feat: support for passing extra_headers to LitellmChatModel

[ashwin153]

My ollama instance runs behind an reverse proxy (Cloudflare Access) which uses custom headers (cf-access-client-id / cf-access-client-secret or cf-access-token) to authenticate the user. I'm able to access my instance using litellm in the following way, but I'd like to switch to using magnetic.

import litellm

response = litellm.completion(
    model="ollama/deepseek-coder-v2",
    messages=[{"content": "generate a simple sql query", "role": "user"}],
    api_base="...",
    extra_headers={"cf-access-token": "..."},
)

print(response)

closes #260

[ashwin153]

The test fails when I run it locally, but I'm not sure why. Did I set something up incorrectly?

>               _is_litellm_router_call = "model_group" in kwargs.get(
                    "metadata", {}
                )  # check if call from litellm.router/proxy
E               TypeError: argument of type 'NoneType' is not iterable

[jackmpcollins]

@ashwin153 Thanks for this! For this tests you'll need to add a vcr recording so running in CI does not query the live APIs. To do that should just require running make test-vcr-once and committing the generated file. You might also want to read this file to ensure there are no API keys or other info in the headers that should remain private.

See here for an overview of vcr: https://vcrpy.readthedocs.io/en/latest/

---

It's also possible this is a bug and not related to vcr!

[ashwin153]

Ok sweet tests pass now. Thanks for the tip!

[ashwin153]

@jackmpcollins Everything ok here?

[jackmpcollins]

@jackmpcollins Everything ok here?

@ashwin153 Thanks for the ping! Looks great. I just changed the test to use openai api for the VCR and to check the litellm callback.