[Feature]: Add support for Llamafile provider

[rupurt]

The Feature

Support connecting to Llamafile models https://github.com/Mozilla-Ocho/llamafile

Motivation, pitch

It's a self contained format to run models as a single binary

Twitter / LinkedIn details

https://twitter.com/rupurt

[krrishdholakia]

@rupurt they're openai-compatible https://github.com/Mozilla-Ocho/llamafile?tab=readme-ov-file#json-api-quickstart

You can already call them with litellm - https://docs.litellm.ai/docs/providers/openai_compatible

Let me know if there's anything i'm missing.

[peteski22]

The problem with just using openai as the provider is that means you're asking LiteLLM to use the OpenAI client from the OpenAI Python API library.

OpenAI client requires an API key either from the args (api_key) or falls back to the environment variable (OPENAI_API_KEY).

If neither are present (e.g. None) then the client errors.

Given that Llamafile doesn't need an API key, then I'd suggest the hosted_vllm provider route ('for OpenAI compatible server'). It seems preferable over having to ensure you send a fake API key in a request (hosted_vllm does it for you).

See: https://docs.litellm.ai/docs/providers/vllm

e.g. ./mistral-7b-instruct-v0.2.Q4_0.llamafile --server

from litellm import completion

msg = "What is the meaning of life?"
messages = [{"content": msg, "role": "user"}]

response = completion(
    model="hosted_vllm/Mistral-7B-Instruct-v0.2",
    base_url="http://127.0.0.1:8080/v1",
    messages=messages,
)

reply = response["choices"][0]["message"]["content"]

print("Completion Result:\n")
print(f"User: {msg}\n\nAssistant: {reply}\n{'-' * 40}")

[krrishdholakia]

Hey @peteski22 that's fair - would you be able to contribute a PR for this?

[peteski22]

Hey @peteski22 that's fair - would you be able to contribute a PR for this?

Thanks @krrishdholakia, I'll take a look into it and see if I can raise a PR (it might not be immediate though) 😄.

[peteski22]

Hey @peteski22 that's fair - would you be able to contribute a PR for this?

Thanks @krrishdholakia, I'll take a look into it and see if I can raise a PR (it might not be immediate though) 😄.

Hey again @krrishdholakia, sorry for the delay. I've created #10203.

If you have any time to review it, I'd be glad to get your feedback and see if anything needs to be done to move it along/get it merged.

Thanks 😄

[krrishdholakia]

Left comments there