app.py

import os

# Set Hugging Face and Torch cache to a guaranteed-writable location
cache_dir = "/tmp/huggingface_cache"
os.environ["HF_HOME"] = cache_dir
os.environ["TORCH_HOME"] = cache_dir
os.environ["MPLCONFIGDIR"] = "/tmp/mplcache"


# Create the directory if it doesn't exist
os.makedirs(cache_dir, exist_ok=True)

import gradio as gr
import torch
from PyPDF2 import PdfReader
from io import BytesIO
from PIL import Image
from transformers import AutoProcessor, Qwen2VLForConditionalGeneration
from olmocr.data.renderpdf import render_pdf_to_base64png
from olmocr.prompts import build_finetuning_prompt
from olmocr.prompts.anchor import get_anchor_text
from ebooklib import epub
import base64
import tempfile



# Load model
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = Qwen2VLForConditionalGeneration.from_pretrained(
    "allenai/olmOCR-7B-0225-preview", torch_dtype=torch.bfloat16 if torch.cuda.is_available() else torch.float32
).eval().to(device)
processor = AutoProcessor.from_pretrained("Qwen/Qwen2-VL-7B-Instruct")


def ocr_page(pdf_path, page_num):
    # Render page to base64 PNG
    image_b64 = render_pdf_to_base64png(pdf_path, page_num + 1, target_longest_image_dim=1024)
    anchor_text = get_anchor_text(pdf_path, page_num + 1, pdf_engine="pdfreport", target_length=4000)
    prompt = build_finetuning_prompt(anchor_text)

    messages = [
        {
            "role": "user",
            "content": [
                {"type": "text", "text": prompt},
                {"type": "image_url", "image_url": {"url": f"data:image/png;base64,{image_b64}"}},
            ],
        }
    ]

    prompt_text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    main_image = Image.open(BytesIO(base64.b64decode(image_b64)))
    inputs = processor(text=[prompt_text], images=[main_image], return_tensors="pt", padding=True)
    inputs = {k: v.to(device) for k, v in inputs.items()}

    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            temperature=0.8,
            max_new_tokens=1024,
            do_sample=True,
        )

    prompt_len = inputs["input_ids"].shape[1]
    new_tokens = outputs[:, prompt_len:]
    decoded = processor.tokenizer.batch_decode(new_tokens, skip_special_tokens=True)
    return decoded[0] if decoded else ""


def convert_pdf_to_epub(pdf_file, title, author, language):
    with tempfile.NamedTemporaryFile(delete=False, suffix=".pdf") as tmp_pdf:
        tmp_pdf.write(pdf_file.read())
        tmp_pdf_path = tmp_pdf.name

    reader = PdfReader(tmp_pdf_path)
    num_pages = len(reader.pages)

    # Create EPUB book
    book = epub.EpubBook()
    book.set_title(title)
    book.add_author(author)
    book.set_language(language)

    # Use first page as cover
    cover_image_b64 = render_pdf_to_base64png(tmp_pdf_path, 1, target_longest_image_dim=1024)
    cover_image_bytes = base64.b64decode(cover_image_b64)
    book.set_cover("cover.jpg", cover_image_bytes)

    # OCR and add pages
    for i in range(num_pages):
        text = ocr_page(tmp_pdf_path, i)
        chapter = epub.EpubHtml(title=f"Page {i+1}", file_name=f"page_{i+1}.xhtml", lang=language)
        chapter.content = f"<h1>Page {i+1}</h1><p>{text}</p>"
        book.add_item(chapter)
        book.spine.append(chapter)

    # Finalize EPUB
    book.add_item(epub.EpubNcx())
    book.add_item(epub.EpubNav())
    epub_path = os.path.join(tempfile.gettempdir(), "output.epub")
    epub.write_epub(epub_path, book, {})

    with open(epub_path, "rb") as f:
        return epub_path, f.read()


def interface_fn(pdf, title, author, language):
    epub_path, epub_bytes = convert_pdf_to_epub(pdf, title, author, language)
    return epub_path


demo = gr.Interface(
    fn=interface_fn,
    inputs=[
        gr.File(label="Upload PDF", file_types=[".pdf"]),
        gr.Textbox(label="EPUB Title", placeholder="e.g. Understanding AI"),
        gr.Textbox(label="Author", placeholder="e.g. Allen AI"),
        gr.Textbox(label="Language", placeholder="e.g. en", value="en"),
    ],
    outputs=gr.File(label="Download EPUB"),
    title="PDF to EPUB Converter (olmOCR)",
    description="Upload a PDF to convert it into a structured EPUB. The first page is used as the cover. OCR is performed with the olmOCR model.",
    allow_flagging="never",
)

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