ComfyUI ROCm Docker

A persistent Docker setup for ComfyUI with ROCm (AMD GPU) support, including a small test model for immediate use. Tested on the Minisforum N5 Pro and should also work on similar AMD-based systems such as the AI X1 Pro.

This repository includes:

• Dockerfile - builds a ROCm-enabled ComfyUI image with Python dependencies and a small test SD model.
• docker-compose.yml - aimplifies container startup with GPU access, bind mounts, and healthchecks.

Features

• Runs ComfyUI on AMD GPUs using ROCm
• Includes a persistent directory structure for models, outputs, inputs, and custom nodes
• Pre-downloads a small Stable Diffusion 1.5 test model
• Exposes web UI on port 8188
• Healthcheck ensures the UI is responding
• Compatible with Minisforum N5 Pro, AI X1 Pro, and other ROCm-enabled AMD systems

Requirements

• Docker >= 24.x
• Docker Compose >= 2.12.x
• AMD GPU with ROCm drivers installed on host
• Linux host (Ubuntu-based or TrueNAS SCALE recommended)

Setup

Build Docker Image

docker build -t comfyui-rocm-n5pro .

This will:

• Install system and Python dependencies
• Clone ComfyUI v0.9.1
• Create required directories (models, output, input, custom_nodes)
• Install Python dependencies for ComfyUI
• Download a small test SD model for immediate use

Start Container with Docker Compose

docker-compose up -d

This uses the prebuilt image (comfyui-basic2) and:

• Maps GPU devices (/dev/kfd and /dev/dri) for ROCm
• Mounts host directories for persistence
• Exposes the UI on http://localhost:8188
• Automatically restarts the container unless stopped

Usage

Access the ComfyUI Web UI:

http://<host-ip>:8188

• Test image generation with the pre-downloaded SD 1.5 model.
• Add your own models to ./models (mapped to /workspace/ComfyUI/models in the container).
• Custom nodes can be placed in /mnt/Storage/apps/comfy3/custom_nodes.

Directory Structure and Bind Mounts

Host Directory	Container Directory	Purpose
/mnt/Storage/apps/comfy3/models	/workspace/ComfyUI/models	Model storage (checkpoints, VAE, LoRAs, etc.)
/mnt/Storage/apps/comfy3/output	/workspace/ComfyUI/output	Generated outputs
/mnt/Storage/apps/comfy3/input	/workspace/ComfyUI/input	Input files for workflows
/mnt/Storage/apps/comfy3/custom_nodes	/workspace/ComfyUI/custom_nodes	Custom node extensions

Health Checks

The container is configured with a healthcheck that verifies the ComfyUI Web UI is responding every 30 seconds. If it fails 3 times in a row, Docker will mark the container as unhealthy.

Notes & Caveats

• Tested and confirmed working on Minisforum N5 Pro; likely compatible with AI X1 Pro or other AMD ROCm setups.
• A small test SD 1.5 model is downloaded by default. Add larger models as needed.
• ROCm runtime does not need to be specified in Compose; devices are exposed directly via /dev/kfd and /dev/dri.
• Shared memory (shm_size) is set to 16GB to improve performance on large models.

Acknowledgements

This project was made possible by Minisforum, who sent over an N5 Pro demo in summer 2025. At first, the system sat on the shelf - a few trade-offs and rough edges kept it there longer than expected.

With time, updates, and a fresh approach - running TrueNAS on bare metal - the N5 Pro revealed its potential as a compact AI powerhouse. GPU support now shines, and the system can really stretch its legs.

While the N5 Pro might not have hit every milestone its creators hoped for, it’s a testament to Minisforum’s willingness to experiment. Their boldness gives users the chance to explore, push boundaries, and find new ways to work with technology.

License

This repository is released under the MIT License.