Olivier Chafik

@pepijndevos thanks for testing things out! For "weak" models, it looks like we need to pass parameters that cause greedy sampling, e.g. `"top_k": 1, "top_p": 1, "temperature": 0` (depending on...

Extradited the Python agent code w/ docker siloing [to this gist](https://gist.github.com/ochafik/9246d289b7d38d49e1ee2755698d6c79), updated this PR's description with simpler instructions. Also, `--jinja` support was merged: https://github.com/ggerganov/llama.cpp/pull/11016

Hey guys, this is now hopefully ready to review (see non-blocking todos below). - Added `common_chat_params` to hold both inference params (prompt, grammar, triggers...) and the output parser (created directly...

> I'm wondering if it can be refactored, for example: > > * Add a new enum `CHAT_PARSER_TYPE` > * Adds a new function `parse_chat(CHAT_PARSER_TYPE, const std::string & input)` that...

> I'm curious if there is any plan going forward to streamline the --chat-template-file usage? @pepijndevos Hopefully eventually all models will bundle the right template(s) in the GGUF and that...

> I'm getting this error while running `test_tool_call.py` btw: @ngxson hopefully fixed (slight hack), the bartowski version of the model i switched to is marking (correctly) `` as a special...

@ngxson I think this is mergeable once you're happy with it; had to disable the plain non-tools jinja test for now (not critical as i've only introduced it to support...