Add default template warning

[Rocketknight1]

Lots of user repos use common tokenizers like LlamaTokenizer even for non-LLaMA models. Unfortunately, these tokenizers come with default chat templates, which may result in user confusion if they use apply_chat_template. We want to avoid a scenario where users think they're getting the right template but actually aren't!

This PR raises a short warning the first time a default_chat_template is read, which happens when apply_chat_template or ConversationalPipeline is called for a model without a chat_template set. The warning tells the user what's happening, and suggests adding an explicit chat template.

Over time, the goal is to eventually deprecate and remove default_chat_template, because we want to avoid using class-level templates entirely, and move to explicit repo level chat_template attributes. This PR starts that process, while still retaining the feature for backward compatibility.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint.

[ArthurZucker]

If we keep this, let's leverage the futur warning. But I am not sure I fully understand what this add, the default template was always used by default for a few of these models, and if we want to improve/grow the template feature we should probably raise and error to make sure the repo owners push a default template (instead of defaulting in the class) which is breaking but helps moving towards template on the hub.
Also if the template is a default template but pushed to the hub should not warn / raise anything.

[LysandreJik]

Hmmm I'm not sure this warning will be useful for the majority of users; wouldn't it be better as an INFO level?

I think having class-level templates is better than not having anything, and IMO it would make sense to have a sensible default (original architecture training scheme) rather than forcing each checkpoint to have their own definition. Or do you expect every checkpoint to have a different template?

[Rocketknight1]

Just from looking at models on the Hub, I think a lot of models use templates totally different from the default class one. For example, if you look at the top trending models on the Hub today, the top model is Mistral-7B-instruct, which uses LlamaTokenizer with a custom chat template, and the second place model is Mistral-7B-OpenOrca, which also uses LlamaTokenizer but with a ChatML template. Neither of these templates match the class-level template.

The problem right now is if a user uses apply_chat_template with a model like that which doesn't have a chat_template attribute, it will appear to work, but they'll silently get the wrong, class-level template instead! This is exactly what we want to avoid with chat templates.

I think raising a warning is the right approach - it won't break existing workflows, but it will flag a potential source of issues, and it might notify users and maintainers to add proper chat templates to the models they're using.

[LysandreJik]

Ok, sounds good to me in that case. Thanks for the explanation!

[Rocketknight1]

Sure! Also @ArthurZucker to respond to your comments, I'm probably going to start with this, then maybe we can begin deprecating default_class_template over time, but I want to do it slowly because it breaks old workflows.

The FutureWarning suggestion was good, though, I'll add that now!

[Rocketknight1]

FutureWarning added!

[Rocketknight1]

Update: After testing, I took FutureWarning out because it clashes with warning_once, and I don't want to spam these messages too much!

[LysandreJik]

Let's merge it!