It uses "ggml", a tensor library written in C to do the math. The models weights are converted to a format that can be used by GGML. The higher level structure

It uses "ggml", a tensor library written in C to do the math. The models weights are converted to a format that can be used by GGML. The higher level structure of the model is created in Rust be leveraging ggml data structures and the weights are passed into the C library. PyTorch is only needed for reading the original weights while converting them.