this post was submitted on 16 Mar 2024
259 points (74.1% liked)
linuxmemes
20880 readers
4 users here now
I use Arch btw
Sister communities:
- LemmyMemes: Memes
- LemmyShitpost: Anything and everything goes.
- RISA: Star Trek memes and shitposts
Community rules
- Follow the site-wide rules and code of conduct
- Be civil
- Post Linux-related content
- No recent reposts
Please report posts and comments that break these rules!
founded 1 year ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
I run Stable Diffusion with ROCm. Who needs CUDA?
What distro are you using? Been looking for an excuse to strain my 6900XT.
I started looking at getting it running on Void and it seemed like (at the time) there were a lot of specific version dependencies that made it awkward.
I suspect the right answer is to spin up a container, but I resent Docker's licensing BS too much for that. Surely by now there'd be a purpose built live image- write it to a flash drive, reboot, and boom, anime ~~vampire princes~~ hot girls
If you don’t like docker take a look at containerd and podman. I haven’t done any cuda with podman but it is supposed to work
Pay us if you're a mid+ sized compan is BS?
I think people don't like dramatic changes in business model. I had installed it for like 3 days, long before the switchover, to test out something from another dev. When they made the announcements, the hammer went down in our org not to use it. But that didn't stop them from sending sales-prospecting/vaguely threateningly worded email to me, who has no cheque-writing authority anyway.
Plus, I'm not a fan of containers.
STOP DOING CONTAINERS.
docker compose up meatball-hero
of something. Please give me Alpine Linux On Musl of it" -- Statements dreamed up by the utterly deranged."Hello, I would like 7.5GB of VM worth of apples please"
THEY HAVE PLAYED US FOR ABSOLUTE FOOLS.
Poor capitalists need to pay for the tools that make them money. Stop it, I'll break down in tears just thinking about the horror of it.
Do you use some different solution, or did you completely avoid containers and orchestration?
They only spam me with promotional material. You used the business email I'm guessing?
Ubuntu native for me..no containers needed.
Arch ofc.
I set mine up on arch. There's an aur package, but it didn't work for me.
After some failed attempts, I ended up having success following this guide.
Some parts are out of date though, so if it fails to install something you'll need to have it target a newer available package. Main example of this is inside the
webui-user.sh
file, it tells you to replace an existing line withexport TORCH_COMMAND="pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/rocm5.1.1"
. This will fail because that version of pytorch is no longer available. So instead you need to replace the download URL with an up to date one from the pytorch website. They've also slightly changed the layout of the file. Right now the correct edit should be to find the# install command for torch
line and change the command under it to:pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm5.7
You may need to swap pip to pip3 too, if you get a pip error. Overall it takes some troubleshooting, look at any errors you get and see if it's calling for a package you don't have or anything like that.
I use stable diffusion on rocm in an ubuntu distrobox container. Super easy to set up and there's a good guide in the opensuse forum for it.
That is exactly what I do too and it works perfectly! This is a link to said guide.
It's effectively install distrobox, save the config, run
distrobox assemble
and thendistrobox enter rocm
and clone the Automatic1111 stable diffusion webui somewhere and runbash webui.sh
to launch it.I use Fedora. It works great, with some tweaks to the startup script.
I can confirm that it works just fine for me. In my case I'm on Arch Linux btw and a 7900XTX, but it needed a few tweaks:
- Having xformers installed at all would sometimes break startup of stable-diffusion depending on the fork
- I had an internal and an external GPU, I want to set HIP_VISIBLE_DEVICE so that it only sees the correct one
- I had to update torch/torchvision and set HSA_OVERRIDE_GFX_VERSION
I threw what I did into https://github.com/icedream/sd-multiverse/blob/main/scripts/setup-venv.sh#L381-L386 to test several forks.
Then show us your anime titty pics!
CUDA?! I barely even know'a!