Any clues to actually get ROCm working on Windows at this point?

[valfaun]

Hi, so, i tried using the ROCm backend again after the new 2025-11-06 update, using the --use-rocm --reinstall flags, and image generation failed 3/3 times, each in a different way:

1. During the first attempt image generation just halted halfway through and the server had seemingly frozen, not taking any input whatsoever. Performance up to that point looked noticeably faster than ZLUDA, though.
2. The second attempt threw an error about some invalid value somewhere, which i didn't bother taking note of as it didn't seem to be reproducible anyway.
3. And the third attempt at generating an image essentially killed all video output from the system and i had to hard reset it. I went back to ZLUDA afterwards.

I have an RX 9070 XT and HIP/ROCm 6.4.2 installed. The installed driver is the "AMD Software: PRO Edition" 25.Q3.1, which the AMD Install Manager at some point has apparently decided was better than the Adrenalin Edition that i chose to keep when installing HIP, but whatever, i guess.

The SDnext documentation/wiki still seems quite barren regarding "ROCm on Windows", probably because nothing's official yet, so i'm wondering if anyone got it working and if so, how? Which steps are necessary to prepare the system and which settings in SDnext are either required or critically incompatible?

I noticed in the application console that it had referenced ROCm 7.0 during startup, which i guess i don't have as it's not available on the HIP SDK download page yet. I've also seen some kinda "PyTorch on Windows" preview driver that was released a while ago, but i honestly don't understand what exactly that's for since the support page at the download location links back to ROCm on Linux information and also instructions on how to install PyTorch, which i never manually did to run SDnext

[vladmandic]

best i can say is to check out our discord community, there is a very active rocm channel talking about (daily improvements to ) rocm-for-windows.
https://discord.com/invite/VjvR2tabEX

[valfaun]

thank you for the reply and link, I'll keep it in mind if necessary.
I'd prefer to try and solve this here first, so that others might find the information too. Not your fault of course, but that's the huge downside to relocating such things to Discord servers now. They can't be indexed and so none of the stuff written in those is visible from outside

[vladmandic]

true - and nothing replaces good docs. i'm just saying that community is dealing with nightly builds, miopen issues, hip vs driver version issues, and other stuff which would be impossible to catch for daily updates. once it stabilizes, for sure we need to a) make it all work out-of-the-box, b) update docs.

[klu1982]

Hello, AMD announce Windows PyTorch Preview on September 24 and I have been using ROCm version staring October.
Officially, AMD HIP SDK is still at 6.4.2. ROCm on Windows is still 6.4.4.
But we do not need to track these because SDNext will install or update them for you automatically.
I was using ZLUDA before october, my steps to migrate to ROCm was as below.

1. I was using Python 3.11 for ZLUDA. So, I installed Python 3.12 based on SD.Next wiki.
2. I completed remove driver with DDU and installed 25.20.01.14 driver. This is the official Windows PyTorch Preview driver.
   https://www.amd.com/en/resources/support-articles/release-notes/RN-AMDGPU-WINDOWS-PYTORCH-PREVIEW.html
3. Git clone and rest is the same as wiki.
4. And that's it. I can generate image without any issue.

[valfaun]

hmm, maybe that particular driver is absolutely necessary then, whatever it does, since I already have Python 3.12.10 installed. I still hardly understand what ROCm, HIP, CUDA, etc even are or why everything runs on Python.

Thank you for the instructions and background info. Does that preview driver seem stable for general use, too? Did you notice any regressions in performance when gaming, if you do any of that, or other undesired quirks?

[klu1982]

I have using that driver playing Battlefield 6 around 100 hours so far.
One thing I just realized I was wrong is that HIP/Torch did not update.
So, I just create seperate virtual environment and did the new installation to compare my stable one.

1. Stable version is running SD.Next 11.8 dev / ROCm 7.2 / HIP 6.4 / Torch 2.8. Speed similar to my ZLUDA. I can run 100 batch generation with HighRes (Force HiRes option) without problem.
2. Newest insallation is running SD.Next 11.15 dev / ROCm 7.2 / HIP 7.2 / Touch 2.9. Really fast like 3 times faster than Stable version. I did 25 batch test run without problem. However, HighRes does not working yet. I bump into driver crash and SD.Next give me no error and generate pictures in full noise.

So, these are my current state on running Windows ROCm SD.Next.

[valfaun]

well, i've installed the driver, reinstalled the HIP SDK, reinstalled Python 3.12.10, switched to the dev branch, deleted the venv folder multiple times, rebooted a few times inbetween those steps but i'm just getting this once again entirely new error message everytime i attempt to launch SDnext:

i've also followed the instructions here:
https://rocm.docs.amd.com/projects/radeon-ryzen/en/latest/docs/install/installrad/windows/install-pytorch.html
but still getting the same error. i even cloned a fresh copy of SDnext in my user folder on the C: drive, just in case the ROCm thing doesnt play well across multiple drives, but same result.
i dont suppose you've run into that error and fixed it, have you?