lcj-group/aitherhub-deploy

Aitherhub Deploy

Repository

• Repo: LCJ-Group/aitherhub
• Clone: gh repo clone LCJ-Group/aitherhub
• Local path: /home/ubuntu/aitherhub-repo
• Branch: master (single branch, push to deploy)

Architecture

| Component | Technology | Hosting | Deploy Trigger | |---|---|---|---| | Frontend | Vite + React | Azure Static Web Apps | Push to master | | Backend (main) | FastAPI | Azure App Service (aitherhubAPI) | Push to master | | GPU Worker | RunPod Serverless (FastAPI) | RunPod Serverless (L4/A5000) | Push to master (gpu-worker/serverless/) |

Deployment Workflows (GitHub Actions)

All workflows trigger on push to master branch automatically.

1. Frontend (deploy-swa-frontend.yml)

• Builds: cd frontend && npm install && npm run build
• Deploys to: Azure Static Web Apps
• Typical time: ~2 minutes

2. Backend (master_aitherhubapi.yml)

• Builds: Python app with requirements.txt
• Deploys to: Azure App Service aitherhubAPI
• Typical time: ~5-10 minutes
• Note: After deploy completes, Azure App Service may take 1-2 extra minutes to restart.

3. GPU Worker (build-gpu-worker.yml)

• Builds: Multi-stage Dockerfile (gpu-worker/serverless/Dockerfile)
• Deploys to: GitHub Container Registry (ghcr.io/lcj-group/aitherhub-gpu-worker:latest)
• Trigger: Push to gpu-worker/serverless/ path in master branch.
• Typical time: ~20-30 minutes

GPU Worker (RunPod Serverless)

Overview

The GPU Worker has been migrated from a persistent RunPod Pod to a RunPod Serverless endpoint. This provides auto-scaling, eliminates the need for manual deployments, and ensures high availability without managing a specific Pod.

The core of this architecture is a self-contained Docker image that includes all necessary code (FaceFusion, MuseTalk), dependencies, and essential models. This removes the dependency on a persistent /workspace/ volume.

Endpoint Details

| Setting | Value | |---|---| | Endpoint ID | 2noptqoq7n8f8g | | Docker Image | ghcr.io/lcj-group/aitherhub-gpu-worker:latest | | GPUs | L4, RTX A5000, etc. (auto-scales) | | Cold Start Time | ~1-2 minutes |

Deployment Process

Deployment is now fully automated:

1. Push Changes: Any changes pushed to the gpu-worker/serverless/ directory in the master branch will automatically trigger the build-gpu-worker.yml GitHub Actions workflow.
2. Build Docker Image: The workflow builds a new multi-stage Docker image, including all code and dependencies. This takes approximately 20-30 minutes.
3. Push to GHCR: The newly built image is pushed to the GitHub Container Registry.
4. RunPod Auto-Update: The RunPod Serverless endpoint is configured to watch the latest tag of the Docker image. When a new image is pushed, RunPod automatically initiates a rolling update, replacing old workers with new ones without downtime.

There are no manual steps required to deploy the GPU worker anymore.

Backend ↔ GPU Worker Connection

• The backend (aitherhubAPI) now uses the RunPodServerlessService to send jobs to the Serverless endpoint.
• The service uses the RunPod API to submit jobs (/run or /runsync) and check their status (/status).
• Configuration: The RUNPOD_API_KEY and RUNPOD_ENDPOINT_ID are configured as fallback values directly in runpod_serverless_service.py. This was done to avoid complexities with Azure App Service environment variable management.

Health & Verification

• Backend Health: The backend's health can be checked to verify its connection to the Serverless endpoint.
  • URL: https://aitherhubapi-cpcjcnezbgf5f7e2.japaneast-01.azurewebsites.net/api/v1/digital-human/musetalk/health
  • Header: X-Admin-Key: aither:hub
  • A cold_start status is normal if the worker is idle.
• Direct Endpoint Health: The Serverless endpoint itself can be checked directly.
  • URL: https://api.runpod.ai/v2/2noptqoq7n8f8g/runsync
  • Method: POST
  • Body: {"input": {"action": "health"}}
  • Authorization: Bearer <RUNPOD_API_KEY>

API Base URLs

• Main API: https://aitherhubapi-cpcjcnezbgf5f7e2.japaneast-01.azurewebsites.net
• Frontend: https://www.aitherhub.com
• GPU Worker API: https://api.runpod.ai/v2/2noptqoq7n8f8g