Hi there! We’ve been using Gemma via LiveKit Inference in production since August 24th around 5pm PDT. Before deploying the Gemma-powered agent to production we tested it internally including uploading images which Gemma was able to handle natively completely fine. For less than a day after our deployment to production, users were able to upload images completely fine. On August 25th around 8:55 am PDT we started getting this error in our logs:
APIStatusError: message='image input is not enabled: message 24 contains image_url content but image (vision) input is currently disabled for this model', status_code=400, retryable=False, body=image input is not enabled: message 24 contains image_url content but image (vision) input is currently disabled for this model
We didn’t change anything on our end and were wondering whether there were some changes with livekit’s Gemma deployment through LiveKit Inference affecting image handling.
Looks like image support was removed because of stability issues.
I would (personally) classify image support as an ‘undocumented feature’ during the initial rollout, since we never claimed support for it in the docs.
Let me see if I can get an official word on Images for Gemma via LiveKit Inference, to see if it’s on the roadmap.
Hi @darryncampbell - this is useful clarification. I do think it’s worth calling out this limitation on the docs. None of the model/provider pages in Livekit’s docs specifically mention multimodal capabilities, and since Gemma is itself a multimodal model it seemed reasonable to assume those capabilities would work as they do with other inference providers.
Thanks, that’s fair feedback. And the other way also - once we get Images re-enabled we should also have a code sample in the docs showing how to use them.
