transformers/docs/source/en/model_doc/fuyu.md

This model was released on {release_date} and added to Hugging Face Transformers on 2023-10-19 and contributed by Molbap.

Fuyu

Fuyu is a small, open-source multimodal model designed for AI agents that can handle both text and images. Unlike most multimodal models, it uses a decoder-only Transformer without a separate image encoder, projecting image patches directly into the transformer and supporting arbitrary image resolutions. This simplified architecture allows fast inference—under 100 milliseconds for large images—and streamlines training by removing multiple specialized stages. Despite its small size, Fuyu-8B achieves competitive performance on standard image understanding benchmarks like VQAv2, OKVQA, COCO Captions, and AI2D, outperforming larger models on several metrics while being easier to scale and deploy.

import torch
from transformers import pipeline

pipeline = pipeline(task="text-generation", model="adept/fuyu-8b", dtype="auto")
pipeline("Plants generate energy through a process known as  ")

import torch
import requests
from PIL import Image
from transformers import AutoProcessor, AutoModelForCausalLM

processor = AutoProcessor.from_pretrained("adept/fuyu-8b")
model = AutoModelForCausalLM.from_pretrained("adept/fuyu-8b", dtype="auto")

url = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/pipeline-cat-chonk.jpeg"
image = Image.open(requests.get(url, stream=True).raw)
prompt = "Generate a coco-style caption.\n"

inputs = processor(images=image, text=prompt, return_tensors="pt")
outputs = model(**inputs)
generated_ids = model.generate(**inputs, max_new_tokens=7)
generation_text = processor.batch_decode(generated_ids[:, -7:], skip_special_tokens=True)
print(generation_text[0])

Usage tips

• Fuyu models trained with bfloat16, but original inference uses float16. Hub checkpoints use dtype='float16'. The AutoModel API casts checkpoints from torch.float32 to torch.float16.
• Online weight dtype matters only when using dtype="auto". The model downloads first (using checkpoint dtype), then casts to torch's default dtype (torch.float32). Specify your desired dtype or it defaults to torch.float32.
• Don't fine-tune in float16. It produces NaN values. Fine-tune in bfloat16 instead.
• Clone the original repository to convert the model: git clone https://github.com/persimmon-ai-labs/adept-inference.
• Pass inputs through a specific Processor for correct formats. A processor needs an image_processor and a tokenizer.
• Fuyu uses a sentencepiece-based tokenizer with a Unigram model. It supports bytefallback (available in tokenizers==0.14.0 for the fast tokenizer). [LlamaTokenizer] wraps sentencepiece as a standard wrapper.
• Use this prompt for image captioning: f"Generate a coco-style caption.\\n".

FuyuConfig

autodoc FuyuConfig

FuyuForCausalLM

autodoc FuyuForCausalLM - forward

FuyuModel

autodoc FuyuModel - forward

FuyuImageProcessor

autodoc FuyuImageProcessor - call

FuyuProcessor

autodoc FuyuProcessor - call