app.py

import gradio as gr
import torch
import pypdfium2
from PIL import Image
from transformers import Qwen2VLProcessor, Qwen2VLImageProcessor, AutoTokenizer, Qwen2VLModel

# Load model and processor
model_name = "Qwen/Qwen-VL"  # You may replace with your preferred VL model
image_processor = Qwen2VLImageProcessor.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)
processor = Qwen2VLProcessor(image_processor=image_processor, tokenizer=tokenizer)
model = Qwen2VLModel.from_pretrained(
    model_name,
    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32
)
model.eval()

# Convert PDF to list of PIL images (one per page)
def pdf_to_images(pdf_path):
    pdf = pypdfium2.PdfDocument(pdf_path)
    return [page.render().to_pil() for page in pdf]

# Generate text from each image using the vision-language model
def process_pdf(pdf_file):
    images = pdf_to_images(pdf_file.name)
    results = []

    for image in images:
        inputs = processor(images=image, return_tensors="pt").to(model.device)
        with torch.no_grad():
            outputs = model.generate(**inputs, max_new_tokens=256)
        text = processor.batch_decode(outputs, skip_special_tokens=True)[0]
        results.append(text.strip())

    return "\n\n".join(results)

# Gradio UI
demo = gr.Interface(
    fn=process_pdf,
    inputs=gr.File(type="file", file_types=[".pdf"]),
    outputs="text",
    title="olmOCR PDF Processor"
)

if __name__ == "__main__":
    demo.launch()