SuryaOCR RunPod H100 Serverless

Production-ready SuryaOCR deployment on RunPod H100 GPU with 0.5s per image OCR speed.

🚀 Quick Setup (2 Options)

Option 1: Pre-built Docker Image (RECOMMENDED - Instant Startup)

Best for production: Models are pre-downloaded, no cold start delays!

The Docker image is automatically built by GitHub Actions and pushed to GitHub Container Registry.

Just use this image in RunPod:

ghcr.io/gunitbindal/surya-runpod-h100:latest

Setup:

1. Create RunPod Template:

   • Go to: RunPod Console → Serverless → Templates → New Template
   • Container Image: ghcr.io/gunitbindal/surya-runpod-h100:latest
   • Docker Command: (leave empty)
   • GPU: H100 80GB or H100 PCIe

2. Deploy Endpoint:

   • Template: Select your template
   • Active Workers: 1
   • Max Workers: 1
   • Endpoint Type: Queue

Benefits:

• ✅ Instant cold starts (2-3 seconds)
• ✅ No model download wait time
• ✅ No pip installs on startup
• ✅ Production-ready
• ✅ Automatically updated on every push to main

Option 2: GitHub Handler (Simple but Slower)

Good for testing: Downloads on every worker start.

1. Create Template:

   • Container Image: runpod/pytorch:2.8.0-py3.11-cuda12.8.1-cudnn-devel-ubuntu22.04
   • Docker Command:

   bash -c "pip install --no-cache-dir surya-ocr runpod pillow && curl -sSL https://raw.githubusercontent.com/GunitBindal/surya-runpod-h100/main/handler_final.py -o handler.py && python -u handler.py"

2. Deploy Endpoint (same as Option 1, step 3)

Drawbacks:

• ⏱️ First request: 60-90 seconds (downloads models)
• 📦 Installs packages every worker start
• 💰 Wastes compute time on setup

📝 Usage

📋 Key Points

• First request: 60-90 seconds (downloads Surya models ~500MB)
• Subsequent requests: ~0.5 seconds per image
• Active Workers = 1: Keeps worker warm, prevents queue issues
• Logs visible: Handler prints status to worker logs

📁 Files

• handler_final.py - Optimized handler with logging
• docker_command.txt - RunPod Docker command
• test_client.py - Python test client

🔧 Troubleshooting

Requests stuck IN_QUEUE?

• Set Active Workers = 1 in endpoint settings
• Wait 90 seconds for first model download
• Check worker logs for errors

Worker crashing?

• Verify H100 GPU is selected
• Ensure Docker command is exactly as shown above
• Check logs for memory issues

Current Status

✓ Handler updated with better logging
✓ Simplified deployment (no Docker build)
✓ Single-command setup via GitHub