app.py

import gradio as gr
import spaces
import os
import torch
from transformers import AutoProcessor, MllamaForConditionalGeneration
from PIL import Image

# Hugging Face token
hf_token = os.getenv("HUGGING_FACE_HUB_TOKEN")
if not hf_token:
    raise ValueError("HUGGING_FACE_HUB_TOKEN not found.")

# Model
model_name = "meta-llama/Llama-3.2-11B-Vision-Instruct"
model = MllamaForConditionalGeneration.from_pretrained(
    model_name,
    token=hf_token,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
processor = AutoProcessor.from_pretrained(model_name, use_auth_token=hf_token)

@spaces.GPU
def predict(image, text):
    messages = [
        {"role": "user", "content": [
            {"type": "image"},
            {"type": "text", "text": text}
        ]}
    ]
    
    input_text = processor.apply_chat_template(messages, add_generation_prompt=True)
    inputs = processor(image, input_text, return_tensors="pt").to(model.device)
    outputs = model.generate(**inputs, max_new_tokens=250)
    response = processor.decode(outputs[0], skip_special_tokens=True)
    # Split the response at the first occurrence of "assistant" and return only the part after it
    response = response.split("assistant", 1)[1].strip()
    return f"\n{response}"

# Example photos and prompts
examples = [
    {"image": "Cowboy Hat.jpg", "text": "Describe the photo"},
    {"image": "Kynda Coffee.jpg", "text": "Search for the business name on his t-shirt to get a description of where the person is."},
    {"image": "Norway.jpg", "text": "Where is this person?"}
]

# Load example images
example_images = [Image.open(example["image"]) for example in examples]

# Gradio
interface = gr.Interface(
    fn=predict,
    inputs=[
        gr.Image(type="pil", label="Image Input"),
        gr.Textbox(label="Text Input")
    ],
    outputs=gr.Textbox(label="Output"),
    title="Llama 3.2 11B Vision Instruct Chat",
    description="Image + text chat.",
    examples=[{"image": image, "text": example["text"]} for image, example in zip(example_images, examples)]
)

interface.launch()