Implement logit_bias correctly

[juanwisz]

Feature request

Currently, logit_bias is labeled as unused https://github.com/huggingface/text-generation-inference/blob/main/router/src/lib.rs

Also, the documentation states that a JSON is needed with a mapping from token_ids to a number between -100 and 100. But this is misaligned with the code's typing, that asks for a vector of floats.

Motivation

Logit_bias is a very important parameter, the documentation in InferenceClient from huggingface_hub states that it can be used, but it does not work.
Also see:
huggingface/huggingface_hub#2720

Your contribution

I can definitely help building a PR but I will need details on what type of solution is expected.

[joetaylor94]

Just seconding the fact that having this parameter implemented would be an incredibly useful feature! It would be great if we could consider implementing it!

[danecor]

I would love to have this parameter as well, any guidance on implementing it?

[eric-tramel]

Strong agree

[drbh]

Thanks everyone for highlighting this! I've started a PR here: #3186