adding litellm integration into toolkit ENG-1086#30
Conversation
acompa
left a comment
There was a problem hiding this comment.
I generally worry about bundling external code with an SDK wholesale in this manner. Let's work together during next sprint to find a lightweight path to introducing litellm support without bringing over their main.py or __init__.py files.
| "content": "Hey", | ||
| }, | ||
| ] | ||
| for model in ND_MODEL_LIST: |
There was a problem hiding this comment.
I missed this in Devansh's PR, but we should probably use @pytest.mark.parametrize here to separate these tests (and optionally run them in parallel via pytest -n, see details).
I think one way to reduce copied code is to only copy |
|
@r0ymanesco Do we still need the enormous |
This is the minimum amount of |
Yep, I actually spent part of my flight up north reviewing this further. I agree. Approving to unblock you, and we can chat with the LiteLLM team about this contribution experience. |
This PR adds our
litellmintegration PR into the ND python SDK. I implemented this in the most brute forced way by copying themain.pyfile inlitellmdirectly into ourtoolkitmodule. This is becauselitellmis implemented purely modularly without any classes so I couldn't just inherit a class and modify some methods. The up side is that this is a complete replacement forlitellm. A user can simply use it as they are using the regularlitellmwith the added benefit of supporting notdiamond.