Jacob Kahn

@zucchini-nlp — great to hear things are working. In terms of generating output from images + text reliably, one strategy that works is to narrow the scope of questions asked...

Two quick notes: 1. The assertion in the conversion script for the 7B model for the test generation fails when run on CPU (still generates reasonable output) ``` AssertionError: Generations...

@zucchini-nlp yay!! Almost there. A few things: - 7B model looks great overall, works across a bunch of setups too - I can't seem to run the 30B model: I...

@Joeycho this is inference code modeled roughly off of https://github.com/facebookresearch/chameleon and the existing Llama recipe in Transformers.

> Should the code in this pull request be able to generate images, provided that the model is further fine-tuned not to avoid generating images or a new model is...

> I'm curious whether I can fine-tune the Chameleon model. Ofc, I have to come up with how to fine-tune (which parts to be frozen and which part to train)....

@zucchini-nlp — just made the 7B and 30B repos public; will populate them once this PR is merged. @ArthurZucker any other blockers to merge?

@zucchini-nlp — weights are up in https://huggingface.co/facebook/chameleon-7b and https://huggingface.co/facebook/chameleon-30b, we can run slow CI now with those repo slugs