Skip to content

deployment.md

Latest commit

 

History

History
48 lines (34 loc) · 2.19 KB

deployment.md

File metadata and controls

48 lines (34 loc) · 2.19 KB

Deploying Stable Diffusion Server

This document describes how to build Docker images using GitHub Actions and deploy them to RunPod and Google Cloud Run with GPU support.

GitHub Actions workflow

The repository includes a workflow located at .github/workflows/docker-build.yml. This workflow builds two Docker images whenever changes are pushed to main or a pull request is opened. It installs dependencies with uv and caches the build layers for faster rebuilds:

  • runpod:latest – image intended for RunPod
  • cloudrun:latest – image intended for Google Cloud Run

Both images are published to the GitHub Container Registry under ghcr.io/<OWNER>/<REPO>.

The provided Dockerfile installs packages via uv pip and uses BuildKit cache mounts to speed up dependency installation.

Running on RunPod

  1. Ensure you have a RunPod account and have created a container workspace with GPU access.
  2. In RunPod, configure your deployment to pull the runpod:latest image from GHCR.
  3. Expose port 8000. The server will start automatically with the default command from the Dockerfile.
  4. Optionally mount any model or data volumes required by the application.

Running on Google Cloud Run with L4 GPUs

  1. Enable the Cloud Run API and create a new service with GPU support. L4 GPUs are available in us-central1 or other supported regions.
  2. Grant Cloud Run permission to access Artifact Registry or GHCR where the image is stored.
  3. Deploy using the cloudrun:latest image: 
    gcloud run deploy sd-server \
    --image=ghcr.io/<OWNER>/<REPO>/cloudrun:latest \
    --region=us-central1 \
    --gpu=1 \
    --gpu-type=nvidia-l4 \
    --memory=8Gi \
    --min-instances=0 \
    --max-instances=1
  4. Make sure to allocate sufficient memory and enable GPUs for the service.

Local testing

To test the Docker image locally with GPU support, install the NVIDIA Container Toolkit and run:

docker run --gpus all -p 8000:8000 ghcr.io/<OWNER>/<REPO>/runpod:latest

The API will be available at http://localhost:8000.