[Issue]: AMD GPU is not being utilized, although ZLUDA loads successfully

[Kolopsel]

Issue Description

I'm unfortunately having a bit of a problem trying to get the newest version of SDNext running on my AMD GPU utilizing ZLUDA. The following output log is from a fresh SDNext installation (dev-branch) on Windows 11 with the newest AMD display drivers (25.5.1) and ROCM 6.2 (HIP-SDK) installed.

I tried all of this on the master branch of SDNext aswell, but unfortunately got the exact same result. During my first boot of SDNext I used the following startup-arguments "--use-zluda --use-nightly --debug".

17:36:06-588150 INFO     Starting SD.Next
17:36:06-588150 INFO     Logger: file="C:\Users\User\Desktop\SDNext\sdnext-dev\sdnext\sdnext.log" level=DEBUG host="DESKTOP-9PQ20S2" size=83 mode=create
17:36:06-588150 INFO     Python: version=3.11.9 platform=Windows bin="C:\Users\User\Desktop\SDNext\sdnext-dev\sdnext\venv\Scripts\python.exe"
                         venv="C:\Users\User\Desktop\SDNext\sdnext-dev\sdnext\venv"
17:36:06-721881 INFO     Version: app=sd.next updated=2025-05-08 hash=55b1cb8c branch=dev url=https://github.com/vladmandic/sdnext.git/tree/dev ui=main
17:36:06-721881 DEBUG    Branch mismatch: sdnext=dev ui=main
17:36:06-876008 DEBUG    Branch sync failed: sdnext=dev ui=dev
17:36:07-557417 INFO     Version: check latest available hash=9141e060 updated=2025-05-06T18:17:05Z
17:36:07-568105 INFO     Platform: arch=AMD64 cpu=AMD64 Family 25 Model 97 Stepping 2, AuthenticAMD system=Windows release=Windows-10-10.0.26100-SP0 python=3.11.9 locale=('English_Germany',
                         '1252') docker=False
17:36:07-568105 DEBUG    Packages: prefix=venv site=['venv', 'venv\\Lib\\site-packages']
17:36:07-568105 WARNING  Packages: site="venv\Lib\site-packages" invalid=['~umpy', '~umpy.libs'] removing
17:36:07-586254 INFO     Args: ['--use-zluda', '--use-nightly', '--debug']
17:36:07-586254 DEBUG    Setting environment tuning
17:36:07-586254 DEBUG    Torch allocator: "garbage_collection_threshold:0.80,max_split_size_mb:512"
17:36:07-596247 DEBUG    Torch overrides: cuda=False rocm=False ipex=False directml=False openvino=False zluda=True
17:36:07-616162 INFO     ROCm: AMD toolkit detected
17:36:08-276185 INFO     ROCm: agents=['gfx1100', 'gfx1036']
17:36:08-276185 INFO     ROCm: version=6.2, using agent gfx1100
17:36:08-276185 INFO     Python: version=3.11.9 platform=Windows bin="C:\Users\User\Desktop\SDNext\sdnext-dev\sdnext\venv\Scripts\python.exe"
                         venv="C:\Users\User\Desktop\SDNext\sdnext-dev\sdnext\venv"
17:36:08-288268 INFO     ZLUDA load: path='C:\Users\User\Desktop\SDNext\sdnext-dev\sdnext\.zluda' nightly=True
17:36:10-183246 INFO     Diffusers update: package=diffusers 0.34.0.dev0 current= target=8c661ea586bf11cb2440da740dd3c4cf84679b85
17:36:36-647650 INFO     Install: verifying requirements
17:36:36-686731 DEBUG    Timestamp repository update time: Thu May  8 16:14:15 2025
17:36:36-689482 INFO     Startup: standard
17:36:36-689482 INFO     Verifying submodules
17:36:38-440967 DEBUG    Git submodule: extensions-builtin/sd-extension-chainner / main
17:36:38-506486 DEBUG    Git submodule: extensions-builtin/sd-extension-system-info / main
17:36:38-566676 DEBUG    Git submodule: extensions-builtin/sd-webui-agent-scheduler / main
17:36:38-623246 DEBUG    Git submodule: extensions-builtin/sdnext-modernui / dev
17:36:38-682587 DEBUG    Git submodule: extensions-builtin/stable-diffusion-webui-rembg / master
17:36:38-740609 DEBUG    Git submodule: modules/k-diffusion / master
17:36:38-800487 DEBUG    Git submodule: wiki / master
17:36:38-838614 DEBUG    Register paths
17:36:38-886689 DEBUG    Installed packages: 160
17:36:38-888703 DEBUG    Extensions all: ['Lora', 'sd-extension-chainner', 'sd-extension-system-info', 'sd-webui-agent-scheduler', 'sdnext-modernui', 'stable-diffusion-webui-rembg']
17:36:39-018203 DEBUG    Extension installer: builtin=True file="C:\Users\User\Desktop\SDNext\sdnext-dev\sdnext\extensions-builtin\sd-webui-agent-scheduler\install.py"
17:36:39-118608 DEBUG    Extension installer: builtin=True file="C:\Users\User\Desktop\SDNext\sdnext-dev\sdnext\extensions-builtin\stable-diffusion-webui-rembg\install.py"
17:36:39-169439 DEBUG    Extensions all: []
17:36:39-170443 INFO     Extensions enabled: ['Lora', 'sd-extension-chainner', 'sd-extension-system-info', 'sd-webui-agent-scheduler', 'sdnext-modernui', 'stable-diffusion-webui-rembg']
17:36:39-170443 INFO     Install: verifying requirements
17:36:39-171957 DEBUG    Setup complete without errors: 1746718599
17:36:39-179981 DEBUG    Extension preload: {'extensions-builtin': 0.0, 'extensions': 0.0}
17:36:39-179981 INFO     Installer time: total=62.94 pip=26.36 diffusers=26.36 torch=2.60 git=2.29 submodules=2.15 latest=0.96 amd=0.70 branch=0.42 version=0.24 base=0.18 ui=0.15
                         sd-webui-agent-scheduler=0.06 files=0.06 requirements=0.05 installed=0.05 stable-diffusion-webui-rembg=0.05
17:36:39-184089 INFO     Command line args: ['--use-zluda', '--use-nightly', '--debug'] use_nightly=True use_zluda=True debug=True args=[]
17:36:39-184089 DEBUG    Env flags: []
17:36:39-186610 DEBUG    Starting module: <module 'webui' from 'C:\\Users\\User\\Desktop\\SDNext\\sdnext-dev\\sdnext\\webui.py'>
17:36:57-090553 INFO     Torch: torch==2.6.0+cu118 torchvision==0.21.0+cu118
17:36:57-097165 INFO     Packages: diffusers==0.34.0.dev0 transformers==4.51.3 accelerate==1.6.0 gradio==3.43.2 pydantic==1.10.21
17:36:58-767012 DEBUG    ONNX: version=1.21.1, available=['AzureExecutionProvider', 'CPUExecutionProvider']
17:36:59-525761 DEBUG    Settings: fn="C:\Users\User\Desktop\SDNext\sdnext-dev\sdnext\config.json" created
17:36:59-526766 INFO     Engine: backend=Backend.DIFFUSERS compute=cpu device=cpu attention="Scaled-Dot-Product" mode=no_grad
17:36:59-526766 DEBUG    Save: file="C:\Users\User\Desktop\SDNext\sdnext-dev\sdnext\config.json" json=0 bytes=2 time=0.000
17:36:59-530097 DEBUG    Read: file="html\reference.json" json=77 bytes=39555 time=0.003 fn=_call_with_frames_removed:<module>
17:36:59-530097 DEBUG    Torch attention: type="sdpa" flash=True memory=True math=True
17:36:59-536777 INFO     Torch parameters: backend=cpu device=cpu config=Auto dtype=torch.bfloat16 context=no_grad nohalf=False nohalfvae=False upcast=False deterministic=False
                         tunable=[False, True] fp16=pass bf16=pass optimization="Scaled-Dot-Product"
17:36:59-538522 INFO     Device:
17:37:00-582999 DEBUG    Entering start sequence

First thing I noticed was ROCM detected both my 7900 XT and my integrated graphics card, as you can see in these lines:

17:36:07-616162 INFO ROCm: AMD toolkit detected 17:36:08-276185 INFO ROCm: agents=['gfx1100', 'gfx1036'] 17:36:08-276185 INFO ROCm: version=6.2, using agent gfx1100

however here:

17:36:59-526766 INFO Engine: backend=Backend.DIFFUSERS compute=cpu device=cpu attention="Scaled-Dot-Product" mode=no_grad

and in the line right before the end:

17:36:59-538522 INFO Device:

I can see that SDNext is not able to utilize my GPU and falls back to the CPU. I'm unfortunately not really sure what the problem here seems to be - I saw a couple of threads with ZLUDA failing to load, causing SDNext to utilize the CPU instead, but my problem seems to be a little different.

This is a screenshot of my "System Info" tab, which again shows SDNext utilizing my CPU instead:

I really appreciate any help in solving this issue. Thank you!

If more information is needed, I would be happy to supply it.

Version Platform Description

Windows 11, Python 3.11.9, ROCM 6.2, SDNext 55b1cb8 (dev-branch)

Relevant log output

Backend

Diffusers

Compute

ZLUDA

UI

Standard

Branch

Dev

Model

Other

Acknowledgements

• I have read the above and searched for existing issues
• I confirm that this is classified correctly and its not an extension issue

[milanboers]

Same problem since today. It looks like a driver update (AMD 25.3.1 / 32.0.21001.9024) is the culprit. After rolling back the driver (to 32.0.13031.3015 in my case), gpu is in use again.

[Kolopsel]

Same problem since today. It looks like a driver update (AMD 25.3.1 / 32.0.21001.9024) is the culprit. After rolling back the driver (to 32.0.13031.3015 in my case), gpu is in use again.

That makes a lot of sense! I also tested an installation of Stable Diffusion WebUI AMD Forge on my system and it also suddenly broke like clockwork. Seems like a rollback might be the only viable option right now, thanks for the quick reply!

[mjp9523]

I'd just like to confirm that a driver rollback worked for me also, 25.2.1 works. I'm new to using zluda but after working through a few other issues and errors I had a similar situation where just the cpu was being used. Driver rollback then allowed the GPU to start working. Really crazy how much faster zluda is over directml. I'm using a rx6800.

[mikharju]

For me it only stopped working when updating to 25.5.1 and using system restore point to get back to 25.3.1 works again. (I'm on Win11, 7900 XTX)

[0xDELUXA]

Yes guys the driver rollback works but we need a solution for the newest driver because we cant use 25.4.1 or earlier for eternity

[EchoesEqu]

Согласен, нужен фикс.

[Flaxee]

There was a commit yesterday in ZLUDA repo that should fix this issue.

I tried to remove the .zluda folder inside SD.Next to force a upgrade of ZLUDA, but it didn't help for me. I assume this doesn't download ZLUDA from the upstream master branch?

[0xDELUXA]

I've found a workaround. Download Ishqqytiger's zluda fork, the latest version, v3.9.5: zluda-windows-rocm6(or rocm5, which one you have already installed)-amd64.zip, and extract it to sd.next's .zluda directory replacing the files there. It will work.

[EchoesEqu]

Yes, indeed if you replace the files everything will work - https://github.com/lshqqytiger/ZLUDA/releases
At least for now it works.

[vladmandic]

zluda 3.9.5 is in sdnext dev branch.
note: delete .zluda folder so that newest zluda will be installed if you are using the latest AMD Adrenaline driver

this update will be merged to master release this week.

[EchoesEqu]

Launching Web UI with arguments: --use-zluda --skip-torch-cuda-test
E:\STB\stable-diffusion-webui-amdgpu-forge-main\venv\lib\site-packages\torch\cuda_init_.py:118: UserWarning: CUDA initialization: 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. (Triggered internally at ..\c10\cuda\CUDAFunctions.cpp:108.)
return torch._C.cuda_getDeviceCount() > 0
Traceback (most recent call last):
File "E:\STB\stable-diffusion-webui-amdgpu-forge-main\launch.py", line 54, in
main()
File "E:\STB\stable-diffusion-webui-amdgpu-forge-main\launch.py", line 50, in main
start()
File "E:\STB\stable-diffusion-webui-amdgpu-forge-main\modules\launch_utils.py", line 677, in start
import webui
File "E:\STB\stable-diffusion-webui-amdgpu-forge-main\webui.py", line 21, in
initialize_forge()
File "E:\STB\stable-diffusion-webui-amdgpu-forge-main\modules_forge\initialization.py", line 58, in initialize_forge
from backend import memory_management
File "E:\STB\stable-diffusion-webui-amdgpu-forge-main\backend\memory_management.py", line 133, in
total_vram = get_total_memory(get_torch_device()) / (1024 * 1024)
File "E:\STB\stable-diffusion-webui-amdgpu-forge-main\backend\memory_management.py", line 100, in get_torch_device
return torch.device(torch.cuda.current_device())
File "E:\STB\stable-diffusion-webui-amdgpu-forge-main\venv\lib\site-packages\torch\cuda_init.py", line 778, in current_device
lazy_init()
File "E:\STB\stable-diffusion-webui-amdgpu-forge-main\venv\lib\site-packages\torch\cuda_init.py", line 293, in _lazy_init
torch._C._cuda_init()
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.

[vladmandic]

this is not stable-diffusion-webui-amdgpu-forge, lets stay on topic.