Radeon 5700xt Queries

[LostAndDead]

To be clear I have yet to try anything, I have just been reading and looking around.

So I have a Radeon 5700xt (50th anniversary edition) and wish to try and set this up. After looking around I see many threads with many different issues and fixes for Radeon cards but I'm yet to just find an answer or guide.

I see many people able to get it to work on the CPU but they all comment that it takes forever on the CPU so I would very much like it working with the GPU.

I am running Windows 11/Ubuntu 22.04 (Dual boot) so a method for either will work.
Just hopping someone can point me in the right direction or give me a small guide. I'm very new to AI and this kind of stuff so don't expect me to know too much.

Thanks in advance,

[Sakura-Luna]

Refer to this https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Install-and-Run-on-AMD-GPUs.

I recommend that users who don't know anything about Python try to use a packaged version from a trusted source to avoid a lot of problems.

[LostAndDead]

I want to comment and say I also have a 5700XT and have been running web-ui on Ubuntu 22.04 for months with no issues. I get 6its/sec on a standard 512x512 image.

I have it working now on Windows 11, for me a 512x512 runs at about 8-11its/sec which is odd that it differs so much. Do you have simular issues with upscalling and larger Images?

You'll be able to enjoy it when PyTorch and the project are updated.

Ill just chill around and keep an eye out then.

[Soulreaver90]

Interesting, I run mine in linux but using Docker, I like to keep my main system clean and run it in its own environment. I wonder if running in docker slows it down a bit. And I can run 1280x720 or 1024x1024, depending on the model or configs, somtimes it craps out at 1024. I tend to just run everything at 768 and upscale.

[LostAndDead]

Interesting, I run mine in linux but using Docker, I like to keep my main system clean and run it in its own environment. I wonder if running in docker slows it down a bit. And I can run 1280x720 or 1024x1024, depending on the model or configs, somtimes it craps out at 1024. I tend to just run everything at 768 and upscale.

I would guess the docker would have some impact but not that much, what startup command line arguments are you using?

[Soulreaver90]

Just --no-half-vae

[Rainz3]

I want to comment and say I also have a 5700XT and have been running web-ui on Ubuntu 22.04 for months with no issues. I get 6its/sec on a standard 512x512 image.

I have it working now on Windows 11, for me a 512x512 runs at about 8-11its/sec which is odd that it differs so much. Do you have simular issues with upscalling and larger Images?

You'll be able to enjoy it when PyTorch and the project are updated.

Ill just chill around and keep an eye out then.

Sampling method has a big hit on speed. Euler a I get 1.3it/s, DPM SDE ++ I get like 2s/it. I can't use high res fix at all and I can't go above 768x768. Also, I can only use openpose and canny in controlnet. Everything else either runs out of memory or errors out

[Sakura-Luna]

PyTorch just released 2.0, now you can update it and WebUI to use --opt-sdp-attention to experience improvements in speed and VRAM usage. Limited by your VRAM size, the saved VRAM may not be significant, and running on Linux is still the optimal solution.

[Sakura-Luna]

It may just overflow the VRAM under some influence, it is difficult to compare different environments.

[Soulreaver90]

Very true. Anyway, I am content with my current setup. If you hear of any improvements with AMD, do let me know.

[Sakura-Luna]

If I notice that PyTorch has updated Flash Attention support, I'll post it.

[iDeNoh]

Is it possible to update to torch 2.0 on windows yet? or are we just talking on linux?

[Sakura-Luna]

@iDeNoh The ROCm version for AMD GPUs is not available for Windows, Nvidia GPUs do not have this limitation.

[moophlo]

I built an hassle free docker image to run AUTOMATIC1111 on RX5700XT on linux (I guess the docker compose can be adjusted for Windows but I didn't test it)
Check it out here: https://hub.docker.com/r/moophlo/stable-diffusion

[MDBrothers]

Thanks for positing this, I want to give it a try as soon as I can. When you say to place a model, you mean a weights file?

[moophlo]

After your comment I updated the docker hub page that was not up to date. Now you can manage username and password via environment variables. Also now is using the version 1.10 of SD and the latest pytorch (2.1.0) possible for rocm 5.2 .
you can use your own command line arguments using the environment variable COMMANDLINE_ARGS . If you find any issue please drop me a DM

P.S.: about models, once you start the container the first time a folder structure is created in the models directory, You'll use the subdirectories according to what you're downloading (LoRA, checkpoint, ...)

[m-aljammali]

Hi moophlo,

I am running into an issue.
The command in Dockerfile "ADD run.sh ."
fails with " "/run.sh": not found"

Do you have any idea why ? if you could explain to me what its purpose I might be able to figure out what is going wrong.
thanks

[moophlo]

Is going bad that I'm a forgetful person and so I forgot to add to the instructions the content of the run.sh . Allow me a couple of hours and I'll update the instructions on docker hub

EDIT: The description was updated to include the run.sh file