I did a similar thing for Julia: Llama2.jl contains vanilla Julia code [1] for training small Llama2-style models on the CPU.

[1] https://github.com/cafaxo/Llama2.jl/tree/master/src/training