Chameleon: add model

[zucchini-nlp]

What does this PR do?

Fixes #31505.

Adds Chameleon, a vision language model from Meta AI.

from transformers import ChameleonForCausalLM, ChameleonProcessor
from PIL import Image
import requests
import torch

model_path = "MODEL_PATH"
model = ChameleonForCausalLM.from_pretrained(model_path, torch_dtype=torch.bfloat16, low_cpu_mem_usage=True, device_map="auto")
processor = ChameleonProcessor.from_pretrained(model_path)

prompt = "I'm very intrigued by this work of art:<image>Please tell me about the artist."
image = Image.open(requests.get("https://uploads4.wikiart.org/images/paul-klee/death-for-the-idea-1915.jpg!Large.jpg", stream=True).raw)

inputs = processor(prompt, images=[image], return_tensors="pt").to(model.device, dtype=torch.bfloat16)
out = model.generate(**inputs, max_new_tokens=40, do_sample=False)
generated_text = processor.batch_decode(out, skip_special_tokens=False)[0]
print(f"Generated text: {generated_text}")

# Multi-image example
prompt = "I used to know a lot about constellations when I was younger, but as I grew older, I forgot most of what I knew. These are the only two constellations that I really remember now.<image><image>I would like for you to tell me about 3 more constellations and give me a little bit of history about the constellation."
image = Image.open(requests.get("https://nineplanets.org/wp-content/uploads/2020/12/the-big-dipper-1.jpg", stream=True).raw)
image_2 = Image.open(requests.get("https://www.kxan.com/wp-content/uploads/sites/40/2020/10/ORION.jpg", stream=True).raw)

inputs = processor(prompt, images=[image, image_2], return_tensors="pt").to(model.device, dtype=torch.bfloat16)
out = model.generate(**inputs, max_new_tokens=200, do_sample=False)
generated_text = processor.batch_decode(out, skip_special_tokens=True)[0]
print(f"Generated text: {generated_text}")
>>> 

Project repo: https://github.com/facebookresearch/chameleon
Paper: https://arxiv.org/abs/2405.09818v1

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[jsm69]

@amyeroberts @ArthurZucker

[ArthurZucker]

you should be able to bake that by updating the alpha and beta for model parallelisme

[ArthurZucker]

IMO we should permute everything in the weights

[ArthurZucker]

does this mean we are computing over say "head_dim" ?
How different is this from normal nn.layer_norm((head_dim,))
?

[ArthurZucker]

Ah the weights are different (hidden_size) but the applied dim is hidden_size is that it?

[zucchini-nlp]

Yes, it's only the weights that are different for 30B model. The 7B has simple repeated weights over all heads

[leloykun]

@zucchini-nlp @ArthurZucker We should also set token["special"] = False so that we can decode this token.

What do you guys think?

(It's what I'm currently doing in my PR btw and I haven't encountered any errors yet)

[zucchini-nlp]

Hi! Special tokens should be decodable with skip_special_tokens=False

For my understanding, why do we need to decode the image token? Afaik it shouldn't affect image generation because it's a token we added manually to keep track of where to add an image in the text

[ArthurZucker]

🔥 great work