[InferenceClient] flag chat_completion()'s logit_bias as UNUSED

[hanouticelina]

Related to #2720.
logit_bias is not implemented in TGI (see the comment here), this PR updates the parameter documentation to flag it as unused.

[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.

[juanwisz]

Could we consider actually implementing logit_bias? It's a very valuable feature for controlling model outputs, I've opened in TGI to discuss implementation details: huggingface/text-generation-inference#2869

[hanouticelina]

Could we consider actually implementing logit_bias? It's a very valuable feature for controlling model outputs, I've opened in TGI to discuss implementation details: huggingface/text-generation-inference#2869

@juanwisz This PR only aligns the client documentation with the current TGI implementation. Once/If logit_bias is implemented on TGI side, we'll update the documentation accordingly 🤗

[LysandreJik]

Is it possible to remove that doc attribute altogether? If not, I would at least specify why it's unused and how to have users request for this argument to be used

[hanouticelina]

Is it possible to remove that doc attribute altogether? If not, I would at least specify why it's unused and how to have users request for this argument to be used

@LysandreJik logit_bias and n (i.e number of chat completion choices to generate for each input message) which is already flagged as unused, are implemented in the OpenAI API, and just like we expose these parameters in TGI to maintain OpenAI compatibility, we want to keep them in the client for the same reason —making InferenceClient a proper drop-in replacements for OpenAI client. I've updated their documentation to indicate they are currently not implemented in TGI while explaining why they are still included.

[LysandreJik]

Ok great! Thank you Célina !