[Support missing] ROCm for Windows 11

[johnmart19]

RuinedFoocus typically is trying to set "DirectML" for Windows, however it is running on Pytorch with ROCm support with 6.2 ROCm environment

Logs with
"call .\venv\Scripts\activate.bat
python.exe launch.py --offline --gpu-only"

Skip checking python modules.
Python 3.11.7 (tags/v3.11.7:fa7a6f2, Dec 4 2023, 19:24:49) [MSC v.1937 64 bit (AMD64)]
RuinedFooocus version: 25.07.13
Skip pip check.
Torch platform: Windows: directml
Skip check of required modules.
Starting webui

:: Checking package versions...
Found pydantic: 2.11.4, pydantic-settings: 2.10.1
:: Pydantic packages are compatible, skipping reinstall
Installed version of comfyui-frontend-package: 1.23.4
Installed version of comfyui-workflow-templates: 0.1.35
Installed version of av: 15.0.0
Installed version of comfyui-embedded-docs: 0.2.4
:: Package version check complete.
:: CUDA device detected: AMD Radeon RX 7900 XTX
--------------------------------------------------------

Using default pipeline.
Loading base model: flux1-schnell.safetensors
WARNING:root:Warning, This is not a checkpoint file, trying to load it as a diffusion model only.

• Running on local URL: http://127.0.0.1:7860
• To create a public link, set share=True in launch().
  WARNING:root:WARNING: No VAE weights detected, VAE not initalized.
  Loading CLIP: ['clip_l.safetensors', 't5-v1_1-xxl-encoder-Q3_K_L.gguf']
  gguf qtypes: Q3_K (120), F32 (50), Q5_K (48), Q6_K (1)
  WARNING:root:clip missing: ['text_projection.weight']
  Loading VAE: flux-vae.safetensors

[johnmart19]

Additional information

Installation type was pretty silmilar to the following: PyTorch + ROCm + Windows 11
For ComfyUI: ComfyUI with ROCm on Windows 11

[johnmart19]

P.s I am pretty new in this topic, so could missunderstood various stuffs, feel free to share opinions / details

[yownas]

Not sure if I understood what the question was. :) But...

• RF usually use its own python 3.10, don't see any reason why 3.11 wouldn't work, but it is nothing we've tested.
• Since me and Rune only have NVIDIA cards, AMD isn't anything we've been able to test. But we use https://github.com/easydiffusion/torchruntime to try to figure out "the best" settings.
• The warnings you get are nothing to worry about, if RF tries to load a model that is missing the text encoders (sometimes called CLIP) or VAE, it will try to find suitable versions to load. (Many (most even?) Flux models only have the unet part so that is why it happens.) But, again, nothing to worry about. It is just RF trying to be helpful.
• Some of the output makes it look like you don't use RF's comfyui repo, there shouldn't be any issue to use any reasonably new version, but it is nothing we've tested.

[johnmart19]

understood what the question was.

If launching on Ubuntu (native 24.04), default environment is set to rocm6.2, while on Windows it always defaulted to "DirectML", however custom env with ROCm modification used, perf on Ubuntu is even worse in same scenario on my HW, mostly was curious if check for ROCm availability can be added to Windows and show respectfully to the environment

[johnmart19]

@yownas To launch on AMD was doing following steps:

RuinedFoocus AMD GPU (In my case RX 7900 XTX)

To make write instructions also did clean Windows installation

Preparation:

1. Install latest AMD GPU Driver (in my case Adrenalin Edition 25.6.3)
2. Install latest AMD HIP SDK for Windows (in my case ROCm 6.2.4)
   P.s In HIP SDK I selected to install all packages fully as usual, without overwriting 25.6.3 latest driver by PRO 24.20.52 which is obviously older
3. Installed python 3.11.9 (3.11.9 is needed for ROCm wheels)
4. Installed git for Windows
5. Adding HIP SDK to Path:

• Settings - system - about system - Additional system settings
• Advaned - Environment Variables - System Variables - Create
  Name: HIP_PATH
  Variable value: C:\Program Files\AMD\ROCm\6.2\
• Additional step:
  Open "Path" by double-click - New -
  C:\Program Files\AMD\ROCm\6.2\bin\
  ** Reboot PC afterwards **

Starting installation of RuinedFoocus dependencies and some fixups:

1. Clone RuinedFoocus repo by going to folder where you want to have RuinedFoocus to be located and clone it by opening cmd and pasting
   git clone https://github.com/runew0lf/RuinedFooocus

Go to downloaded folder
2) cd RuinedFooocus

Create virtual python environment (venv)
3) py -3.11 -m venv venv

Activate virtual python environment
4) .\venv\Scripts\activate

Start basic dependencies installation
5) python entry_with_update.py

Installation typically completes with various errors due to directml packages and "AssertionError: Torch not compiled with CUDA enabled" in the end

6. Downloading Pytorch wheels for gfx110x, gfx1151, and gfx1201 and copying them to RuinedFoocus path
   In my case the files are:
   torch-2.7.0a0+rocm_git3f903c3-cp311-cp311-win_amd64.whl
   torchaudio-2.7.0a0+52638ef-cp311-cp311-win_amd64.whl
   torchvision-0.22.0+9eb57cd-cp311-cp311-win_amd64.whl

Install ROCm Torch, TorchAudio, TorchVision
7) pip install "torch-2.7.0a0+rocm_git3f903c3-cp311-cp311-win_amd64.whl" "torchaudio-2.7.0a0+52638ef-cp311-cp311-win_amd64.whl" "torchvision-0.22.0+9eb57cd-cp311-cp311-win_amd64.whl"

Remove DirectML which is not supported
8) pip uninstall torch-directml

Install compatible version of Numpy for ComfyUI (removing numpy-2.2.6 and installing numpy-1.26.4)
9) pip install numpy==1.26.4 --force-reinstall

Downgrade OpenCV to work with current Numpy
10) pip install opencv-python-headless==4.11.0.86

Start RuinedFoocus with bypass checks of current packages
11) python entry_with_update.py --offline

For future starts typically making AMD_Launch.bat file:

@echo off
cd D:\RuinedFooocus
call .\venv\Scripts\activate
python entry_with_update.py --offline --gpu-only

If nothing changed, here are logs from 3.10:
RuinedFoocus_Python_3.10_clean_run_failed.txt
And 3.11.9 installation fails:
RuinedFoocus_Python_3.11.9_clean_run_failed.txt

P.s on 3.10 Custom wheels were not working "Not supported platform", on 3.11.9 works fine

[johnmart19]

After described details, works fine on clean install

[johnmart19]

:: Checking package versions...
Found pydantic: 2.11.4, pydantic-settings: 2.10.1
:: Pydantic packages are compatible, skipping reinstall
Installed version of comfyui-frontend-package: 1.23.4
Installed version of comfyui-workflow-templates: 0.1.35
Installed version of av: 15.0.0
Installed version of comfyui-embedded-docs: 0.2.4
:: Package version check complete.
:: CUDA device detected: AMD Radeon RX 7900 XTX
--------------------------------------------------------

This was with ComfyUI-Zluda

Installation was just on top of regular ComfyUI:
Going to path

1. cd repositories/ComfyUI
   Fetch & Reset
2. git fetch https://github.com/patientx/ComfyUI-Zluda && git reset --hard FETCH_HEAD
   Installing dependencies (it creates own venv folder)
3. install.bat
   P.s it will give error "RuntimeError: CUDA unknown error - this may be due to an incorrectly set up environment, e.g. changing env variable CUDA_VISIBLE_DEVICES after program start. Setting the available devices to be zero." which is normal
   ComfyUI-Zluda installation logs .txt

Installing Zluda patch (pick link for matching from https://github.com/lshqqytiger/ZLUDA/releases)
4) patchzluda2.bat -> in my case ZLUDA-windows-rocm6-amd64.zip
Example:
ComfyUI-Zluda patchzluda2 logs.txt

Finally launch ComfyUI
5) comfyui.bat

We are getting initial succesfull launch with WebUI:
ComfyUI-Zluda initial succesfull launch.txt

Reviewing RuinedFoocus env:
Go back to RuinedFoocus folder

1. cd ../..
   Re-launch venv just in case
   2) .\venv\Scripts\activate
   Start RuinedFoocus (offline to prevent python packages reinstallation)
2. python entry_with_update.py --offline --gpu-only

Succesfull RuinedFoocus launch with ComfyUI-Zluda working on bg:
RuinedFoocus_with_ComfyUI-Zluda.txt

[johnmart19]

With that setup it is working as is and is typically working with all the models that I tried, only failed with Flux models as they does not fit in 24GB VRam and are taking way too much time to generate anything
Pony, illusion, LLM, other models typically working without any issues

"PowerUP" features also works fine

User:

Abandoned Church, architectural installation, reveals the haunting beauty of decay, weathered stone and crumbling masonry exude a rustic charm, muted earth tones and faded pastels highlight the structure's age. The church is framed by arched doorways and broken stained glass, shafts of light stream through, casting ethereal patterns on the floor. The background features overgrown foliage and encroaching vines, enveloping the church, creating a poignant contrast between nature and the remnants of faith.

art maker: (Thinking...)

Abandoned Church, architectural installation, reveals the haunting beauty of decay, weathered stone and crumbling masonry exude a rustic charm, muted earth tones and faded pastels highlight the structure's age. The church is framed by arched doorways and broken stained glass, shafts of light stream through, casting ethereal patterns on the floor. The background features overgrown foliage and encroaching vines, enveloping the church, creating a poignant contrast between nature and the remnants of faith.
Wooden frame with metal accents, natural wood tones, soft golden lighting, moss-covered stone base

Time taken: 5.03 seconds
LoRAs loaded: []
100%|██████████████████████████████████████████████████████████████████████████████████| 30/30 [00:08<00:00, 3.48it/s]
Time taken: 9.16 seconds Pipeline process

Img2img, Upscale and other features also typically works fine, same with custom resolutions and LoRA's

That's not the best, but at least it's working, but yep, someone with better knowledge is needed to revisit what can be reworked to make it better

[yownas]

Thanks for all the testing. :) I wonder if this is something @iwr-redmond would be interested in?
I think torchruntime just see 7xxx and Windows and defaults to directml. Even if there seem to be better options.

I also changed so that --offline is "more offline", so it will skip the torch-check completely. (not pushed to main yet, need to test it some more.)

EDIT: Also testing downgrading numpy so it should work on on AMD ootb.

[johnmart19]

@yownas From strange stuff, if I use AMD GPU as primary (7900xtx 24gb) and GTX 1060 3GB for second monitor, RuinedFoocus defaults to cuda env while still using rx 7900 xtx

[johnmart19]

(load on gtx1060 is not from RuinedFoocus, it's just showing YT videos on second 1080p@240hz monitor)

[iwr-redmond]

Thanks for all the testing. :) I wonder if this is something @iwr-redmond would be interested in? I think torchruntime just see 7xxx and Windows and defaults to directml. Even if there seem to be better options.

See torchruntime#17. The decision @cmdr2 will need to make is at what point TheRock builds will be judged as good enough that they can be rolled out to ordinary Windows users of targeted AMD architectures. I reckon it's a bit early yet, as the builds are still alpha-quality.

[johnmart19]

torchruntime#17.

TorchVision is not available in https://github.com/ROCm/TheRock repo for Windows, so most likely not usable environment

Need or to wait for official support
Or to make own fork, figure out patches used in https://github.com/scottt/rocm-TheRock for the latest version, recompile, and potentially make pull requests to main branches and if approved - then will be "Stable" versions for rx 7900 xtx etc

[iwr-redmond]

Support varies by architecture, e.g. TorchVision is available for Strix Halo already. However, I agree that integrating these builds is premature for everyday use.

Updated: erroneous claim

[iwr-redmond]

Torchvision for Windows is now up for gfx1151 (Strix Halo). Scroll to the very bottom of the page for the links.