README_RUNPOD.md

MinerU RunPod Serverless Deployment

Overview

This deployment includes MinerU models directly in the Docker image for fast cold starts on RunPod Serverless.

Build and Deploy

1. Build Docker Image

./build_runpod.sh

This will:

• Build the Docker image with all MinerU models included
• Download models during build (this takes ~10-15 minutes)
• Result in a Docker image of approximately 5-10GB

2. Push to Docker Hub

docker login
docker push marcosremar2/mineru-runpod:latest

3. Deploy on RunPod

1. Go to RunPod Serverless
2. Click "New Template"
3. Configure:
   • Container Image: marcosremar2/mineru-runpod:latest
   • Container Disk: 20 GB (to be safe)
   • Volume Size: 0 GB (not needed, models in image)
   • GPU: Any GPU with 8GB+ VRAM
   • Max Workers: Based on your needs
   • Idle Timeout: 5 seconds
   • Execution Timeout: 120 seconds

4. Test the Deployment

python test_runpod.py test.pdf https://api.runpod.ai/v2/YOUR_ENDPOINT_ID YOUR_API_KEY

API Usage

Request Format

{
  "input": {
    "pdf_base64": "base64_encoded_pdf_content",
    "filename": "document.pdf"
  }
}

Response Format

{
  "output": {
    "markdown": "# Converted Document\n\nContent here...",
    "filename": "document.pdf",
    "status": "success",
    "pages": 5
  }
}

Cost Estimation

• Cold Start: ~5-10 seconds (models already in image)
• Processing: ~10-30 seconds per PDF
• GPU Cost: ~$0.00024/second
• Total per PDF: ~$0.01-0.02

Optimization Tips

1. Reduce Image Size: Remove unnecessary models from Dockerfile
2. Use Active Workers: For consistent load, keep 1-2 active workers
3. Adjust Timeout: Increase for larger PDFs
4. Monitor Usage: Use RunPod dashboard to track costs

Troubleshooting

1. Out of Memory: Use larger GPU (16GB+ VRAM)
2. Timeout: Increase execution timeout in template
3. Model Loading: Check MINERU_MODEL_PATH environment variable