yesterday's updates broke all image processing via litellm custom endpoints for vision models?

[gmerritt]

Title pretty much sums it up -- vision models behind litellm used to take images just fine, but after yesterday's updates all models just come back with "I didn't get anything" or "I am a language model and cannot see" etc.

Accessing the same vision models directly without litellm (LC --> vendor endpoint, like OpenAI) still works fine for image uploads.

Has anyone else tried yesterday's update with vision models via litellm?

Thanks!

[danny-avila]

Hi @gmerritt

I think my "fix" for the jais chat model from yesterday was too aggressive and created an issue with non-text blocks. Will fix this shortly.