Can't get Stable Diffusion DirectML to work on my GPU

[CanisDirusPrime]

I am running a 5800X3d CPU with 32g of ram and a AMD RX6950XT. I've tried every explanation I've found both on the AMD forums and here to get Stable Diffusion DirectML to run on my GPU. It just refuses to do so. Been trying for a week now, to no avail. I do not have an ONNX tab nor an Olive tab in the UI, either (both of which have been downloaded several times). Any help would be greatly appreciated. Here's links to some of the things I've tried (just a few of the ones I've tried, btw):

https://community.amd.com/t5/ai/how-to-automatic1111-stable-diffusion-webui-with-directml/ba-p/649027

https://www.youtube.com/watch?v=mKxt0kxD5C0

https://github.com/microsoft/Olive/blob/main/examples/directml/stable_diffusion/README.md#setup

And of course the README for SD and for DiretML versions, too.

I've downloaded the Stable-Diffusion-WebUI-DirectML, the k-diffusion and Stability-AI's stablediffusion Extensions, also. None of these seem to make a difference.

I've enabled the ONNX runtime in settings, enabled Olive in settings (along with all the check boxes required) added the sd_unet checkpoint model thing (whatever you call it) under quick settings.

My webui-user.bat file ARGS are: --no-half --precision full --no-half-vae --medvram --opt-sub-quad-attention --opt-split-attention-v1 --skip-torch-cuda-test. Tried adding --use-directml, but that just throws an error and it won't start at all.

I've also tried it with and without xformers

The one thing that gives me hope is my GPU usage flashes up to between 4 and 7% for a spit second then the CPU goes to 40-50% and it sits there for 3-5 minutes before i get the image.

Finally, I have tried both the standard stable_diffusion_webui and the stable_diffusion_webui_diretml versions with all of the options, to no avail. I've tried running them from miniconda and python 3.10.6 directly and in different environments (I have a couple, olive-env and automatic_dmlplugin, mainly)

Here's Conda code that runs at startup:
(automatic1111_olive) C:\stable-diffusion-webui-directml>webui-user.bat
venv "C:\stable-diffusion-webui-directml\venv\Scripts\Python.exe"
fatal: No names found, cannot describe anything.
Python 3.10.13 | packaged by Anaconda, Inc. | (main, Sep 11 2023, 13:24:38) [MSC v.1916 64 bit (AMD64)]
Version: 1.7.0
Commit hash: 7ed2ff1
no module 'xformers'. Processing without...
no module 'xformers'. Processing without...
No module 'xformers'. Proceeding without it.
C:\stable-diffusion-webui-directml\venv\lib\site-packages\pytorch_lightning\utilities\distributed.py:258: LightningDeprecationWarning: pytorch_lightning.utilities.distributed.rank_zero_only has been deprecated in v1.8.1 and will be removed in v2.0.0. You can import it from pytorch_lightning.utilities instead.
rank_zero_deprecation(
Invalid version: 'onnxruntime-directml'
Warning: Failed to install onnxruntime-directml package, DirectML extension will not work.
Launching Web UI with arguments: --no-half --precision full --no-half-vae --medvram --opt-sub-quad-attention --opt-split-attention-v1 --skip-torch-cuda-test
Style database not found: C:\stable-diffusion-webui-directml\styles.csv
Warning: caught exception 'Found no NVIDIA driver on your system. Please check that you have an NVIDIA GPU and installed a driver from http://www.nvidia.com/Download/index.aspx', memory monitor disabled
C:\stable-diffusion-webui-directml\venv\lib\site-packages\diffusers\utils\outputs.py:63: UserWarning: torch.utils._pytree._register_pytree_node is deprecated. Please use torch.utils._pytree.register_pytree_node instead.
torch.utils._pytree._register_pytree_node(
C:\stable-diffusion-webui-directml\venv\lib\site-packages\diffusers\utils\outputs.py:63: UserWarning: torch.utils._pytree._register_pytree_node is deprecated. Please use torch.utils._pytree.register_pytree_node instead.
torch.utils._pytree._register_pytree_node(
ONNX: selected=DmlExecutionProvider, available=['DmlExecutionProvider', 'CPUExecutionProvider']
ControlNet preprocessor location: C:\stable-diffusion-webui-directml\extensions\sd-webui-controlnet\annotator\downloads
2024-02-16 15:33:08,478 - ControlNet - INFO - ControlNet v1.1.440
2024-02-16 15:33:08,592 - ControlNet - INFO - ControlNet v1.1.440
*** Error loading script: img2img.py
Traceback (most recent call last):
File "C:\stable-diffusion-webui-directml\modules\scripts.py", line 469, in load_scripts
script_module = script_loading.load_module(scriptfile.path)
File "C:\stable-diffusion-webui-directml\modules\script_loading.py", line 10, in load_module
module_spec.loader.exec_module(module)
File "", line 883, in exec_module
File "", line 241, in _call_with_frames_removed
File "C:\stable-diffusion-webui-directml\extensions\stablediffusion\scripts\img2img.py", line 16, in
from imwatermark import WatermarkEncoder
ModuleNotFoundError: No module named 'imwatermark'

---

*** Error loading script: txt2img.py
Traceback (most recent call last):
File "C:\stable-diffusion-webui-directml\modules\scripts.py", line 469, in load_scripts
script_module = script_loading.load_module(scriptfile.path)
File "C:\stable-diffusion-webui-directml\modules\script_loading.py", line 10, in load_module
module_spec.loader.exec_module(module)
File "", line 883, in exec_module
File "", line 241, in _call_with_frames_removed
File "C:\stable-diffusion-webui-directml\extensions\stablediffusion\scripts\txt2img.py", line 14, in
from imwatermark import WatermarkEncoder
ModuleNotFoundError: No module named 'imwatermark'

---

Loading weights [ec41bd2a82] from C:\stable-diffusion-webui-directml\models\Stable-diffusion\photon_v1.safetensors
2024-02-16 15:33:09,002 - ControlNet - INFO - ControlNet UI callback registered.
Creating model from config: C:\stable-diffusion-webui-directml\configs\v1-inference.yaml
Running on local URL: http://127.0.0.1:7860

To create a public link, set share=True in launch().
Startup time: 4.9s (prepare environment: 7.8s, initialize shared: 1.4s, load scripts: 1.1s, create ui: 0.8s, gradio launch: 0.5s).
Applying attention optimization: sub-quadratic... done.
Model loaded in 3.9s (load weights from disk: 1.0s, create model: 0.5s, apply weights to model: 2.0s, apply float(): 0.4s, calculate empty prompt: 0.1s).

When generating a txt2img prompt (Flying golden dragon), this is the code I see:
100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [02:27<00:00, 7.35s/it]
Total progress: 100%|██████████████████████████████████████████████████████████████████| 20/20 [02:24<00:00, 7.21s/it]
Total progress: 100%|██████████████████████████████████████████████████████████████████| 20/20 [02:24<00:00, 7.27s/it]

And here's my sensor panel while this is going on:

And the image it produced (not a bad image, actually, for just 3 words):

If ya'll need any other info to help me figure this out, I'm all ready to give it. Thanks in advance!

[waylaa]

You read every guide except for this. Re-install everything from scratch by following the steps from the link.

When you are done installing, edit the webui-user.bat file, navigate to where COMMANDLINE_ARGS is and type --skip-torch-cuda-test --use-directml

[CanisDirusPrime]

I wish I could say you were right, LOL! That is the most recent one I read and am still having the same issue. And every time, it's a fresh install from scratch. This is the guide that told me to use the --use-directml commandline arg that crashes on start. Even tried it with just the skip cuda test ARG and still CTD.

It takes well over an hour to install from scratch - no fibre connection where i live and no starlink (yet). Thanks for the advice though! If y you've got anything else to try, I'm all ears.

[teflon1104]

hey, i have same PC as you and i have same problem... :( I have question for u: When now after fix this graphics card get a load now when generating a photo?

[CanisDirusPrime]

A1111 never accessed my card. It worked in ComfyUI, but it was never great (it took anywere from 3 to 5 minutes to generate an image). Have permanently switched over to Comfy and now am the proud owner of an EVGA RTX3090 which only takes 20-30 seconds to generate an image and roughly 45-60 seconds with the HIRes fix (upscale) turned on. Which, by the way, I don't turn on very often cause there's no need. The generated images are awesome. Here's a couple with Godzilla. I'm now learning perspectives in Comfy and images like this will only improve.