⚠️ branch based on #238

This is a follow-up on #240.

This is not ready to merge, but should be a good starting point to start adapting the structure to support image generation.
The Bagel model is not super clean and has quite a few specific modules which make it difficult to rationalize, but it is IMO a good candidate to explore new modalities (image generation, thinking, etc.)

This also allows to test simple BNB quantization, allowing to fit the whole model on a 24GB GPU (unlike the official code which offloads parts to CPU). For reference, without any optimization, image generation runs at approx. 3 seconds per timestep on a 3090 (+ 5950x cpu) -- 30-50 timesteps being the sweet spot it seems.

What works

• simple vision understanding query (e.g. GDP image + prompt) -> test_bagel_understanding.py
• simple image generation -> test_bagel_generation.py

What needs to be fixed/rationalized

• positions handling currently breaks other vision models, we need to find a proper condition (maybe split out classes as previously discussed)
• image autoencoder settings are hardcoded/copy-pasted from the official code
• some settings are not properly grabbed from config yet
• image transform logic could probably be factorized a bit with current logic
• the image generation codepath triggers an early exit in inference.decode_and_generate, we should probably find a cleaner way to support this (+ support in serving mode)

What needs to be implemented/tested

• image "edition" use case (+ image cfg and co)
• "thinking" step support (might be useful for other models as well)
• multiple image handling (understanding path, probably not supported in generation path yet)
• batch mode