wiki/ZLUDA.md

ZLUDA (CUDA Wrapper) for AMD GPUs in Windows

Warning

ZLUDA does not fully support PyTorch in its official build. So ZLUDA support is so tricky and unstable. Support is limited at this time. Please don't create issues regarding ZLUDA on github. Feel free to reach out via the ZLUDA thread in the help channel on discord.

Installing ZLUDA for AMD GPUs in Windows.

Note

This guide assumes you have git and python installed, have used SD.Next before, and are comfortable using the command prompt, navigating Windows Explorer, renaming files and folders, and working with zip files.

Compatible GPUs

A list of compatible GPUs can be found here. If your GPU is not on the list, you may need to build your own roclabs,please follow the instructions in Rocm Support guide.
(Note: including intergrated GPUs)

Note: If you have an integrated GPU (iGPU), you may need to disable it, or use the HIP_VISIBLE_DEVICES environment variable. Learn more here.

Install Visual C++ Runtime

Note: Most everyone would have this anyway, since it comes with a lot of games, but there's no harm in trying to install it.

Grab the latest version of Visual C++ Runtime from https://aka.ms/vs/17/release/vc_redist.x64.exe (this is a direct download link) and then run it.
If you get the options to Repair or Uninstall, then you already have it installed and can click Close. Otherwise, install it.

Install ZLUDA

ZLUDA is now auto-installed, and automatically added to PATH, when starting webui.bat with --use-zluda.

Install HIP SDK

Install HIP SDK 5.7 from https://www.amd.com/en/developer/resources/rocm-hub/hip-sdk.html
So long as your regular AMD GPU driver is up to date, you don't need to install the PRO driver HIP SDK suggests.

Add folders to PATH

Add %HIP_PATH%bin to your PATH.
https://github.com/brknsoul/ROCmLibs/wiki/Adding-folders-to-PATH

Note: %HIP_PATH%bin typically relates to "C:\Program Files\AMD\ROCm\5.7\bin", assuming Windows is installed on C:.

Replace HIP SDK library files for GPU architectures gfx1031 and gfx1032

Go to https://rocm.docs.amd.com/projects/install-on-windows/en/develop/reference/system-requirements.html and find your GPU model.
If your GPU model has a ✅ in both columns then skip to Compilation, Settings, and First Generation.
If your GPU model has an ❌ in the HIP SDK column (LLVM targets gfx1031 and gfx1032) follow the instructions below;

1. Go to %HIP_PATH%bin\rocblas
2. Rename library to something else, like origlibrary
3. Download ROCmLibs.zip
   i. Alternative: If you have a 6600 or 6600xt (gfx1032) GPU, give Optimised_ROCmLibs_gfx1032.7z a go. It seems to be about 50% faster. Thanks FremontDango!
4. Open the zip file.
5. Drag and drop the library folder from ROCmLibs.zip into %HIP_PATH%bin\rocblas (The folder you opened in step 1).
6. Reboot PC

If your GPU model not in the HIP SDK column or not availalbe in the above list, follow the instructions in Rocm Support guide to build your own RocblasLibs;

Install or Update SD.Next

Install SD.Next;

git clone https://github.com/vladmandic/automatic
then
cd automatic
then
webui.bat --use-zluda --debug --autolaunch

or Update SD.Next

(from a current SD.Next install folder)
venv\scripts\activate
pip uninstall -y torch-directml torch
deactivate
git pull
webui.bat --use-zluda --debug --autolaunch --reinstall
(after running successfully once, you can remove --reinstall)

Note: ZLUDA functions best in Diffusers Backend, where certain Diffusers-only options are available

Compilation, Settings, and First Generation

After the UI starts, head on over to System Tab > Compute Settings
Set "Attention optimization method" to "Dynamic Attention BMM"
Now, try to generate something.
This should take a fair while (10-15mins, or even longer; some reports state over an hour) to compile, but this compilation should only need to be done once.
Note: There will be no progress bar, as this is done by ZLUDA and not SD.Next. Eventually your image will start generating.

Comparison (DirectML)

	DirectML	ZLUDA
Speed	Slower	Faster
VRAM usage	More	Less
VRAM GC	❌	✅
Traning	*	✅
Flash Attention	❌	❌
FFT	❓	✅
FFTW	❓	❌
DNN	❓	🚧
RTC	❓	❌
Source Code	Closed	Opened
Python	<=3.10	Same as CUDA

*: Known as possible, but uses too much VRAM to train stable diffusion models/LoRAs/etc.

Compatibility

DTYPE
FP64	✅
FP32	✅
FP16	✅
BF16	✅
LONG	✅
INT8	✅*
UINT8	✅*
INT4	❓
FP8	❌
BF8	❌

*: Not tested.

---

Experimental Settings

Sections below are optional and highly experimental, and aren't required to start generating images. Ensure you can generate images first before trying these.

Experimental Speed Increase Using deepcache (optional)

Start SD.Next, head on over to System Tab > Compute Settings.
Scroll down to "Model Compile" and tick the 'Model', 'VAE', and 'Text Encoder' boxes.
Select "deep-cache" as your Model compile backend.
Apply and Shutdown, and restart SD.Next.

Enabling ZLUDA DNN (partial support)

This is PARTIAL and INCOMPLETE support of a performance library, ZLUDA DNN. Most of the cases, trying this will waste your time.

1. Checkout dev branch.
2. Download ZLUDA.zip from v3.7-pre5-dnn, and unpack it upon your ZLUDA folder. (default: path/to/sd.next/.zluda)
3. Replace path/to/sd.next/venv/Lib/site-packages/torch/lib/cudnn64_8.dll with path/to/sd.next/.zluda/cudnn.dll. (rename cudnn.dll and overwrite torch/lib/cudnn64_8.dll)
4. Download 5.7.zip from the same release, and unpack it upon your HIP SDK folder. (%HIP_PATH%)
5. Tick Enable ZLUDA DNN, and restart webui.

Simple Comparison

RX 7900 XTX, not a optimal settings.

• without DNN: 16.66 ~ 16.9it/s
• with DNN: 17.1 ~ 17.4it/s

---

If you get a "ZLUDA: failed to automatically patch torch" error;

• Manually download the ZLUDA release marked with a green "latest" tag from https://github.com/lshqqytiger/ZLUDA/releases/
• Unzip it somewhere, like C:\ZLUDA
• Add C:\ZLUDA to your PATH following this guide.
  • If there's a ZLUDA path there already, be sure to remove it, before closing the dialog box
• Manually copy and rename the following files from C:\ZLUDA to [SD.Next Install Folder]\venv\Lib\site-packages\torch\lib, overwriting the originals;
  cublas.dll -> cublas64_11.dll
  cusparse.dll -> cusparse64_11.dll
  nvrtc.dll -> nvrtc64_112_0.dll