Document the difference between model and base_url

[Wauplin]

cc @MoritzLaurer who pinged me about this on slack (private link).

It is wrong to say that base_url and model have the exact same behavior since #2410 have been merged. The base URL is used to build the complete URL by appending (/v1)?/chat/completions to it. On the contrary, when an URL is passed as model, we use it directly without appending anything (e.g. we don't consider it as a "base url").

This PR fixes this confusion in the docs.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[MoritzLaurer]

Thank you for adding the clarification!
Small point: I assume that not all users will be aware of chat completion and the chat/completions suffix. Maybe you could add a note like "... appended to the base URL (see the TGI Messages API documentation for details) ...".

[MoritzLaurer]

Does the use of base_url always append the chat completions suffix and is therefore only compatible with models in a TGI container? if so, this could maybe be made more explicit (If using base_url for chat completion sounds like I can also use base_url for other things than chat completion).

[Wauplin]

Thanks for the review @MoritzLaurer! I think we are good to merge now :)

Small point: I assume that not all users will be aware of chat completion and the chat/completions suffix. Maybe you could add a note

Good idea! Addressed it in b589816

Does the use of base_url always append the chat completions suffix and is therefore only compatible with models in a TGI container?

Yes it always append the chat completions suffix when using client.chat_completion. This means it's compatible with TGI but also all OpenAI-compatible providers.
If the user sets base_url for another task (say text_to_image), no suffix will be appended so it should work as well. This is not really a recommended/expected behavior but "it works" (passing base_url or model for any non-Chat Completion method is strictly the same).