comfyui-strix-docker

ComfyUI on AMD Strix Halo (RDNA 3.5 / gfx1151) via Docker. Ubuntu 26.04 LTS + UV-managed Python 3.12 + TheRock ROCm 7.13 wheels. Fixes the silent CPU fallback that Debian / Python 3.13 images hit on gfx1151.

---

What this is

A Docker container for running ComfyUI on AMD's Strix Halo (Ryzen AI Max+ 3xx) APU. It targets gfx1151 directly so the GPU is actually used, instead of the silent CPU fallback you get from generic images.

Why this repo exists

Generic images fail on Strix Halo for three reasons:

1. Permission: the container needs the host's exact render / video group IDs to reach /dev/kfd and /dev/dri.
2. Debian pip: plain pip on Debian installs a CPU-only torch and silently falls back to CPU.
3. Python version: Ubuntu 26.04 ships Python 3.13; the gfx1151 ROCm wheels are built for 3.12.

Solution: Ubuntu 26.04 LTS for a stable, current base, with uv building an isolated Python 3.12 venv inside the container, and TheRock ROCm 7.13 wheels pinned to one coherent build so the install is reproducible.

About the ROCm pin

There is no stable PyPI index for gfx1151, so the Dockerfile pins one same-date build off TheRock's nightly index rather than floating "latest". The three packages must stay on the same line (an ABI set):

package	version	tag
torch	2.9.1	+rocm7.13.0a20260513
torchvision	0.24.0	+rocm7.13.0a20260513
torchaudio	2.9.0	+rocm7.13.0a20260513

To bump, pick a date present for all three on the same line and update the ARGs in the Dockerfile:

for p in torch torchvision torchaudio; do
  curl -fsSL https://rocm.nightlies.amd.com/v2/gfx1151/$p/ | grep rocm7.13.0a<DATE>
done

This is the same ROCm stack the gamentic project runs on the same gfx1151 box.

Prerequisites

• Host OS: Ubuntu 25.10+ / a recent 6.x kernel with amdgpu
• Hardware: AMD Ryzen AI Max+ 3xx (Strix Halo / RDNA 3.5, gfx1151)
• Docker + Compose

Installation

1. Clone

git clone https://github.com/hec-ovi/comfyui-strix-docker.git
cd comfyui-strix-docker

2. Configure permissions (required)

Tell Docker your GPU group IDs:

getent group video  | cut -d: -f3   # -> VIDEO_GID
getent group render | cut -d: -f3   # -> RENDER_GID

Copy the template and fill in your numbers and model path:

cp .envTemplate .env

# .env
VIDEO_GID=44       # your 'video' GID
RENDER_GID=990     # your 'render' GID
MODELS_PATH=/path/to/your/models

3. Launch

docker compose up -d --build

UI: http://localhost:8188

Model management

Map a local folder so models persist across rebuilds. Suggested layout:

~/workspace/models/comfy/
├── checkpoints/  (SD checkpoints)
├── unet/         (FLUX.2 diffusion models)
├── clip/         (FLUX.2 text encoders / T5)
├── vae/          (FLUX.2 VAE)
└── loras/        (LoRAs)

Drop files in and hit "Refresh" in the UI.

Technical details

• Base image: ubuntu:26.04 (LTS)
• Python: 3.12, pinned and built by uv (ignores the system 3.13)
• PyTorch: TheRock ROCm 7.13 wheels for gfx1151, pinned (see the table above)
• GPU runtime env (set in the Dockerfile, overridable from compose):
  • HSA_OVERRIDE_GFX_VERSION=11.5.1 pins the ISA for Strix Halo
  • HSA_ENABLE_SDMA=0 and HSA_USE_SVM=0 are the community stability fixes for unified memory (avoid GPU ring timeouts / checkerboard artifacts during VAE decode)
• The build runs an import torch; assert 'rocm' in torch.__version__ check, so a CPU-only torch fails the build instead of failing silently at runtime.

Verified on a Ryzen AI Max+ 395 (gfx1151, June 2026): the image builds against ubuntu:26.04 with torch 2.9.1+rocm7.13.0a20260513, and a FLUX.2 klein graph renders a 1024x1024 image in about 22s with the weights in GTT (unified RAM) and the VRAM carve-out left at its idle baseline.

Troubleshooting

• "Torch not compiled with CUDA enabled": the GIDs in .env are wrong, or your kernel is too old for amdgpu on gfx1151.
• Permission denied on the GPU: add your user to the host render and video groups.

---

License

MIT for the build glue in this repository (Dockerfile, docker-compose.yml, entrypoint.sh, .envTemplate, README, scripts).

This repository does not redistribute ComfyUI source. The Dockerfile clones ComfyUI from github.com/comfyanonymous/ComfyUI at build time. ComfyUI is licensed under GPL-3.0. Any image built from this repository that bundles ComfyUI is therefore a derivative work of ComfyUI and is subject to GPL-3.0 when distributed. Building locally for personal use is fine; redistributing the resulting image, or any modified version of ComfyUI, requires complying with GPL-3.0.