Feature Request: Support for Meta Chameleon 7B and 34B

[arch-btw]

Prerequisites

• I am running the latest code. Mention the version if possible as well.
• I carefully followed the README.md.
• I searched using keywords relevant to my issue to make sure that I am creating a new issue that is not already open (or closed).
• I reviewed the Discussions, and have a new and useful enhancement to share.

Feature Description

"Meta Chameleon is a family of models that can combine text and images as input and output any combination of text and images with a single unified architecture for both encoding and decoding. While most current late-fusion models use diffusion-based learning, Meta Chameleon uses tokenization for text and images. This enables a more unified approach and makes the model easier to design, maintain, and scale. The possibilities are endless—imagine generating creative captions for images or using a mix of text prompts and images to create an entirely new scene."

Motivation

This would be a great addition to llama.cpp!

The image features look interesting but it can also simply do Text -> Text and a lot of other combinations:

• Text -> Text
• Image -> Image
• Text -> Image
• Image -> Text
• Image -> Text + Image
• Text + Image -> Text
• Text -> Text + Image
• Text + Image -> Text + Image

chameleon.mp4

[EliEron]

Here are some relevant links:

• Paper
• Blog Post
• Code
• Model Form
• Hugging Face

Given it's a completely new architecture, and a multimodal one at that, I imagine adding support for it will not be easy. But I'm also very excited to see this supported.

Edit: According to Meta researcher Armen Aghajanyan the architecture is actually similar:

Similar architecture to LLaMa (apart from QK-norm), get fast inference working.

[0wwafa]

Yes! Please do! Also because as of now there is no way to run it on CPU only.

[SolvAI]

yum yum yum :p

[ann-brown]

Since it uses similar to Medusa architecture is that likely to be supportable at the same time for the self-speculative decoding side of inference? It sounded like it could run without that, but it'd be neat to have that available too.

[jacobkahn]

Let me know if we can answer any questions about the architecture, inference, etc. Our reference implementation in https://github.com/facebookresearch/chameleon should be clear. Differences from the Llama architecture are minor:

• qk-norm in attention after the initial qk weight transformations
• swin-norm (in the 30B model only) which is normalizes inputs before attention and feedforward blocks

[typedrat]

Considering the VQGAN is public, it should be possible for llama.cpp to reinstate the image output capabilities.

[chigkim]

+1000! I'd love to run Chameleon with llama.cpp!

[freeqaz]

In the interim: Is anybody aware of any quantized weights for this model currently? Would love to try to 30b version on my 2x 3090s.

Found this issue because I know llama.cpp is a common way to shrink these models down.

[github-actions]

This issue was closed because it has been inactive for 14 days since being marked as stale.