ZLUDA Crashing on torch Torch attention status

[coodinii]

Sometimes it's crashing python 3.12.0 completely. Using an RX6900xt. This worked before but I was silly and tried to update and broke the whole thing. I can't figure out what I'm doing wrong. Was using python 3.10.6 on the old version, not sure if that's the problem, I uninstalled that one because it was giving me some pytorch error after updating it. The new docs seem to recommend using python 3.12. Is my python version bad?

It simply loads up to "2025-12-07 00:59:56,308 | DESKTOP-REB3D51 | sd | DEBUG | devices | Torch attention status: flash=False flash3=False aiter=False sage=False flex=True npu=False xla=False xformers=False" then sometimes it tells me "press any key to continue", sometimes python crashes, sometimes it just closes the window, sometimes it goes to "press any key to continue" and then goes to the next line allowing me to type but the server isn't running.

The steps I followed to install are from here https://vladmandic.github.io/sdnext-docs/ZLUDA/#install-sdnext:
git clone https://github.com/vladmandic/sdnext
cd sdnext
.\webui.bat --use-zluda --debug --autolaunch

I did install the hip sdk 6.2.4 for windows and it lists my gpu as supported from what I see.

sdnext.log

If this just doesn't work are there any alternatives similar in performance? Would directml be a better attempt and more compatible? I've tried installing without the zluda argument it launches but when I press generate nothing happens. It's acting like it's doing something and then just vanishes and gives me updates once in a while that the site instance is running.

Thanks for any help I've been struggling for a month to get this working but I cannot find the problem. I've tried tens of tutorials and ways to get this working, I don't know how to read the log to understand why it doesn't work.

[vladmandic]

everything you said suggests that python itself crashed and in thar case, reading sdnext log will not show much - it cannot log if there is no python.

check windows event log for reason of crash.

[coodinii]

Hey thanks for answering. I went ahead and checked the event log, but sometimes it doesn't crash however so there's no log. It just goes to the press any key to continue and python doesn't crash.

"12:19:58-315947 DEBUG Torch attention installed: flashattn=False sageattention=False
12:19:58-317948 DEBUG Torch attention status: flash=False flash3=False aiter=False sage=False flex=True npu=False
xla=False xformers=False
Press any key to continue . . ."

In the event log for the error crash I found this:
"Faulting application name: python.exe, version: 3.12.10150.1013, time stamp: 0x67f515a7
Faulting module name: amdhip64.dll, version: 10.0.3592.0, time stamp: 0x6567627f
Exception code: 0xc0000005"

Is this indicative of the hip sdk crashing for whatever reason? I've used the same steps before to install it and it worked when I was using python 3.10.6, should I attempt a newer version of the hip sdk or would you say that's incompatible. Like would python 14 work with this UI, or hip sdk 6.4.2?

I also went ahead and upgraded python to 3.12.2 which is supposedly the last exe release of the 3.12 series but this didn't change anything.

Lastly I went ahead and installed the ROCm version which launched and it appeared to work and I assumed it was the first time it took longer but it took 4 hours or more to generate one image. It generated at 1iteration every 760s. Once done the image was there and I attempted a second generation however the result appears to go the exact same direction so I stopped it.

With the older version I used it would load the model and sometimes compile the model if I was switching them in about 3-10 minutes, and generate an image every 30 seconds maybe at default settings. For reference I was finally able to install comfyUI using DirectML and that loaded and image and generated one in about 1-2 minutes I think. However I am failing their install and most tutorials use RDNA3. I'd prefer to use zluda if I can get it working again because it's much faster.

Would you have any recommendations as to where I'm maybe going wrong? Like it worked before on python 3.10.6 but something changed. Could it be my graphics driver all of a sudden? I am running a 5800x3D so there should be no iGPU either.

PS: I'll attach a log of the ROCm version I trried.
sdnext.log

I found "sd | ERROR | shared | ROCm initialization failed: Expected a cuda device, but got: cpu" on line 60. Is it using my cpu to generate?
On line 21 it seems to have found my 6900xt or is this something else? "sd | INFO | installer | Torch detected: gpu="AMD Radeon RX 6900 XT" vram=16368 arch=(10, 3) cores=40".

[vladmandic]

but sometimes it doesn't crash however so there's no log. It just goes to the press any key to continue and python doesn't crash.

this mean exactly that process crashed.

Faulting module name: amdhip64.dll, version: 10.0.3592.0, time stamp: 0x6567627f

yes, this means that hip sdk crashed

Is this indicative of the hip sdk crashing for whatever reason? I've used the same steps before to install it and it worked when I was using python 3.10.6, should I attempt a newer version of the hip sdk or would you say that's incompatible. Like would python 14 work with this UI, or hip sdk 6.4.2?

I also went ahead and upgraded python to 3.12.2 which is supposedly the last exe release of the 3.12 series but this didn't change anything.

python 3.13 is not recommended. python 3.14 is not supported at all.
python 3.12 is recommended.
this is fully documented.
which version of hip sdk is recommended is unfortunately very dependend on gpu model. best to discuss with amd gpu users on sdnext discord server. i wish that amd get its dev in order, there is no reason why there is such a version nightmare.

I found "sd | ERROR | shared | ROCm initialization failed: Expected a cuda device, but got: cpu" on line 60. Is it using my cpu to generate?

everything else you wrote indicates its using cpu because gpu failed to initialize so yes, it will be really slow. there is no point of even looking at cpu generation numbers.

[coodinii]

Awesome I managed to find the sliders. Not sure if this is a bug or intended but in the video generator the output is displayed in the models tab for me, not in the prompts tab which would be similar as to how it is for the image generator.

One last question for now, is it possible to have it save the prompt as a .txt file with the file name which shows the prompts just in case I want to return to them?

[vladmandic]

next to a prompt you have small buttons, one of them is save style - it saves it as json.

[coodinii]

Was able to do it thanks.

Seems my limit on a 6900xt with 32gb of ram is the wan 2.1 1.3b model. Anything else seems to crash. Is there any documentation or source you know off that shows what models people are running on their machines with provided hardware specs, so I can plan out my upgrades?

[vladmandic]

are you using quantization? sdnq uint4 (with or without svd) would help a lot with resource utilization.

[coodinii]

are you using quantization? sdnq uint4 (with or without svd) would help a lot with resource utilization.

Hello, I apologize for necroing this I've been away for a while unable to spend more time on this, but I don't know if by quantization models you mean "WAN 2.2 A14B SDNQ T2V https://huggingface.co/Disty0/Wan2.2-T2V-A14B-SDNQ-uint4-svd-r32" or if there's better options out there that I should be using with a RX6900XT(16gb vram)? Is there some way for me to find what's best for my gpu? For the settings I'm using the default I believe 832x480 but I do see in the example on huggingface they use 1288x720.