FastHunyuan RunPod Worker

This is a RunPod worker implementation for FastHunyuan, a fast inference version of the Hunyuan text-to-video model.

Development Guide

Prerequisites

• RunPod account with API access
• NVIDIA GPU with CUDA support
• (Optional) VSCode for remote development

Development Setup

1. Create a development pod on RunPod:

   • Go to RunPod Console
   • Click "Deploy"
   • Select base image: runpod/pytorch:2.1.2-py3.10-cuda12.1.0 (includes SSH and JupyterLab)
   • Choose your GPU type
   • Name your pod (e.g., "fasthunyuan-dev")
   • Deploy

2. Connect to your pod (choose one method):

   a. Using RunPod Web Terminal:

   • Click "Connect" on your pod in RunPod console
   • Select "SSH Terminal" or use the web terminal

   b. Using VSCode (Recommended for development):

   • Follow the official RunPod guide for VSCode setup
   • This allows you to develop directly on the pod with full IDE support

3. Install dependencies:

   cd /workspace
   chmod +x builder/install.sh
   ./builder/install.sh

4. Test the handler:

   python src/handler.py

Development Tips

• The handler is located in src/handler.py
• Dependencies are managed in builder/requirements.txt
• System dependencies and installation steps are in builder/install.sh
• Using VSCode remote development provides a better development experience with:
  • Integrated terminal
  • Code completion
  • Debugging capabilities
  • Git integration

Production Guide

Building the Docker Image

1. Build the image:

   docker build -t your-registry/fasthunyuan-worker:version .

   Note: For production, we use a minimal base image (nvidia/cuda:12.1.0-cudnn8-devel-ubuntu22.04) without development tools.

2. Push to your registry:

   docker push your-registry/fasthunyuan-worker:version

Deploying on RunPod

1. Create a template on RunPod:

   • Base image: your-registry/fasthunyuan-worker:version
   • Container disk: at least 10GB
   • HTTP endpoints: enabled
   • Environment variables:

     HUGGING_FACE_HUB_TOKEN=your_token
     

2. Deploy serverless endpoints using the template

API Usage

Example request:

{
  "input": {
    "prompt": "A cinematic video of a beautiful landscape",
    "height": 720,
    "width": 1280,
    "num_frames": 45,
    "num_inference_steps": 6,
    "seed": 1024,
    "fps": 24
  }
}

Example response:

{
  "video_path": "/tmp/output_1024.mp4"
}

Project Structure

.
├── builder/
│   ├── install.sh        # Development installation script
│   └── requirements.txt  # Python dependencies
├── src/
│   └── handler.py        # RunPod handler implementation
├── Dockerfile           # Production container definition
└── README.md           # This file

License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.