[Bug]: gpt-3.5-turbo-instruct not supported

[StanGirard]

What happened?

Tried to use gpt-3.5-turbo-instruct on Quivr

Relevant log output

https://smith.langchain.com/public/7814d1a6-d387-4333-b893-35af5838d4ca/r

Twitter / LinkedIn details

No response

[krrishdholakia]

cc: @ishaan-jaff

[krrishdholakia]

@StanGirard acknowledging the bug, I'll pick this up and revert back with a fix in a few hours.

[krrishdholakia]

@StanGirard which version of litellm are you using. gpt-3.5-turbo-instruct was added a few days ago. If you could bump the version and let me know if that works, it'd be great!

I think the issue is that the model list in the older version doesn't contain the instruct model name.

[krrishdholakia]

Sharing as proof. This works for me in 0.1.714

[krrishdholakia]

As a potential improvement, we could decouple the model lists from the local package. This would make it easy to update lists and ensure everyone has the latest version.

thoughts @ishaan-jaff / @StanGirard

[StanGirard]

I'm on litellm==0.1.531

That is probably why ;)

Yeah that would be amazing to decouple it.

[StanGirard]

It works

[krrishdholakia]

Sounds good - I'll close this issue, and open a new one to track the decoupling idea.

[aleclarson]

In my PR for integrating LiteLLM with Aider, I'm updating the backup whenever the backup is older than 12 hours on Aider startup. If the backup is newer than 12 hours, I set the LITELLM_LOCAL_MODEL_COST_MAP to True (before importing LiteLLM) so the backup is used instead of hitting the network. I wonder if this approach should be merged into LiteLLM itself?

https://github.com/aleclarson/aider/blob/41ca019854a81cc49eb99135fc6977c0e0d03354/aider/main.py#L596-L629