llama.cpp

This is a wrapper of llama.cpp implemented as per the discussion Integration of llama.cpp and whisper.cpp:

• Use the llama.cpp C interface in llama.h
• Reimplement the common library

As mentioned in the discussion the (maybe distant) future plan is to ditch llama.cpp by reimplementing entirely with vanilla ggml and a C++ interface.

Reimplementation Notes:

• Better error handling, please
• GGUF metadata access (llama_model_meta_*) is not great. We should provide a better interface
• llama_chat_apply_template does not handle memory allocation optimally. There's a lot of room for improvement
  • as a whole, chat management is not very efficient. llama_chat_format_single doing a full chat format for a single message is terrible
• Chat templates can't be used to escape special tokens. If the user actually enters some, this just messes-up the resulting formatted text.
• Give vocab more visibility
• Token-to-text can be handled much more elegantly by using plain ol' string_view instead of copying strings. It's not like tokens are going to be modified once the model is loaded
  • If we don't reimplement, perhaps keeping a parallel array of all tokens to string would be a good idea
• llama_batch being used for both input and output makes it hard to propagate the constness of the input buffer. This leads to code having to use non-const buffers, even if we know they're not going to be modified. We should bind the buffer constness to the batch struct itself.
• The low-level llama context currently takes a rng seed (which is only used for mirostat sampling). A reimplemented context should be deterministic. If an operation requires random numbers, a generator should be provided from the outside.
  • For now we will hide the mirostat sampling altogether and ditch the seed
• As per this discussion we should take into account how we want to deal with asset storage and whether we want to abstract the i/o away.