DeepSeek-OCR-2 API Service

This repository provides an API service and Docker support for DeepSeek-OCR-2, utilizing the V2 model with "Visual Causal Flow". This is a port of the API structure from dpsk-ocr adapted for the newer model architecture.

Features

• DeepSeek-OCR-2 Model: Uses the latest V2 architecture for improved OCR performance and layout analysis.
• FastAPI Backend: Efficient, async-capable API for PDF and image processing.
• Dockerized: Ready-to-deploy Docker image with CUDA 11.8, PyTorch 2.6, and vLLM 0.8.5 optimization.
• Task Queue: Background processing for handling large PDF files without blocking the API.
• Layout Detection: Returns structured markdown, detection coordinates, and layout visualizations.

Prerequisites

• NVIDIA GPU (Ampere or newer recommended, e.g., A100, A10, RTX 3090/4090)
• roughly 16GB+ VRAM for the model.
• Linux with CUDA drivers installed.

Installation

Option 1: Docker (Recommended)

1. Build the image:

   docker build -t deepseek-ocr2 .

2. Run the container:

   docker run --gpus all -p 8000:8000 -v $(pwd)/data:/app/data deepseek-ocr2

Option 2: Local Conda Environment

1. Create Environment:

   conda create -n deepseek-ocr2 python=3.12 -y
   conda activate deepseek-ocr2

2. Install GPU Dependencies:

   pip install torch==2.6.0 torchvision==0.21.0 torchaudio==2.6.0 --index-url https://download.pytorch.org/whl/cu118
   
   # Install vLLM (Compatible wheel for Python 3.12/CUDA 11.8)
   wget https://github.com/vllm-project/vllm/releases/download/v0.8.5/vllm-0.8.5+cu118-cp38-abi3-manylinux1_x86_64.whl
   pip install vllm-0.8.5+cu118-cp38-abi3-manylinux1_x86_64.whl
   
   pip install flash-attn==2.7.3 --no-build-isolation

3. Install Python Requirements:

   pip install -r requirements.txt

4. Run the API:

   python serve_pdf.py

Usage

API Endpoints

The API runs at http://localhost:8000. Full interactive documentation is available at http://localhost:8000/docs.

1. Process a PDF

POST /process_pdf

• Upload a PDF file to start processing.
• Returns a job_id.

2. Check Status

GET /result/{job_id}/status

• Check if the job is pending, processing, completed, or failed.

3. Get Results

• Markdown: GET /result/{job_id}/markdown
• Markdown with Coordinates: GET /result/{job_id}/markdown_det
• Layout Visualization (PDF): GET /result/{job_id}/layout_pdf
• Extracted Images: GET /result/{job_id}/images

CLI Scripts

You can also run the original scripts for direct processing:

• Single Image: python run_dpsk_ocr2_image.py
• PDF: python run_dpsk_ocr2_pdf.py
• Batch Eval: python run_dpsk_ocr2_eval_batch.py

Note: Configure config.py to set input/output paths for CLI scripts.

Project Structure

• serve_pdf.py: Main API application.
• DeepSeek-OCR2-vllm/: Core model code.
• config.py: Configuration settings.
• database.py & task_queue.py: Job management.
• pdf_utils.py & processing_utils.py: Helper functions for image/PDF handling.

License

This project follows the license of the original DeepSeek-OCR-2 repository.