implement groq completions

[stargazing-dino]

Implements groq api. See their quickstart

Which for the most part follows the openAI schema.

One relatively larger change this PR made was to the AdapterKind::from_model.

The models from groq do not have any common prefix or suffix to do partial matching on so I had to change it to just match the exact string. For that I made the models pub. I thought to use Adapter::list_models but it required kind to be passed through which was unfeasible since we don't have one at that point and it's also async so it'd require some more changes than just making the models pub to the crate.

I tested this locally and it seemed to work fine.

[jeremychone]

First, thanks so much for this PR.

The only thing I'm not a fan of right now is losing the "open" matching for the models that have clear prefixes.

For the from_mod, could we make it so that only Grok has the fixed name match, and the others remain as before?

If you do not have time to make the change, I will do it, no problem.

Having the fixed names for the list_models is okay for now. That will change later as I would like to have them query when the API is available.

Note: by design, the list_models is decoupled from the from_models, as the first one is more for user info and tries to be as live as possible, and the second is more for a wider match focused on in-memory/static processing.

[stargazing-dino]

I moved back to the prefix solution :)

Note: by design, the list_models is decoupled from the from_models, as the first one is more for user info and tries to be as live as possible, and the second is more for a wider match focused on in-memory/static processing.

ah makes sense

[jeremychone]

Perfect, I will merge this.

So Grok supports the OpenAI API to the letter?
I found one bug in the Ollama OpenAI compatibility layer (it does not support multiple system messages), but I worked around it, so it all works. Hopefully, Grok does not have those types of issues.

[stargazing-dino]

So Grok supports the OpenAI API to the letter?

I think so, this is even their endpoint https://api.groq.com/openai/v1/chat/completions

it does not support multiple system messages

I've tested groq and from what I remember it supports multiple system messages (for the llama models I tested at least) fine so no issue there hopefully

[jeremychone]

Cool. I found this page: https://console.groq.com/docs/openai

So it seems OpenAI API is their API strategy, so they should support it pretty well.

I am going to add the temperature, max_token, … all in the ChatRequestOptions, so we will get that soon, and that will include Grok.

[jeremychone]

@stargazing-dino Thanks, I just merged it.

[jeremychone]

Btw, very clean code and PR. Thanks!