Make mistral-common dependency optional

[juliendenize]

In PRs #14737 #15420, we added our format support for conversion, and therefore mistral-common dependency for conversion, and made chat templates optional. This is because we noticed that chat templates and custom tokenization are generally not matching mistral-common which can affect model performance.

The goal of these PRs was to provide to the community the freedom to experience the best use of our models with the llama.cpp engine.

However part of this was not well welcomed by the community that particularly disliked having mistral-common as a hard dependency as discussed in #16146. This PR aims to remove this hard dependency and instead raise an error if it is not installed. This occurs for converting Mistral models for the following cases:

• the model conversion is done with our format
• the model conversion is done with transformers format except for the tokenizers. This is what happens for our releases now as we do not not release a tokenizer config.

Fix #16146

cc @CISC @ngxson hope this matches what was dicussed in #16146 and that it passes the CI in particular for typing.

[CISC]

Moving the imports won't fix the typings btw. :)

You need to use try/except ImportError instead of importlib.

Edit: Actually, I see it's now complaining about vocab.py as well, so you might have to silence the errors per line as well...

[juliendenize]

Yep I tried moving things a bit and it didn't work so I just added the flags to ignore the relevant lines if that's ok.

[CISC]

Yep I tried moving things a bit and it didn't work so I just added the flags to ignore the relevant lines if that's ok.

Yep, though I think try/except ImportError is still preferable. Also, since we are no longer installing mistral-common, you now need to silence errors in gguf-py/gguf/vocab.py as well.

[LostRuins]

Thank you for taking our concerns into account.

[CISC]

LGTM, thank you! However I'm currently away travelling so great if @ngxson can take a final look and potentially merge.

[ngxson]

Nice! Thanks for addressing the issue