Qwen3-2B-VL OCR Deployment

This project provides a Dockerized setup to run Qwen3-2B-VL with vLLM @ FP8 (testing on a 3060 12GB) and a Streamlit frontend for performing OCR on PDF documents.

---

Features

• Upload PDF files via Streamlit UI
• Rasterizes PDF pages to images
• Sends images to Qwen3-2B-VL for text extraction
• View extracted text per page in the browser
• Download combined OCR output as a .txt file

---

Requirements

• Docker & Docker Compose installed
• GPU with CUDA support (optional but recommended)
• Internet connection to download the model from Hugging Face
• vLLM API running (inside Docker)

---

Project Structure

vllmocrexp/
├── docker-compose.yml
├── streamlit/
│   ├── app.py
│   ├── requirements.txt
│   └── Dockerfile
├── models/          # optional, local model storage
└── tmp/             # temporary PDFs/images (ignored in git)

---

Setup & Deployment

1. Clone the repo

git clone https://github.com/ikantkode/qwen3-2b-ocr-app
cd qwen3-2b-ocr-app

2. Build and start the containers

docker compose build --no-cache && docker compose up
OR
docker compose build --no-cache && docker compose up -d

• qwen-vlm container runs the Qwen3-2B-VL model with vLLM
• streamlit-ui container runs the frontend at: http://localhost:8501

3. Upload a PDF in the Streamlit UI

• Each page will be displayed as an image
• OCR text is extracted for each page
• Full text can be downloaded as ocr_output.txt

---

Clean-up

To remove temporary files, Docker volumes, and images:

docker compose down -v
rm -rf tmp/

---

License

MIT License.