Installing ComfyUI on AMD APUs

This step-by-step guide shows how to install ComfyUI as a systemd service running under a dedicated system user on Strix Halo or Strix Point machines. It has been written and tested on Arch Linux, but the instructions should be generic enough to apply to other distributions.

The goal of this guide is to end up with a configuration that works well, runs fast, and allows you to tinker when and if you feel like it.

Prerequisites

Make sure you have the following installed:

uv - A fast Python package installer and manager
git - Version control system

You can install uv and git on most Linux distributions:

# Arch Linux
sudo pacman -S uv git

# Ubuntu/Debian
sudo apt install uv git

Creating the ComfyUI User

We want to create a dedicated system user to run ComfyUI. This is a good security practice. This user won’t be able to log in and doesn’t have a password. It will only be used to run ComfyUI. Run these commands:

sudo useradd --system -s /usr/bin/nologin -d /var/lib/comfyui comfyui
sudo mkdir /var/lib/comfyui
sudo chown comfyui:comfyui /var/lib/comfyui

You can change all instances of /var/lib/comfyui in the commands above to your preferred location.

Installing ComfyUI

Now let’s install ComfyUI. Run these commands:

sudo -u comfyui bash
cd
git clone https://github.com/Comfy-Org/ComfyUI .
uv venv --python 3.13
uv pip install -r requirements.txt --extra-index-url https://rocm.nightlies.amd.com/v2-staging/gfx1151
uv pip install -r manager_requirements.txt
exit

Note 1: Python 3.13 is recommended for the time being; uv will handle this automatically.

Note 2: For Strix Point, the URL becomes https://rocm.nightlies.amd.com/v2-staging/gfx1150.

Running with systemd

We now need to create a systemd unit file to run ComfyUI. Open /etc/systemd/system/comfyui.service:

sudo nano /etc/systemd/system/comfyui.service

Then paste the following content:

[Unit]
Description=ComfyUI
After=network-online.target

[Service]
Type=simple
User=comfyui
Group=comfyui
StateDirectory=comfyui
CacheDirectory=comfyui
ProtectHome=false
WorkingDirectory=/var/lib/comfyui
Environment=VIRTUAL_ENV=/var/lib/comfyui/.venv
Environment=PATH=/var/lib/comfyui/.venv/bin:/usr/bin:/bin
Environment=TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1
Environment=COMFYUI_ENABLE_MIOPEN=0
Environment=TORCH_BLAS_PREFER_HIPBLASLT=1
Environment=TORCHINDUCTOR_CACHE_DIR=/var/lib/comfyui/.cache/torchinductor
Environment=CC=/var/lib/comfyui/.venv/bin/amdclang
Environment=CXX=/var/lib/comfyui/.venv/bin/amdclang++
Environment=FLASH_ATTENTION_TRITON_AMD_ENABLE=TRUE
ExecStart=/var/lib/comfyui/.venv/bin/python /var/lib/comfyui/main.py --enable-manager --use-pytorch-cross-attention
Restart=on-failure
RestartSec=5s
KillSignal=SIGINT

[Install]
WantedBy=multi-user.target

Remember to replace all instances of /var/lib/comfyui with the actual path you are using if applicable. Some of the environment variables set here aren’t strictly needed but are included to support future tinkering.

Now, if you want ComfyUI to start automatically at boot, run the following:

sudo systemctl enable --now comfyui

If instead you want to run it just once, run this instead:

sudo systemctl start comfyui

Running on a Local Network

If you plan to access ComfyUI on your local network - for example, on a dedicated machine—replace the ExecStart line in the systemd unit file with the following:

ExecStart=/var/lib/comfyui/.venv/bin/python /var/lib/comfyui/main.py --listen --enable-manager --use-pytorch-cross-attention

Additionally, if you want to use ComfyUI Manager from other computers on your local network, do the following:

sudo -u comfyui bash
cd
nano user/__manager/config.ini

and search for the line:

network_mode = public

and replace it with:

network_mode = personal_cloud

Then you will need to restart ComfyUI:

sudo systemctl daemon-reload
sudo systemctl restart comfyui

Another tip—especially applicable if you only access ComfyUI remotely (for example, if it runs on a headless server)—is to install the Workflow Models Downloader plugin through ComfyUI Manager. This provides a user-friendly interface to download models from Hugging Face and CivitAI.

Updating ComfyUI

Periodically, you will want to update ComfyUI. This can be achieved with the following commands:

sudo -u comfyui bash
cd
git pull
uv pip install -r requirements.txt --extra-index-url https://rocm.nightlies.amd.com/v2-staging/gfx1151
uv pip install -r manager_requirements.txt

Once again, if you are on Strix Point, adjust the URL to https://rocm.nightlies.amd.com/v2-staging/gfx1150.

As the uv cache tends to grow, periodically you may also want to run:

uv cache clean

to recover some disk space.

VRAM Configuration and Performance Tuning

By default, the amdgpu driver allows up to 50% of system memory to be used as VRAM. While this is a safe default, we need more VRAM to run larger models, so we will change this setting. However, I do not advise allowing the GPU to allocate 100% of RAM. My rule of thumb is to reserve at least 16GB of RAM for the system if you use a desktop environment, and at least 8GB if you do not. Here’s how to set this:

VRAM Settings in the BIOS

Set the iGPU VRAM size to the smallest value available (usually 512MB). This may seem counter-intuitive, but the VRAM value set in BIOS is not what we will actually use.

VRAM Settings in Linux

The exact procedure may depend on your distribution. Note that the memory to allocate is expressed in pages, not bytes. This page from the Arch wiki provides a useful table with common sizes. The instructions there should be fairly generic, but please refer to distribution-specific tutorials for guidance.

Tuned

This is suggested in many guides. I find the effect to be very small, but it does seem to exist. Just install the tuned package and run:

sudo systemctl enable --now tuned
sudo tuned-adm profile accelerator-performance

ComfyUI

We can achieve some non-negligible performance gains by enabling torch.compile for models. You can download my workflows for Flux 2 Klein (works with 9b and 4b), Z-Image (works with Turbo and Base), and Anima if you want a quick way to set this up. You will need to download some custom nodes, but ComfyUI Manager should make this very easy.