It's just difficult to wrangle all of the dependencies; I want to be able to wrap an entire model in a complete isolated black box that I can call into with a C API or similar.

That is, I'd like something like https://github.com/ggerganov/llama.cpp/blob/master/llama.h without having to rewrite the entire model.

For my use cases, native would be good, but web would be a nice to have. (With enough magic, a native solution could be potentially compiled to WebAssembly?)