mkdir -p ~/SD
cd ~/SD
docker build -t sd-mrocm-5-6 .
docker-compose up -d
docker logs -f sd-mrocm-5-6
docker start sd-mrocm-5-6
The web UI will pull things in the background and may act... Odd the first few times you run things. WAIT for a while after a click, watch your GUI and network load to see what's up. Sometimes just hitting F5 to reload the page will fix things.
| version: "3.9" | |
| services: | |
| sd: | |
| image: sd-mrocm-5-6 | |
| container_name: sd-mrocm-5-6 | |
| ports: | |
| - "7860:7860" | |
| volumes: | |
| - ./models:/SD/stable-diffusion-webui/models/ | |
| - ./repositories:/SD/stable-diffusion-webui/repositories/ | |
| - ./extensions:/SD/stable-diffusion-webui/extensions/ | |
| - ./outputs:/SD/stable-diffusion-webui/outputs/ | |
| devices: | |
| - '/dev/kfd:/dev/kfd' | |
| - '/dev/dri:/dev/dri' | |
| security_opt: | |
| - seccomp:unconfined | |
| group_add: | |
| - video |
| #docker run -ti rocm/composable_kernel:ck_ub20.04_rocm5.6 bash | |
| FROM rocm/composable_kernel:ck_ub20.04_rocm5.6 | |
| RUN mkdir /SD | |
| # Clone SD | |
| WORKDIR /SD | |
| RUN git clone https://github.com/AUTOMATIC1111/stable-diffusion-webui | |
| WORKDIR /SD/stable-diffusion-webui | |
| RUN git reset --hard 22bcc7be428c94e9408f589966c2040187245d81 | |
| RUN apt update && apt install -y python3.8-venv | |
| RUN python3 -m venv venv | |
| # Activate VENV | |
| ENV VIRTUAL_ENV=/SD/stable-diffusion-webui/venv | |
| RUN python3 -m venv $VIRTUAL_ENV | |
| ENV PATH="$VIRTUAL_ENV/bin:$PATH" | |
| RUN python3 -m pip install --upgrade pip wheel | |
| ENV HIP_VISIBLE_DEVICES=0 | |
| ENV PYTORCH_ROCM_ARCH="gfx1100" | |
| ENV CMAKE_PREFIX_PATH=/SD/stable-diffusion-webui/venv/ | |
| ENV USE_CUDA=0 | |
| # Remove old torch and torchvision | |
| RUN pip uninstall -y torch torchvision | |
| # pytorch | |
| RUN pip3 install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm5.5 | |
| WORKDIR /SD/stable-diffusion-webui | |
| # Patch requirements.txt to remove torch | |
| RUN sed '/torch/d' requirements.txt | |
| RUN pip install -r requirements.txt | |
| EXPOSE 7860/tcp | |
| # Fix for "detected dubious ownership in repository" by rom1win. | |
| RUN git config --global --add safe.directory '*' | |
| CMD python3 launch.py --listen --disable-safe-unpickle --skip-torch-cuda-test | |
Oh, no, I was talking about automatic1111 being an older build. It was as simple as I thought to change though. I just replaced the git reset hard command hash to the newest build they have. Then i compiled a new docker image with that hash and new name in another folder.
As for not being able to get to 1024 without issue, that is strange. With your provided dockerfiles i was able to do 1024 easily around 16it/s. Highres fix went about 2.5it/sec. It was only eating about 17 to 18gb of vram doing so. When i upgraded to my new docker image i was able to use the invokeai optimizations as opposed to the default doggetx on your dockerfiles and that allowed me to go all the way to 2048 image size with about 4s/it.
Edit : just tested with invokeai optimizations and I was able to batch 8 x 1024 images at once. It used 16gb during both the highres fix portion and first pass. Totat time was 1:01 with facefix. 2.07it/s during regular and 3.55s/it during highres fix. Using dpm++2m karras at 50 steps normal and 10 for highres.
One thing i did do though was stop firefox from using hardware acceleration to prevent it from eating vram for the session. But thats it.
Unfortunately I get 😭
Error response from daemon: error gathering device information while adding custom device "/dev/kfd": no such file or directory
You can also make variations by adjusting the composer file and have different versions at the same time. I think you might be able to run at least 2 at the same time with an XTX. Probably not on the XT.
I tried doing a single 1024x1024 image on my XTX and it used so much vRAM it killed my desktop session. It ended up working, I just lost my login.