Step-by-step guide to run on RDNA3 (Linux)

[wsippel]

AMD is keeping awfully quiet, but I somehow stumbled across a ROCm 5.5 release candidate Docker container that works properly on 7900XT/ 7900XTX cards - but you have to also compile PyTorch yourself. So here's a hopefully correct step-by-step guide, from memory:

mkdir $HOME/docker_sd
docker pull rocm/composable_kernel:ck_ub20.04_rocm5.5_rc4
docker run --name rocm -it -v $HOME/docker_sd:/SD --device=/dev/kfd --device=/dev/dri --security-opt seccomp=unconfined --group-add video rocm/composable_kernel:ck_ub20.04_rocm5.5_rc4
cd /SD
git clone https://github.com/AUTOMATIC1111/stable-diffusion-webui
apt update && apt install python3.8-venv
cd stable-diffusion-webui
python3 -m venv venv
source venv/bin/activate
python3 -m pip install --upgrade pip wheel
export HIP_VISIBLE_DEVICES=0
export PYTORCH_ROCM_ARCH="gfx1100"
export CMAKE_PREFIX_PATH=/SD/stable-diffusion-webui/venv/
pip install -r requirements.txt
pip remove torch torchvision
mkdir repositories
cd repositories
wget https://github.com/pytorch/pytorch/releases/download/v2.0.0/pytorch-v2.0.0.tar.gz
wget https://github.com/pytorch/vision/archive/refs/tags/v0.15.1.tar.gz
tar -xzvf pytorch-v2.0.0.tar.gz && cd pytorch-v2.0.0
pip install cmake ninja
pip install -r requirements.txt
pip install mkl mkl-include
python3 tools/amd_build/build_amd.py
python3 setup.py install
cd ..
tar -xzvf v0.15.1.tar.gz
cd vision-0.15.1
python3 setup.py install
cd /SD/stable-diffusion-webui
python3 launch.py --listen

You might not need HIP_VISIBLE_DEVICES and PYTORCH_ROCM_ARCH unless you have another unsupported AMD GPU in your machine - the integrated GPU in Zen 4 CPUs for example.

[collsni]

Dude, you totally rock. I really appreciate all the time you obviously spent on this. Thank you for sharing, worked the first try. 7900xtx SD1.5

Pretty sure I heated my house up a few degrees with that compile.

[wsippel]

Glad I could help! And yeah, compiling Torch takes a hot minute. Not as bad as installing gentoo back in the day on a single core machine, but still. And we only have to compile for one target. I don't envy the Arch maintainers who'll have to compile Torch for nine targets once ROCm 5.5 officially releases (it's six right now, ROCm 5.5 adds another three)...

[ZhenyaPav]

Tried to follow your steps, launching SD fails with the following message:

python3 launch.py --listen
Python 3.8.10 (default, Mar 13 2023, 10:26:41) 
[GCC 9.4.0]
Commit hash: 22bcc7be428c94e9408f589966c2040187245d81
Installing requirements for Web UI
Launching Web UI with arguments: --listen
/SD/stable-diffusion-webui/venv/lib/python3.8/site-packages/torchvision/io/image.py:13: UserWarning: Failed to load image Python extension: 'libc10_cuda.so: cannot open shared object file: No such file or directory'If you don't plan on using image functionality from `torchvision.io`, you can ignore this warning. Otherwise, there might be something wrong with your environment. Did you have `libjpeg` or `libpng` installed before building `torchvision` from source?
  warn(
Traceback (most recent call last):
  File "/SD/stable-diffusion-webui/venv/lib/python3.8/site-packages/transformers/utils/import_utils.py", line 1093, in _get_module
    return importlib.import_module("." + module_name, self.__name__)
  File "/usr/lib/python3.8/importlib/__init__.py", line 127, in import_module
    return _bootstrap._gcd_import(name[level:], package, level)
  File "<frozen importlib._bootstrap>", line 1014, in _gcd_import
  File "<frozen importlib._bootstrap>", line 991, in _find_and_load
  File "<frozen importlib._bootstrap>", line 975, in _find_and_load_unlocked
  File "<frozen importlib._bootstrap>", line 671, in _load_unlocked
  File "<frozen importlib._bootstrap_external>", line 848, in exec_module
  File "<frozen importlib._bootstrap>", line 219, in _call_with_frames_removed
  File "/SD/stable-diffusion-webui/venv/lib/python3.8/site-packages/transformers/modeling_utils.py", line 78, in <module>
    from accelerate import __version__ as accelerate_version
ImportError: cannot import name '__version__' from 'accelerate' (/SD/stable-diffusion-webui/venv/lib/python3.8/site-packages/accelerate/__init__.py)

The above exception was the direct cause of the following exception:

Traceback (most recent call last):
  File "launch.py", line 356, in <module>
    start()
  File "launch.py", line 347, in start
    import webui
  File "/SD/stable-diffusion-webui/webui.py", line 21, in <module>
    import pytorch_lightning # pytorch_lightning should be imported after torch, but it re-enables warnings on import so import once to disable them
  File "/SD/stable-diffusion-webui/venv/lib/python3.8/site-packages/pytorch_lightning/__init__.py", line 35, in <module>
    from pytorch_lightning.callbacks import Callback  # noqa: E402
  File "/SD/stable-diffusion-webui/venv/lib/python3.8/site-packages/pytorch_lightning/callbacks/__init__.py", line 14, in <module>
    from pytorch_lightning.callbacks.batch_size_finder import BatchSizeFinder
  File "/SD/stable-diffusion-webui/venv/lib/python3.8/site-packages/pytorch_lightning/callbacks/batch_size_finder.py", line 24, in <module>
    from pytorch_lightning.callbacks.callback import Callback
  File "/SD/stable-diffusion-webui/venv/lib/python3.8/site-packages/pytorch_lightning/callbacks/callback.py", line 25, in <module>
    from pytorch_lightning.utilities.types import STEP_OUTPUT
  File "/SD/stable-diffusion-webui/venv/lib/python3.8/site-packages/pytorch_lightning/utilities/types.py", line 27, in <module>
    from torchmetrics import Metric
  File "/SD/stable-diffusion-webui/venv/lib/python3.8/site-packages/torchmetrics/__init__.py", line 14, in <module>
    from torchmetrics import functional  # noqa: E402
  File "/SD/stable-diffusion-webui/venv/lib/python3.8/site-packages/torchmetrics/functional/__init__.py", line 82, in <module>
    from torchmetrics.functional.text.bleu import bleu_score
  File "/SD/stable-diffusion-webui/venv/lib/python3.8/site-packages/torchmetrics/functional/text/__init__.py", line 30, in <module>
    from torchmetrics.functional.text.bert import bert_score  # noqa: F401
  File "/SD/stable-diffusion-webui/venv/lib/python3.8/site-packages/torchmetrics/functional/text/bert.py", line 24, in <module>
    from torchmetrics.functional.text.helper_embedding_metric import (
  File "/SD/stable-diffusion-webui/venv/lib/python3.8/site-packages/torchmetrics/functional/text/helper_embedding_metric.py", line 26, in <module>
    from transformers import AutoModelForMaskedLM, AutoTokenizer, PreTrainedModel, PreTrainedTokenizerBase
  File "<frozen importlib._bootstrap>", line 1039, in _handle_fromlist
  File "/SD/stable-diffusion-webui/venv/lib/python3.8/site-packages/transformers/utils/import_utils.py", line 1083, in __getattr__
    module = self._get_module(self._class_to_module[name])
  File "/SD/stable-diffusion-webui/venv/lib/python3.8/site-packages/transformers/utils/import_utils.py", line 1095, in _get_module
    raise RuntimeError(
RuntimeError: Failed to import transformers.modeling_utils because of the following error (look up to see its traceback):
cannot import name '__version__' from 'accelerate' (/SD/stable-diffusion-webui/venv/lib/python3.8/site-packages/accelerate/__init__.py)

[Beinsezii]

Try unpinning the accelerate version from requirements_versions.txt then updating it with pip.

[wsippel]

UserWarning: Failed to load image Python extension: 'libc10_cuda.so: cannot open shared object file: No such file or directory looks very wrong. We're on AMD, of course there's no CUDA library. Are you sure you didn't accidentally skip a step or two?

[Beinsezii]

there's no CUDA library

If you install auto1111's requirements first before installing Torch/Vision it grabs the cuda versions from pypi. Probably torch+rocm didnt get built & installed so its still running the cuda version.

Side note, you can omit fetching requirements.txt for auto1111 since launch.py reads requirements_versions.txt anyways. That way the stock Torch from pypi and its dependencies wont get installed just to be immediately overwritten

[wsippel]

Now that you mention it, I think I did pip remove torch torchvision before installing my own compiles, but as I wrote, those instructions were written from memory, and I wasn't sure if it was necessary.

[In-line]

@wsippel

Hi, can you publish the container somewhere? It seems AMD deleted it.

❯ docker pull rocm/composable_kernel:ck_ub20.04_rocm5.5_rc4
Error response from daemon: manifest for rocm/composable_kernel:ck_ub20.04_rocm5.5_rc4 not found: manifest unknown: manifest unknown

[mvasilchuk]

Torrent: https://mega.nz/file/EUQAFLjb#KhklqKIdDd85KazEdsQdL0avWVutpFKoyrXLOncFOOE
I will seed for the next few hours or until a few people download it.

[In-line]

@mvasilchuk You need to add a tracker to the torrent. It's a bit hard to find peers without it.

[mvasilchuk]

It already has udp://tracker.opentrackr.org:1337/announce

[Beinsezii]

Github didnt like me pushing a massive tarball so I put it directly on Mega. It's compressed with xz to get around the 5gb download limit mega has.

https://mega.nz/folder/yldQjRwb#PUFz7DD5BBpwt6GxWmcicQ

[wszgrcy]

maybe we can build image by https://github.com/ROCmSoftwarePlatform/composable_kernel/tree/develop?

[In-line]

Was someone able to come up with a working Dockerfile that does everything?

[In-line]

Docker + Docker Compose setup with ready to use SD UI: https://gist.github.com/In-line/c1225f05d5164a4be9b39de68e99ee2b

[ZhenyaPav]

This one works for me, thank you!

[Dnak-jb]

I do wish i was a smart enough nerd to be able to run things like this. I managed to install a fresh copy of lunar lobster nightly and get docker to give me a hello world successfully, but beyond that I have no idea how to build/execute this stuff :( I do have the rocm 5.5 RC but I have no idea where to place it, whether or not to keep it zipped, or if i have to do more.

docker run --name rocm -it -v $HOME/docker_sd:/SD --device=/dev/kfd --device=/dev/dri --security-opt seccomp=unconfined --group-add video rocm/composable_kernel:ck_ub20.04_rocm5.5_rc4
As soon as I try to do this, everything falls apart.. I need someone to hold my hand and teach me the ways!

[ZhenyaPav]

Thy this
https://gist.github.com/In-line/c1225f05d5164a4be9b39de68e99ee2b
I am not sure how to install the docker image though tbh

[xGigaa]

docker load -i '/path/to/rocm5.5.tar.gz' should do the trick.
Then you can just "up" the compose file like normal.

[PiotrKlimczak]

You don't need docker if you don't want to.
Got it working natively without docker:

1. Install ubuntu 20.04
2. Install AMDGPU PRO driver from AMD website
3. Restart ubuntu to pick up new driver
4. Install Anaconda from the net
5. $ conda create -n anaconda38 python=3.8
6. $ conda activate anaconda38
7. $ export HIP_VISIBLE_DEVICES=0
8. $ export PYTORCH_ROCM_ARCH="gfx1100"
9. pull "rocm/composable_kernel:ck_ub20.04_rocm5.5_rc5" docker image and start it (unfortunately not avaialble anymore as they keep deleting them)
10. Copy /opt/rocm* from docker image to your local /opt
11. Install PyTorch 2.0.0 as per https://github.com/pytorch/pytorch#from-source to your conda env

I was unable to get PyCharm working with docker image, while natively + conda it is fully working- thus motivation for this work.
Please note that I was lucky to grab 5.5rc5 image.
Everything seems to be working perfectly fine, including torchvision.
Tested on PyTorch examples.

Hope that helps.
Piotr.

[Sp3ci3s8472]

The reason you might still want docker is to keep your system clean, no other reason for me.

[PennyFranklin]

it's said that the current speed 15~16it/s it same as the 3080's speed, and I think 7900xtx would be at a higher speed if optimization fine

[Marcelo5444]

Have you tested other networks different than stable diff?

[Sp3ci3s8472]

See the following benchmarks
https://www.tomshardware.com/news/stable-diffusion-gpu-benchmarks
I thought I was getting 18-21 it/s on my 7900XTX, but I would need to recheck that to be sure.

[tmode-1960]

rocm/composable_kernel:ck_ub20.04_rocm5.5_release is available at dockerhub

[In-line]

It seems to currently segfault

[Sp3ci3s8472]

I tried the release version today and it indeed seems to segfault. Does the rocm5.5_rc4 not have this problem? Since I will try that one next then (if I can find the image....)

[Spooky309]

I tried the release version today and it indeed seems to segfault. Does the rocm5.5_rc4 not have this problem? Since I will try that one next then (if I can find the image....)

Segfaults here too, did you manage to find an rc4 image? All the links I can find are dead.

[mvasilchuk]

rc4 image
magnet:?xt=urn:btih:dc9b3ec573f4deb765952ddfa764b1d9e9512c6f&xt=urn:btmh:12203daea6d9ea069933e19403d4050463b3f014c034237a2934cbd01f8aa8f75b36&dn=rocm5.5.tar.gz&tr=udp%3a%2f%2ftracker.dler.org%3a6969%2fannounce&tr=udp%3a%2f%2ftracker.openbittorrent.com%3a6969%2fannounce&tr=http%3a%2f%2ftracker.openbittorrent.com%3a80%2fannounce&tr=https%3a%2f%2fopentracker.i2p.rocks%3a443%2fannounce&tr=udp%3a%2f%2f9.rarbg.com%3a2810%2fannounce&tr=udp%3a%2f%2ftracker.opentrackr.org%3a1337%2fannounce&tr=udp%3a%2f%2fexplodie.org%3a6969%2fannounce&tr=udp%3a%2f%2ftracker.torrent.eu.org%3a451%2fannounce&tr=udp%3a%2f%2fopen.stealth.si%3a80%2fannounce&tr=udp%3a%2f%2ftracker.tiny-vps.com%3a6969%2fannounce&tr=udp%3a%2f%2fopen.demonii.com%3a1337%2fannounce&tr=udp%3a%2f%2ftracker.moeking.me%3a6969%2fannounce&tr=https%3a%2f%2ftracker.lilithraws.org%3a443%2fannounce

[Sp3ci3s8472]

I did find a working image, it has some minor things like no VAE. Took me a while to find it because google search is nothing compared to how good it was 5 years ago.
https://gist.github.com/In-line/c1225f05d5164a4be9b39de68e99ee2b
and
https://evshiron.github.io/are-we-gfx1100-yet/post/a1111-webui/

This seems to be with the 5.5 release version.
Just noticed that the gist is from In-Line the user I replied to lol.

[idkwhodatis]

worked on my native Ubuntu 22.04.02, thanks a lot

[yangvw]

I'm getting segfault in Docker too, then I turned to native Ubuntu and it worked fine. Pointer: https://are-we-gfx1100-yet.github.io/post/a1111-webui/

[Scarecrow928]

not works when using rocm/composable_kernel:ck_ub20.04_rocm5.5_release

Launching Web UI with arguments: -f --precision full --no-half --skip-torch-cuda-test
terminate called after throwing an instance of 'c10::Error'
  what():  Tried to register an operator (image::decode_png(Tensor _0, int _1, bool _2) -> Tensor _0) with the same name and overload name multiple times. Each overload's schema should only be registered with a single call to def(). Duplicate registration: registered by RegisterOperators. Original registration: registered by RegisterOperators
Exception raised from registerDef at /stable-diffusion-webui/tmp/pytorch-v2.0.0/aten/src/ATen/core/dispatch/Dispatcher.cpp:207 (most recent call first):
......
frame #53: python3() [0x5c4ef0]

Aborted (core dumped)

[sunng87]

I'm getting same error with torchvision. This error happens when import torchvision

[wsippel]

There's no reason to use the Docker containers anymore, this guide was for ROCm 5.5 release candidates and is completely outdated ever since 5.5 was released to the public. We're at 5.6 now, with 5.7 around the corner.

[adrian369]

I want to run pytorch with my 7900xtx on my windows ubuntu 22.04.3, would following what was talked about do the job?

[adrian369]

I want to run pytorch with my 7900xtx on my windows ubuntu 22.04.3, would following what was talked about do the job?*

[fulcrumEFX]

I tried following this using (K)Ubuntu 23.10. I used the docker image for rocm5.7.

When trying to launch SD I get the following error:

Python 3.8.10 (default, May 26 2023, 14:05:08) 
[GCC 9.4.0]
Version: v1.6.0
Commit hash: 5ef669de080814067961f28357256e8fe27544f4
Launching Web UI with arguments: --listen
no module 'xformers'. Processing without...
No SDP backend available, likely because you are running in pytorch versions < 2.0. In fact, you are using PyTorch 2.0.0a0+git5ef669d. You might want to consider upgrading.
no module 'xformers'. Processing without...
Traceback (most recent call last):
  File "launch.py", line 48, in <module>
    main()
  File "launch.py", line 44, in main
    start()
  File "/SD/stable-diffusion-webui/modules/launch_utils.py", line 432, in start
    import webui
  File "/SD/stable-diffusion-webui/webui.py", line 13, in <module>
    initialize.imports()
  File "/SD/stable-diffusion-webui/modules/initialize.py", line 33, in imports
    from modules import shared_init
  File "/SD/stable-diffusion-webui/modules/shared_init.py", line 5, in <module>
    from modules import shared
  File "/SD/stable-diffusion-webui/modules/shared.py", line 5, in <module>
    from modules import shared_cmd_options, shared_gradio_themes, options, shared_items, sd_models_types
  File "/SD/stable-diffusion-webui/modules/options.py", line 74, in <module>
    class Options:
  File "/SD/stable-diffusion-webui/modules/options.py", line 77, in Options
    def __init__(self, data_labels: dict[str, OptionInfo], restricted_opts):
TypeError: 'type' object is not subscriptable

Does anyone know how to fix this?

[lukeab]

This is down to the dependencies of the webui project not supporting older pyton, you need 3.10 of some point release, check the requirements.

[Aipathon]

I tried this on Arch Linux and got 2 Errors:
1: ERROR: Could not open requirements file: [Errno 2] File or Directory not found: 'requirements.txt'
2:File "/home/******/stable-diffusion-webui/modules/launch_utils.py", line 384, in prepare_environment
raise RuntimeError(
RuntimeError: Torch is not able to use GPU; add --skip-torch-cuda-test to COMMANDLINE_ARGS variable to disable this check

How can i fix those two errors? I did changed the apt commands to pacman. changing the python3.8-env package(which isnt available on arch) to one of the 3 statet in the wiki(https://wiki.archlinux.org/title/Python/Virtual_environment) didnt change anything.