Zluda with Polaris GPUs (RX470, RX570, RX580)

[TheFerumn]

Thanks to the https://github.com/advanced-lvl-up user we have possibility now to use Zluda without memory leaks with old Polaris GPU's.

To make it work you have to install HIP SDK 5.7.1: https://www.amd.com/en/developer/resources/rocm-hub/hip-sdk.html

After installation download: https://github.com/advanced-lvl-up/Rx470-Vega10-Rx580-gfx803-gfx900-fix-AMD-GPU/releases/download/v1.0.0/rocm.gfx800-gfx900-for.hip.sdk.5.7.1-and-6.2.4.7z

From unpacked .7z file put rocblas.dll to where HIP SDK is installed. Usually in: c:/program files/AMD/ROCm/5.7/bin/
Then replace the library folder or files in c:/program files/AMD/ROCM/5.7/bin/rocblas/library

Restart PC

Follow standard installation of Stable Diffusion:
git clone https://github.com/lshqqytiger/stable-diffusion-webui-amdgpu-forge
cd stable-diffusion-webui-amdgpu-forge

Don't close cmd window yet. Download this version of ZLUDA: https://github.com/lshqqytiger/ZLUDA/releases/download/rel.5e717459179dc272b7d7d23391f0fad66c7459cf/ZLUDA-windows-rocm5-amd64.zip
And swap files in stable-diffusion-webui-amdgpu-forge/.zluda
(I am not sure if swapping ZLUDA files are required or it will choose correct version after detecting version of HIP SDK)

Go back to the cmd and run:
.\webui.bat --use-zluda --debug

After everything will install close CMD window and reopen it again in stable-diffusion-webui-amdgpu-forge/venv/scripts
Type activate to activate your virtual environment and follow commands:

pip uninstall torch torchvision torchaudio -y
pip install --force-reinstall torch==2.2.1 torchvision==0.17.1 torchaudio==2.2.1 --index-url https://download.pytorch.org/whl/cu118
pip install --force-reinstall numpy==1.26.4
pip install --force-reinstall transformers==4.51.3 safetensors==0.5.3

And its almost finished. Last step is to add some quality of life commands into your webui-user.bat
Edit it with notepad or other text editor and in the set COMMANDLINE_ARGS= add --all-in-fp16 --always-low-vram --use-zluda --cuda-stream --attention-quad --skip-ort --disable-gpu-warning
Use --always-low-vram only with 4GB VRAM GPU. I doubt you need it with 8 GB versions

Enjoy around 20% faster generations comparing to DML

Edit: I forget to mention that you can finally use Inpainting in FP16. No more --full precision or --no-half commands !!