stablediffusion-webui-rocm-docker/README.md

# Stable Diffusion WebUI Docker with AMD ROCm Support

This Docker setup provides a containerized version of Stable Diffusion WebUI with full AMD ROCm GPU acceleration support.

## Features

- **AMD ROCm 5.7** support for GPU acceleration
- **gfx1100 (ROCm 11.0.0)** GPU target configuration
- **PyTorch 2.0** with ROCm compatibility
- **Docker Compose** for easy deployment
- **Persistent storage** for models, outputs, and configurations
- **Optimized** Docker build with multi-stage construction

## Compatibility Notes

This configuration uses older versions for compatibility with the upstream Stable Diffusion WebUI project:
- **ROCm 5.7**
- **Ubuntu 22.04**
- **PyTorch 2.0**

**GPU Compatibility**: This setup may not work with GPUs newer than the Radeon RX 7900 XTX. For newer GPUs, users may wish to try updating to newer ROCm versions (7.11+) and Ubuntu versions (24.04+).

For newer hardware compatibility, consider:
1. Updating ROCm to version 7.11 or later
2. Using Ubuntu 24.04 LTS instead of 22.04
3. Updating PyTorch to a newer version compatible with newer ROCm versions

This approach prioritizes stability and compatibility with the upstream project over newer hardware support.

## Requirements

- Docker Engine (20.10+ recommended)
- Docker Compose (v2+ recommended)
- AMD GPU with ROCm support (RDNA2/RDNA3 recommended)
- Linux host system (for proper GPU passthrough)

## Quick Start

### 1. Build the Docker image

```bash
docker-compose build
```

### 2. Start the container

```bash
docker-compose up -d
```

### 3. Access the Web UI

Open your browser to: `http://localhost:7860`

## Configuration

### Environment Variables

You can customize the setup using environment variables in the `docker-compose.yml` file:

- `HSA_OVERRIDE_GFX_VERSION`: GPU target version (default: `11.0.0` for gfx1100)
- `HIP_VISIBLE_DEVICES`: Which GPU devices to use (default: `0`)
- `PYTORCH_ROCM_ARCH`: PyTorch ROCm architecture (default: `gfx1100`)
- `CLI_ARGS`: Additional command line arguments for launch.py

### Volumes

The following directories are persisted as Docker volumes:

- `./models` - Model files
- `./outputs` - Generated images and outputs
- `./extensions` - WebUI extensions
- `./config.json` - Configuration file

### GPU Configuration

The Docker setup automatically configures:

- `/dev/kfd` - AMD KFD device for GPU acceleration
- `/dev/dri` - Direct Rendering Infrastructure for display
- Proper group permissions (video, render)

## Customization

### Change GPU Target

To use a different GPU target (e.g., gfx1030 for Navi 2x):

1. Edit `docker/docker-compose.yml`
2. Change `HSA_OVERRIDE_GFX_VERSION` to your target version
3. Change `PYTORCH_ROCM_ARCH` accordingly

### Add Custom Arguments

Modify the `CLI_ARGS` environment variable in `docker-compose.yml` to add custom launch arguments:

```yaml
environment:
  - CLI_ARGS=--listen --port 7860 --xformers --opt-sdp-attention --your-custom-arg
```

## Building for Different ROCm Versions

To use a different ROCm version:

1. Edit `docker/Dockerfile`
2. Change the `ROCM_VERSION` environment variable
3. Update the PyTorch installation command to match the ROCm version
4. Rebuild the container

## Troubleshooting

### GPU Not Detected

- Ensure your host system has ROCm properly installed
- Check that the `video` and `render` groups exist on your host
- Verify that `/dev/kfd` and `/dev/dri` devices exist

### Build Failures

- Ensure you have enough disk space (ROCm installation requires ~5GB)
- Check your internet connection for downloading ROCm packages
- Try cleaning the Docker build cache: `docker system prune -a`

### Performance Issues

- Make sure your GPU is properly supported by ROCm 5.7
- Check that the correct `HSA_OVERRIDE_GFX_VERSION` is set for your GPU
- Monitor GPU usage with `rocm-smi` on the host system

## Updating

To update to the latest version:

```bash
docker-compose down
git pull
docker-compose build --no-cache
docker-compose up -d
```

## License

This Docker configuration is provided under the same license as the Stable Diffusion WebUI project (AGPL-3.0).