ComfyUI-on-RX-9060-XT-RDNA-4---Docker-Setup-Guide

This guide documents how to run ComfyUI on the AMD Radeon RX 9060 XT (RDNA 4 / gfx1200) using Docker on Linux (Fedora/Ubuntu).

The Problem: RDNA 4 cards require ROCm 7.0+ libraries, which are not yet available as of now 2025 in standard Linux repositories or the default Docker images. The Solution: We use the standard ROCm Docker container but manually upgrade the internal PyTorch version to the "Nightly" build specifically compiled for gfx120x (RDNA 4).

Prerequisites Docker installed and running. GPU drivers installed on the host.

Step 1: Host System Permissions Before starting the container, we must ensure Docker has permission to read/write to the external drive where ComfyUI is stored. (Run these on your host machine)

Bash

1. Set the target directory variable

export COMFY_DIR="/run/media/hasan/ComfyUI/native_comfy"

2. Fix SELinux context (Crucial for Fedora)

sudo chcon -Rt svirt_sandbox_file_t "$COMFY_DIR"

3. Ensure read/write permissions

chmod -R 777 "$COMFY_DIR"

Step 2: Create the Persistent Container We create a container named comfy-rdna4 that runs in the background. We map the ComfyUI folder to /workspace.

Bash

docker run -d --name comfy-rdna4
--restart unless-stopped
--device=/dev/kfd --device=/dev/dri --group-add=video
-p 8188:8188
-v "$COMFY_DIR":/workspace
-e HSA_OVERRIDE_GFX_VERSION=12.0.0
rocm/pytorch:latest
tail -f /dev/null

Note: HSA_OVERRIDE_GFX_VERSION=12.0.0 is the recommended override for RDNA 4 if native detection fails.

Step 3: Install RDNA 4 Drivers (One-Time Setup) We need to replace the standard PyTorch version inside the container with the RDNA 4 compatible version.

1. Enter the container shell:

Bash docker exec -it comfy-rdna4 /bin/bash

2. Run these commands INSIDE the container:

Bash

A. Uninstall the incompatible stable version

pip uninstall -y torch torchvision torchaudio

B. Install the RDNA 4 Nightly version (Specific Index URL)

pip install --pre torch torchvision torchaudio --index-url https://rocm.nightlies.amd.com/v2/gfx120X-all/

C. Install ComfyUI dependencies

pip install -r requirements.txt

D. Install missing dependencies for Custom Nodes (Fixes Impact Pack crashes)

pip install scikit-image

Step 4: Usage Once setup is complete, you can launch ComfyUI anytime using this single command:

Bash docker exec -it comfy-rdna4 python main.py --listen