app.py

import gradio as gr
from transformers import (
    PaliGemmaProcessor,
    PaliGemmaForConditionalGeneration,
)
from transformers.image_utils import load_image
import torch
import os
import spaces  # Import the spaces module
import requests
from io import BytesIO
from PIL import Image


def load_model():
    """Load PaliGemma2 model and processor with Hugging Face token."""

    token = os.getenv("HUGGINGFACEHUB_API_TOKEN")  # Retrieve token from environment variable

    if not token:
        raise ValueError(
            "Hugging Face API token not found. Please set it in the environment variables."
        )

    # Load the processor and model using the correct identifier
    model_id = "google/paligemma2-10b-pt-448"
    processor = PaliGemmaProcessor.from_pretrained(model_id, use_auth_token=token)
    device = "cuda" if torch.cuda.is_available() else "cpu"
    model = PaliGemmaForConditionalGeneration.from_pretrained(
        model_id, torch_dtype=torch.bfloat16, use_auth_token=token
    ).to(device).eval()

    return processor, model


@spaces.GPU(duration=120)  # Increased timeout to 120 seconds
def process_image_and_text(image_pil, text_input, num_beams, temperature, seed):
    """Extract text from image using PaliGemma2."""
    try:
        processor, model = load_model()
        device = "cuda" if torch.cuda.is_available() else "cpu"

        # Load the image using load_image
        image = load_image(image_pil)

        # Use the provided text input
        model_inputs = processor(images=image, return_tensors="pt").to(
            device, dtype=torch.bfloat16
        )
        input_len = model_inputs["input_ids"].shape[-1]
        
        torch.manual_seed(seed) # Set random seed for reproducibility

        with torch.inference_mode():
            generation = model.generate(**model_inputs, max_new_tokens=500, do_sample=True, num_beams=num_beams, temperature=temperature)
            generation = generation[0][input_len:]
            decoded = processor.decode(generation, skip_special_tokens=True)

        return decoded
    except Exception as e:
        print(f"Error during GPU task: {e}")
        raise gr.Error(f"GPU task failed: {e}")


if __name__ == "__main__":
    iface = gr.Interface(
        fn=process_image_and_text,
        inputs=[
            gr.Image(type="pil", label="Upload an image"),
            gr.Slider(minimum=1, maximum=10, step=1, value=1, label="Number of Beams"),
            gr.Slider(minimum=0.1, maximum=2.0, step=0.1, value=1.0, label="Temperature"),
            gr.Number(label="Random Seed", value=0, precision=0),
        ],
        outputs=gr.Textbox(label="Generated Text"),
        title="PaliGemma2 Image and Text to Text",
        description="Upload an image and enter a text prompt. The model will generate text based on both.",
    )
    iface.launch()