corundex/comfyui-rocm repository overview

⁠ComfyUI ROCm Docker Image

🔥 ComfyUI with AMD ROCm support - Run ComfyUI on AMD GPUs with optimized ROCm-compatible dependencies.

corundex/comfyui-rocm GitHub repository⁠

ComfyUI running on AMD ROCm with sample workflow and generated landscape image

⁠📋 Version Information

• Base Image: rocm/pytorch:rocm6.4.1_ubuntu24.04_py3.12_pytorch_release_2.6.0
• Python: 3.12.10
• PyTorch: 2.6.0+git684f6f2
• ROCm: 6.4.43483-a187df25c
• ComfyUI: v0.3.43 (e18f53c, 2025-06-27)

⁠📦 What's Different from Standard ComfyUI?

This ROCm-optimized build:

• ✅ Uses ROCm PyTorch instead of CUDA PyTorch
• ✅ Pre-tested dependencies on real AMD hardware

⁠🧪 Tested Package Compatibility

All packages in this container have been tested on real AMD hardware. See requirements_rocm.txt for the complete lists:

⁠✅ ROCm-Compatible Packages (18 total):

• comfyui-frontend-package==1.23.4
• comfyui-workflow-templates==0.1.30
• comfyui-embedded-docs==0.2.3
• torchsde
• einops
• transformers>=4.37.2
• tokenizers>=0.13.3
• sentencepiece
• safetensors>=0.4.2
• pyyaml
• Pillow
• alembic
• SQLAlchemy
• kornia>=0.7.1
• spandrel
• av>=14.2.0
• pydantic~=2.0
• pydantic-settings~=2.0

⁠📦 Pre-installed in ROCm Base (9 total):

• torch
• torchvision
• numpy>=1.25.0
• aiohttp>=3.11.8
• yarl>=1.18.0
• scipy
• tqdm
• psutil
• soundfile

⁠🌙 Incompatible packages (1 total):

• torchaudio - Excluded from the image

⁠🚀 Quick Start

# Pull and run ComfyUI with ROCm support
docker run -d \
  --device=/dev/kfd \
  --device=/dev/dri \
  --group-add=video \
  -p 8188:8188 \
  -v $(pwd)/models:/workspace/ComfyUI/models \
  -v $(pwd)/output:/workspace/ComfyUI/output \
  corundex/comfyui-rocm:latest

Access ComfyUI at: http://localhost:8188⁠

⁠🎯 What's Included

• ComfyUI - Latest stable version
• ROCm-optimized PyTorch - AMD GPU acceleration
• Filtered dependencies - Only ROCm-compatible packages
• Ready-to-use workflow - Sample workflow included
• Persistent storage - Models and outputs preserved

⁠📋 Requirements

⁠Hardware

• AMD GPU with ROCm support (RX 6000+, RX 7000+, etc.)
• 8GB+ VRAM (16GB+ recommended for larger models)

⁠Software

• Docker with GPU support
• ROCm drivers 6.4+ installed on host
• Linux (Ubuntu 24.04+ recommended)

⁠🔧 Installation & Setup

⁠1. Install ROCm Drivers

# Ubuntu/Debian
curl -fsSL https://repo.radeon.com/rocm/rocm.gpg.key | sudo gpg --dearmor -o /etc/apt/keyrings/rocm.gpg
echo "deb [arch=amd64 signed-by=/etc/apt/keyrings/rocm.gpg] https://repo.radeon.com/rocm/apt/6.4 jammy main" | sudo tee /etc/apt/sources.list.d/rocm.list
sudo apt update && sudo apt install rocm-dkms
sudo usermod -a -G render,video $USER

⁠2. Verify ROCm Installation

rocm-smi
# Should show your AMD GPU(s)

⁠3. Run ComfyUI Container

docker run -d \
  --name comfyui-rocm \
  --device=/dev/kfd \
  --device=/dev/dri \
  --group-add=video \
  -p 8188:8188 \
  -v ./models:/workspace/ComfyUI/models \
  -v ./output:/workspace/ComfyUI/output \
  -v ./input:/workspace/ComfyUI/input \
  -v ./custom_nodes:/workspace/ComfyUI/custom_nodes \
  corundex/comfyui-rocm:latest

⁠📁 Volume Mounts

⁠🎛️ Smart Model Management

The ComfyUI Docker image includes an intelligent model management system that automatically downloads models based on your needs.

⁠Model Download Modes

Control model downloading with the MODEL_DOWNLOAD environment variable:

⁠MODEL_DOWNLOAD=default (Default)

Downloads essential model to get started:

• Stable Diffusion 1.5 - Most popular base model (4GB)

⁠MODEL_DOWNLOAD=common

Downloads comprehensive starter set:

• Stable Diffusion 1.5 + SD 1.5 VAE
• SDXL Base 1.0 + SDXL Refiner 1.0
• ControlNet models (Canny, Depth, OpenPose)
• Real-ESRGAN upscalers (4x General + 4x Anime)
• EasyNegative embedding

⁠MODEL_DOWNLOAD=realistic

Downloads realistic photo models:

• Realistic Vision v5.1 - High-quality realistic images
• DreamShaper 8 - Versatile realistic model
• Realistic VAE

⁠MODEL_DOWNLOAD=photorealistic

Downloads SDXL-based photorealistic models:

• Juggernaut XL v9 - Top-tier SDXL photorealism
• RealVisXL V4.0 - Excellent SDXL realistic model

⁠MODEL_DOWNLOAD=artistic

Downloads creative/stylized models:

• Deliberate v2 - Artistic and painterly style

⁠MODEL_DOWNLOAD=all

Downloads everything from all model sets above (⚠️ Large download ~100GB+)

⁠MODEL_DOWNLOAD=none

Skips all downloads - uses only existing models in volumes

⁠Custom Model Sets

Add your own sections to models.yaml and use them:

MODEL_DOWNLOAD=mycustom

⁠Docker Usage Examples

⁠Basic Usage (SD 1.5 only)

docker run -d \
  --device=/dev/kfd --device=/dev/dri --group-add=video \
  -p 8188:8188 \
  -v ./models:/workspace/ComfyUI/models \
  corundex/comfyui-rocm:latest

⁠Common Model Set

docker run -d \
  --device=/dev/kfd --device=/dev/dri --group-add=video \
  -p 8188:8188 \
  -e MODEL_DOWNLOAD=common \
  -v ./models:/workspace/ComfyUI/models \
  corundex/comfyui-rocm:latest

⁠All Models (Large Download)

docker run -d \
  --device=/dev/kfd --device=/dev/dri --group-add=video \
  -p 8188:8188 \
  -e MODEL_DOWNLOAD=all \
  -v ./models:/workspace/ComfyUI/models \
  corundex/comfyui-rocm:latest

⁠No Downloads (Use Existing Models)

docker run -d \
  --device=/dev/kfd --device=/dev/dri --group-add=video \
  -p 8188:8188 \
  -e MODEL_DOWNLOAD=none \
  -v ./models:/workspace/ComfyUI/models \
  corundex/comfyui-rocm:latest

⁠Smart Features

⁠Automatic Skip Existing

• Never re-downloads existing models
• Validates file sizes - redownloads corrupted files
• Resume support - continues interrupted downloads

⁠Model Inventory

On startup, see what's available:

[ComfyUI] Current model inventory:
  checkpoints: 3 models
  vae: 1 models
  loras: 2 models
  upscale_models: 1 models
  controlnet: 3 models
  embeddings: 1 models

⁠Customizing Models

⁠Edit models.yaml

Add your own model sections in YAML format:

models:
  mycustom:
    - name: "My Custom Model"
      url: "https://example.com/model.safetensors"
      path: "checkpoints/my_model.safetensors"
      min_size: 2000000000
    - name: "Another Model"
      url: "https://example.com/model2.safetensors"
      path: "checkpoints/model2.safetensors"
      min_size: 4000000000

⁠Model Configuration Format

• name: Human-readable description
• url: Download URL
• path: Relative path from /workspace/ComfyUI/models/
• min_size: Minimum file size in bytes for validation

⁠🎨 Getting Started

⁠Load Sample Workflow

1. Open ComfyUI at http://localhost:8188
2. Click "Load" → "Upload"
3. Upload sample_workflow.json (included in this directory)
4. Click "Queue Prompt" to generate your first image!

⁠🐳 Docker Compose (Recommended)

Create docker-compose.yml:

version: '3.8'

services:
  comfyui-rocm:
    image: corundex/comfyui-rocm:latest
    container_name: comfyui-rocm
    devices:
      - /dev/kfd:/dev/kfd
      - /dev/dri:/dev/dri
    group_add:
      - video
    ports:
      - "8188:8188"
    volumes:
      # Model storage (checkpoints, VAE, LoRA, etc.)
      - ./data/models:/workspace/ComfyUI/models
      # Generated images and outputs
      - ./data/output:/workspace/ComfyUI/output
      # Input images for processing
      - ./data/input:/workspace/ComfyUI/input
      # Custom nodes and extensions
      - ./data/custom_nodes:/workspace/ComfyUI/custom_nodes
      # ComfyUI user settings and workflows
      - ./data/user:/workspace/ComfyUI/user
    environment:
      # Model download behavior: default, common, realistic, photorealistic, artistic, all, none
      - MODEL_DOWNLOAD=default
      # ROCm environment
      - HIP_VISIBLE_DEVICES=0
      - CUDA_VISIBLE_DEVICES=""
    restart: unless-stopped

Run with: docker compose up -d

⁠Recommended Volume Structure

./data/
├── models/           # AI models (checkpoints, VAE, LoRA, etc.)
│   ├── checkpoints/  # Main AI models (.safetensors)
│   ├── vae/         # VAE models for better image quality
│   ├── loras/       # LoRA fine-tuning models
│   ├── embeddings/  # Text embeddings
│   ├── controlnet/  # ControlNet guidance models
│   └── upscale_models/ # Image upscaler models
├── output/          # Generated images and videos
├── input/           # Input images for processing
├── custom_nodes/    # ComfyUI extensions
└── user/            # User settings and saved workflows

⁠🧪 Tested Hardware

• AMD Radeon RX 9060 XT (16GB VRAM) ✅

⁠⚡ Performance Notes

• Image Generation: ~30-60 seconds for 512x512 (depends on model and steps)
• Memory Usage: 4-8GB VRAM for basic operations
• Model Loading: ~30-60 seconds first time, cached afterward
• Batch Processing: Supported for multiple images

⁠🚀 Performance Tips

1. Use Persistent Volumes: Always mount model directories to avoid re-downloading
2. Start with default: Test with basic model first
3. Upgrade to all: Download full set when ready
4. Custom Sets: Create targeted collections for specific workflows
5. External Storage: Use fast SSD for model storage

⁠🔍 Troubleshooting

⁠Container won't start

# Check ROCm drivers
rocm-smi

# Check Docker GPU access
docker run --rm --device=/dev/kfd --device=/dev/dri rocm/pytorch:latest rocm-smi

⁠ComfyUI shows "No GPU detected"

# Verify ROCm in container
docker exec comfyui-rocm python -c "import torch; print(f'ROCm: {torch.cuda.is_available()}')"

⁠Model Download Failures

• Check internet connection
• Verify HuggingFace/source availability
• Check disk space in mounted volumes
• Review logs for specific error messages

⁠Model Not Found

• Ensure model downloaded successfully
• Check file permissions in mounted volumes
• Verify model file isn't corrupted (size check)

⁠Startup Issues

• Check MODEL_DOWNLOAD value is valid
• Ensure volumes are properly mounted
• Review container logs for errors

⁠Out of memory errors

• Reduce batch size in workflow
• Use smaller models
• Ensure sufficient VRAM (8GB minimum)

⁠📄 License

This Docker image packages ComfyUI with ROCm support. ComfyUI is licensed under GPL-3.0.

⁠🤝 Contributing

Found an issue or want to contribute? Visit our GitHub repository⁠ to:

• Report bugs
• Suggest improvements
• Submit pull requests
• Share your AMD GPU test results

⁠📄 License & Credits

This project is licensed under GPL-3.0.

⁠Third-Party Components

• ComfyUI: GPL-3.0 - ComfyUI⁠
• PyTorch: BSD 3-Clause - PyTorch⁠
• ROCm: Various OSS licenses - AMD ROCm⁠

⁠🙏 Acknowledgments

• ComfyUI⁠ - The amazing node-based UI
• AMD ROCm⁠ - Open source GPU computing platform
• ROCm community for AMD GPU AI support