.rs 🦀 https://github.com/phildav/llm.rs

Performances

Full description in the README.md

Here is a quick preview: