AMD ROCm under WINDOWS Status Update. ZLUDA with SD.next as the best alternative (Tutorial).

What is the status of AMD ROCm on Windows - especially with regard to Stable Diffusion?
We install SD.next with ZLUDA to accelerate Stable Diffusion and bridge the waiting time for ROCm on Windows.

If you want to run ComfyUI with ZLUDA, too, watch this video:

PyTorch/ROCm status and GitHub Windows issues:

Links regarding ZLUDA and SD.next installation:

Videos:
AMD ROCm on Linux:
AMD ROCm on Windows Status Details and GIT & MiniConda-Installation:

Chapters:
0:00 About ROCm on Windows and ZLUDA
1:34 Status Update ROCm
2:46 Zluda Preparation
4:13 HIP SDK
7:14 SD.next Installation
7:56 Example Generation

[NextTechandAI]

UPDATE: In case of trouble, check if you have an integrated GPU. Disabling it in your BIOS might help. Additionally, if you have trouble with latest ROCm, try using ROCm HIP 5.71 with an older ZLUDA version as I did in this video.
---
What will you do?
1. Give ZLUDA/SD.next a try?
2. Wait for AMD ROCm on Windows?
3. Use Linux?
4. Else...?

[kassper]

SUBBED. I've been trying to get this local version of SD.NEXT running on my 7800X3D/7900XTX setup since last night, and this video is the one that got me there. Every little thing absolutely mattered and no other help on YT was this complete. Last step was to disable the iGPU in the BIOS and BOOM...compiling. Thank you man! You rock.

[rudolfaeschlimann]

Thank you for this informative video, you helped me a lot with the setup!

[Bucky_Bailey]

Excellent video! I went from 8s/it on normal to 6+it/s on ZLUDA. There is however some features that don't work, or work inconsistently, but generally, image generation is a breeze and up-scaling works well.

[isaacmiller]

I'm running an old potato PC with RX 580, don't even ask about the CPU I have, lol. But anyway, your video is one of only 3/4 I found directly talking about ROCm for old AMD GPU's, on Windows. Honestly felt like I was trapped in a box. Thank you for the video and I hope the updates continue. I subscribed immediately

[heloiamerk]

Thanks! Got it working on my 6800XT with Python 3.10.11, lshqqytiger ZLUDA v3.8.1, HIP SDK ROCm v6.1.2, and the dev branch of SD.next.

[alanreynolds]

Hello, when I try to generate an image I get this error: Building PyTorch extensions using ROCm and Windows is not supported. I have followed everything perfectly. Could it be because I have an RX 7900 GRE, which isn't listed at all?

[taipeiperson]

Thanks, you It's work for my RX 6700 XT.

[duladrop]

Quick question currently most of benchmarking application or any application right now when running Radeon they are using HIP on it. to my knowledge but correct me if I'm wrong, is just an approach to map wave instruction to warp(CUDA) so that it can work with their applications. So if ROCm is available on windows, Will those application that's so dependent on CUDA, can use ROCm now instead of doing HIP-CUDA instruction translation?

[CapaUno]

Thanks for this video, I have AMD in my machine which I am very happy with and I would like very much for the AI stuff to work as best as, it doesn't have to be as fast as my GPU is 20% slower than a rtx3090 but near 50% less expensive! 😅 I have Stable Diffusion running though it could be faster it works and so I'm sure I can make it work more efficiently as I have seen someone with the same card doing just that, so eh thanks again! Have a fun day! ;D

[pinkpig]

I am getting the following error :
Error(displayed on SD.Next site): model not loaded
Time: 0.02s
does this mean zluda is not working? this is my laptop model - AMD Ryzen 7 7730U with Radeon Graphics.
error displayed on terminal: ERROR sd shared Reading failed: ui-config.json [Errno 2] No such file or directory: 'ui-config.json'

[manaphylv]

When you say the first generation takes a long time, do you mean only the very first run after installation, or the first image in every session?
In other words, if I restart the program or my PC after the installation and initial run, do I have to wait it out again?

[lowrider]

First of all thank you very much for your videos and willingness to help with amd video cards. i followed the instructions and i managed to run SD.next. i used HIP SDK 6.1.2 and rocm gfx1031 for hip sdk 6.1.2 optimized.7z (library) since i have a radeon 6750xt. What I noticed is if I select 512x512 with 20 steps and default settings the picture is generated 3-4 seconds. however with 1024x1024 it takes 2.5 minutes.
So I decided to continue using Fooocus which I use with direct-ml, in it 1024x1024 at 30 steps also takes 2.5 minutes and for entertainment purposes to generate a wallpaper or something like that.. i'm fine with it.
Thanks anyway

[Vindexlegamer]

Hello, I have a 7800XT and the ia charges to infinity for a small image...

When I put to the strict minimum, it takes about 2 minutes but it's blurry

[RealKatTea]

I keep getting "Building PyTorch extensions using ROCm and Windows is not supported", anybody else?

[crk]

I've tried it several times now and I always get the same error :( "OS Error: [WinError 126] The specified module could not be found. Error loading or one of it's dependencies." Does anyone know what I can do?

[FluxDigital]

I got everything to work, and generated some images using a model I got on Civit, but what is the Refiner Model option next to the Base model and why don’t I have it?

[ДмитрийЗелень-щщ]

I would like to clarify, am I correct that with this build, every time we change the prompt the first image generation will be extremely long? Or is this true only for one, the first prompt? I'm just impressed by the speed of generation of subsequent prompts

[Киночат-гг]

Updated SD.next today. It worked fine before on 7800 xt with ZLUDA. Now it says: "OSError: Building PyTorch extensions using ROCm and Windows is not supported." Do you know by any chance what the matter is, how to fix it?

[chriswalt]

I've been struggling a lot for the past week now trying to get this or literally ANY AI generation tool to work for my AMD RX 6850m XT on windows. Do I need to use torch with Zluda? also does your video actually apply to my case or do i just have no hope getting this to work?