How to Install Stable Diffusion on AMD GPUs (NEW)

I have got SD installed without an issue but issue start when loading model.

It was too slow for my laptop AMD 5500u. After once image was created it failed to create it again because GPU is not left the chat.

Lenovo laptop which i've 2GB reserve ram for it funny thing is it didn't worked at all with 2GB dedicated ram of gpu but it worked with when limited to 650MB without and issue exclude(full gpu ram usage after image done).

GPU, CPU usage was lower than 3% but RAM full usage, HDD full usage (100%) 😂 when it goes to GPU it was ready fast less than 5 min.

When i set dedicated ram of GPU which was 2GB by default it return Black Images and after generate failed to create it again.

I already had 90% of this tutorial done on my PC. The part that changed it from CPU to GPU was the commandline_args for me. I spent days on this. Thank you sir

Awesome! Worked. Image does take 2m on average but it works.

will try this. Usually it said runtime error because of my gpu. ill reinstall using this guide

It works on my AMD GPU. Thank you very much.

You are a god thanks

Hey dude i used to render images using my CPU but that took around 20 min per image, with this guide i was able to use my AMD GPU and it renders an image in 2 minutes so it was really helpful. but i got a problem after rendering 4-5 images screen goes black GPU fans go ballistic and it dosent work back until i hard reset the case. I can still listen background so i know rest of the pc is working well only GPU is doing this any idea what mightve causing it?

The music used in your video is Better Days, Chill Day, and at least one more song that may have been made by LAKEY INSPIRED. I don't know what it is about this man's music, but nobody gives him credit, and to be honest, I'm kinda sick of it. If you use somebody's music, especially if they allow you to use their music for free, and all they ask is that you give them credit, give them credit.

Can you teach us how to install RVC on AMD GPUs?

I followed the instructions, but get the following error message: RuntimeError: "LayerNormKernelImpl" not implemented for 'Half'
It is still looking for an Nvidia driver apparently.

I had problems with pytorch installing, pytorch will not install on the latest python so you need have python v3.10.6
if you have multiple pythons installed go to venv folder edit pyvenv.cfg and replace:Home, Executable and command with your python 310 location.
C:\Users\USERNAME\AppData\Local\Programs\Python\Python310\python.exe

Absolute chad, thanks!
I normally draw, but i think using AI will help me generate images for my D&D campaign much faster and more aesthetically than what I'm capable of doing on a weekly basis :)))))

Thanks for the video. I am looking at it 4 months after you posted and it's still good. I got it to work using my RX 480 with 8GC VRAM and the argument: --medvram --precision full --no-half --no-half-vae --opt-split-attention-v1 --opt-sub-quad-attention --disable-nan-check . Try to figure out max resolution. Render times are barrable. I am currently using PlagroundAI. Thanks again for a great video.

your link in the description is shortened, i cant copy it

Every time i click generate its not generating anything

Couldn't launch python - exit code 9009
When executing the web-user.bat file

☺👍

The text to copy is wrong in the description. The ... messes it up.

ERROR: Could not find a version that satisfies the requirement torch-directml (from versions: none)
ERROR: No matching distribution found for torch-directml

I’ve been using Happy Diffusion for a while now, and it’s been a game-changer for my image generation work.

So this works for me to an extent. 6800xt, I can get up to a 1080 size file. But further upscaling, even at 1.1 upscale size eventually fails. So looking like I can create images, but may need to find a different solution for upscaling. First day trying though, hopefully I'm just doing something wrong and I'll figure that out.

I should be loading up images lighting fast but my RX 7900 is no faster than my 1070.

It runs putting -skip-torch-cuda-test and --no half, but it´s painfully slow, EXTREMELY slow compared to my last card. I had a 6GB GTX 970 TI. Now I have a RX 6950 XT with 16 GB of ram and it´s like 5 times slower than before.

Please help! when I try to run "webui-user.bat" it closes instantly. I´ve tried running as admin, reinstalling everything, moving the folder to another disks, nothing works.

Finallt got it working thenks to you, good video. Also if some of you are getting trouble whit the text, copy it straigh from the link in the dresciption

Thank you bro

Didn't work. Tells me I don't using NVIDIA drivers (really? wow! didn't know) and rendering with CPU instead of GPU

ModuleNotFoundError: No module named 'basicsr.version' :c

Thanks! Very happy to share with you my joy!

if you add " --autolaunch" will launch the browser for you when loaded, and you have a great tutorial but with the error pointed out before, you need Python 3.10.6 for all the things (torch) in SD to work, and also you did towards the end remembere about the model about 3 mins after, but you forgot to tell people what folder to put the model into at that point, so some could have left the video after that, just for the future so it's more step by step you know, and let's be real you are not the only one with tut for AMD. But anyways, Thank you ;)

Only 1 problem and it states it clearly on the page "(Newer version of Python does not support torch) and you need torch. Python and SD wil still run but without it.

Thank you very much. It worked like a charm.

Where do I put the medium vram?

How can you train your own pics

I have a 4Gb AMD Video card, but this screen appears: RuntimeError: Could not allocate tensor with 1347760 bytes. There is not enough GPU video memory available!

I'm getting a gpu memory error when generating above 512 x 512 on a 7900 xtx.

It won't work with anything but 3.10 version of python, so if you get errors during the installation this is the cause.

it doesn't seem like you can use Lora type checkpoints. Am I missing something?

Okay, so after much troubleshooting, I finally got mine to at least START to generate an image. However, ultimately it fails to return an image and gives me the following message below where the image would appear. This is what it shows 🡻

NotImplementedError: No operator found for `memory_efficient_attention_forward` with inputs: query : shape=(1, 4096, 1, 512) (torch.float32) key : shape=(1, 4096, 1, 512) (torch.float32) value : shape=(1, 4096, 1, 512) (torch.float32) attn_bias : <class 'NoneType'> p : 0.0 `cutlassF` is not supported because: device=cpu (supported: {'cuda'}) `flshattF` is not supported because: device=cpu (supported: {'cuda'}) dtype=torch.float32 (supported: {torch.float16, torch.bfloat16}) max(query.shape[-1] != value.shape[-1]) > 128 `tritonflashattF` is not supported because: device=cpu (supported: {'cuda'}) dtype=torch.float32 (supported: {torch.float16, torch.bfloat16}) max(query.shape[-1] != value.shape[-1]) > 128 Operator wasn't built - see `python -m xformers.info` for more info triton is not available `smallkF` is not supported because: max(query.shape[-1] != value.shape[-1]) > 32 unsupported embed per head: 512
Time taken: 6 min. 9.2 sec.


FYI, my graphics card is an AMD Radion (TM) RX 480 It's 8 bit and running at 59.940 Hz (whatever any of that means, because I'm certainly no tech guru)

In the arguments, here's what I have pasted in there. set COMMANDLINE_ARGS= --xformers --autolaunch --medvram --skip-torch-cuda-test --precision full --no-half
The reason for the "--precision full --no-half" is because I kept getting RuntimeError: "LayerNormKernelImpl" not implemented for 'Half'
Since pasting those in the arguments, I can finally at least start to generate an image, but of course it fails as mentioned above.
Does anyone have a fix for this? Please and thank you.

In the beginning it says you cannot train stable diffusion with amd, is that right? Id like to teach it certain styles :(

Excellent upload, I got it working, on rx580 8 gbram, takes about 1minite to generate 512 x 512, image, now i just have to learn all about the interface. thanks for the upload liked subed,

Now How do we add control net?

Its not working, it says


import diffusers
ModuleNotFoundError: No module named 'diffusers'

" ModuleNotFoundError: No module named 'diffusers' " while trying the installation

While this works (sort of) You are barely scratching the surface of the capabilities of the GPU. In order to get the full benefits, you need ROCM. (barely supports windows ATM) So figure out how to get Docker running with ROCM and stable diffusion. That's the only way at the moment to actually have stab;e diffusion, AMD GPU at it's full potential

you forgot to paste the bat edit

suport 4gb video card?