Llama-3-8B gets always requested/required/loaded

[uahic]

When importing (even selective) from conversation.py, the AutoTokenizer.from_pretrained("meta-llama....") gets triggered and tries to access huggingface, and for me personally I have no access to this model which causes an error.

The issue here is that all Conversations get instantiated as soon as anything from this file is imported. There should be a lazy loading way to do that