SeedVR RunPod Container

A lightweight Docker container for deploying SeedVR video restoration models on RunPod. Supports both SeedVR2-3B and SeedVR2-7B models with a user-friendly Gradio interface.

🚀 Recent Update: Now uses PyTorch 2.7.1 + CUDA 12.6 base image for improved L40 GPU compatibility and reduced complexity. See README_pytorch_approach.md for technical details.

🚀 Features

• Dual Model Support: Choose between SeedVR2-3B (faster) or SeedVR2-7B (higher quality)
• Lightweight Container: Downloads models and dependencies on the remote host
• Gradio Interface: Easy-to-use web interface for video restoration
• RunPod Optimized: Designed specifically for RunPod deployment
• GPU Accelerated: Supports multi-GPU inference with sequence parallelism
• Robust Setup: Enhanced error checking, retry logic, and CUDA verification
• Smart Caching: Caches wheels and models to speed up restarts
• Comprehensive Monitoring: Real-time GPU verification and tensor operation testing

📋 Requirements

Hardware

• GPU: NVIDIA GPU with at least 24GB VRAM (for SeedVR2-3B)
• GPU: NVIDIA GPU with at least 40GB VRAM (for SeedVR2-7B)
• Storage: At least 50GB free space for models and processing
• CUDA: Compatible with CUDA 12.1+

Software

• Docker with NVIDIA GPU support
• NVIDIA Docker runtime
• Python 3.10 (PyTorch base image)
• PyTorch 2.7.1 with CUDA 12.6 support (PyTorch official base)
• Flash-attention 2.8.3 for L40 GPU compatibility
• Minimal dependencies - selective installation vs full requirements.txt

🐳 Docker Hub

The pre-built image is available on Docker Hub:

docker pull gemneye/seedvr-runpod:latest

🔧 Environment Variables

Variable	Default	Description
GRADIO_SHARE	false	Enable Gradio sharing (true/false)
GRADIO_PORT	7860	Port for Gradio interface
GRADIO_HOST	0.0.0.0	Host binding for Gradio
HF_TOKEN	-	Hugging Face token (if needed)

🚀 Quick Start

RunPod Deployment

1. Create a new RunPod instance:

   • Choose a GPU pod with sufficient VRAM
   • Use the custom container option
   • Container image: gemneye/seedvr-runpod:latest
   • Expose port: 7860

2. Set environment variables (optional):

   GRADIO_SHARE=true
   HF_TOKEN=your_token_here

3. Start the container: The container will automatically:

   • Download required models (SeedVR2-3B and SeedVR2-7B)
   • Install flash-attention and apex
   • Launch the Gradio interface on port 7860

4. Access the interface:

   • Open the provided URL in your browser
   • Upload a video and select your preferred model
   • Adjust output resolution and settings
   • Click "Restore Video" to process

Local Development

# Clone the repository
git clone https://github.com/gemneye/SeedVR-RunPod.git
cd SeedVR-RunPod

# Build the Docker image
docker build -t seedvr-runpod .

# Run with GPU support
docker run --gpus all -p 7860:7860 -e GRADIO_SHARE=true seedvr-runpod

📝 Usage Guide

Model Selection

• SeedVR2-3B: Faster processing, good quality, requires ~24GB VRAM
• SeedVR2-7B: Higher quality, slower processing, requires ~40GB VRAM

Input Recommendations

• Format: MP4 videos work best
• Length: Shorter videos (< 30 seconds) process faster
• Resolution: Input resolution affects processing time
• Quality: Best results on videos with compression artifacts or low quality

Output Settings

• Width/Height: Adjust output resolution (512-2048px)
• Seed: Fixed seed for reproducible results
• Processing Time: Varies based on video length and resolution

🛡️ Enhanced Robustness

Based on proven patterns from production deployments, this container includes:

CUDA & GPU Verification

• Real Tensor Operations: Tests actual GPU computations, not just availability
• Memory Detection: Reports GPU memory and capabilities
• Environment Setup: Configures RANK, LOCAL_RANK, WORLD_SIZE for proper inference
• Comprehensive Logging: Detailed status reporting throughout setup

Retry Logic & Error Handling

• Exponential Backoff: Smart retry timing for downloads and installations
• Wheel Caching: Pre-downloads and caches flash_attn and apex wheels
• Resume Downloads: Models resume from interruption points
• Graceful Degradation: Continues setup even if optional components fail

Installation Verification

• Dependency Checking: Verifies each component before proceeding
• Version Validation: Ensures compatible PyTorch/CUDA versions
• Setup Completion: Marks successful setup to avoid re-running
• Storage Monitoring: Reports available space during downloads

Version Management

• Fixed Dependencies: Uses tested, compatible package versions
• CUDA Compatibility: PyTorch 2.4.0 + CUDA 12.1 for optimal performance
• Dependabot: Configured for security patches only (no major version bumps)
• Environment Stability: Python 3.10 ensures compatibility with all components

🔧 Manual Setup

If you need to set up the environment manually:

# Install flash-attention (with retry logic)
pip install https://github.com/Dao-AILab/flash-attention/releases/download/v2.5.9.post1/flash_attn-2.5.9.post1+cu122torch2.4cxx11abiFALSE-cp310-cp310-linux_x86_64.whl

# Install apex (with caching)
pip install https://huggingface.co/ByteDance-Seed/SeedVR2-3B/resolve/main/apex-0.1-cp310-cp310-linux_x86_64.whl

# Download models (with resume capability)
python download_models.py --model both

📊 Performance Notes

GPU Memory Requirements

• SeedVR2-3B: ~24GB VRAM for 720p videos
• SeedVR2-7B: ~40GB VRAM for 720p videos
• Multi-GPU: Sequence parallelism reduces per-GPU memory usage

Processing Times (Approximate)

• 10-second 720p video: 2-5 minutes (3B), 5-10 minutes (7B)
• 30-second 1080p video: 10-15 minutes (3B), 20-30 minutes (7B)

🛠️ Development

Building the Image

docker build -t seedvr-runpod:dev .

GitHub Actions

The repository includes automated building and pushing to Docker Hub:

• Triggers on push to main/develop branches
• Creates tagged releases
• Supports manual workflow dispatch

Required Secrets

Set these in your GitHub repository:

• DOCKERHUB_USERNAME: Your Docker Hub username
• DOCKERHUB_TOKEN: Your Docker Hub access token

📄 License

This project follows the original SeedVR licensing:

• SeedVR: Apache 2.0 License
• Container Code: MIT License

🤝 Contributing

1. Fork the repository
2. Create a feature branch
3. Make your changes
4. Test thoroughly
5. Submit a pull request

📞 Support

• Issues: Report bugs via GitHub Issues
• Discussions: Use GitHub Discussions for questions
• Original SeedVR: Visit the original repository

🙏 Acknowledgments

• ByteDance: For the original SeedVR models and research
• RunPod: For the cloud GPU platform
• Gradio: For the web interface framework

---

Note: This is an unofficial containerization of SeedVR for RunPod deployment. Please refer to the original SeedVR repository for the latest model updates and research.