Skip to content

AMD-INSTALLATION.md

Latest commit

 

History

History
275 lines (189 loc) · 9.13 KB

AMD-INSTALLATION.md

File metadata and controls

275 lines (189 loc) · 9.13 KB

AMD Installation Guide for Windows (TheRock)

This guide covers installation for AMD GPUs running under Windows using TheRock's official PyTorch wheels.

Supported GPUs

Based on TheRock's official support matrix, the following GPUs are supported on Windows:

gfx110X-all (RDNA 3):

  • AMD RX 7900 XTX (gfx1100)
  • AMD RX 7800 XT (gfx1101)
  • AMD RX 7700 XT (gfx1101)
  • AMD RX 7700S / Framework Laptop 16 (gfx1102)
  • AMD Radeon 780M Laptop iGPU (gfx1103)

gfx120X-all (RDNA 4):

  • AMD RX 9060 XT (gfx1200)
  • AMD RX 9060 (gfx1200)
  • AMD RX 9070 XT (gfx1201)
  • AMD RX 9070 (gfx1201)

gfx1151 (RDNA 3.5 APU):

  • AMD Strix Halo APUs

gfx1150 (RDNA 3.5 APU):

  • AMD Radeon 890M (Ryzen AI 9 HX 370 - Strix Point)

Also supported:

gfx103X-dgpu: (RDNA 2)


Note: If your GPU is not listed above, it may not be supported by TheRock on Windows. Support status and future updates can be found in the official documentation.

Requirements

  • Python 3.11 (recommended for Wan2GP - TheRock currently supports Python 3.11, 3.12, and 3.13).
  • Windows 10/11

Installation Environment

This installation uses PyTorch wheels built by TheRock.

Installing Python

Download Python 3.11 from python.org/downloads/windows. Press Ctrl+F and search for "3.11." to find the newest version available for installation.

Alternatively, you can use this direct link: Python 3.11.9 (64-bit).

After installing, make sure python --version works in your terminal and returns 3.11.9

If it doesn’t, you need to add Python to your PATH:

  • Press the Windows key, type Environment Variables, and select Edit the system environment variables.
  • In the System Properties window, click Environment Variables….
  • Under User variables, find Path, then click Edit → New and add the following entries (replace <username> with your Windows username):
C:\Users\<username>\AppData\Local\Programs\Python\Launcher\
C:\Users\<username>\AppData\Local\Programs\Python\Python311\Scripts\
C:\Users\<username>\AppData\Local\Programs\Python\Python311\

Note: If Python still doesn't show the correct version after updating PATH, try signing out and signing back in to Windows to apply the changes.

Installing Git

Download Git from git-scm.com/downloads/windows and install it. The default installation options are fine.

Installation Steps (Windows, using a Python venv)

Note: The following commands are intended for use in the Windows Command Prompt (CMD).
If you are using PowerShell, some commands (like comments and activating the virtual environment) may differ.

Step 1: Download and set up Wan2GP Environment

:: Navigate to your desired install directory
cd \your-path-to-wan2gp

:: Clone the repository
git clone https://github.com/deepbeepmeep/Wan2GP.git
cd Wan2GP

:: Create virtual environment
python -m venv wan2gp-env

:: Activate the virtual environment
wan2gp-env\Scripts\activate

Note: If you have multiple versions of Python installed, use py -3.11 -m venv wan2gp-env instead of python -m venv wan2gp-env to ensure the correct version is used.

Step 2: Install ROCm/PyTorch by TheRock

IMPORTANT: Choose the correct index URL for your GPU family!

For gfx110X-all (RX 7900 XTX, RX 7800 XT, etc.):

pip install --pre torch torchaudio torchvision rocm[devel] --index-url https://rocm.nightlies.amd.com/v2/gfx110X-all/

For gfx120X-all (RX 9060, RX 9070, etc.):

pip install --pre torch torchaudio torchvision rocm[devel] --index-url https://rocm.nightlies.amd.com/v2/gfx120X-all/

For gfx1151 (Strix Halo iGPU):

pip install --pre torch torchaudio torchvision rocm[devel] --index-url https://rocm.nightlies.amd.com/v2/gfx1151/

For gfx1150 (Radeon 890M - Strix Point):

pip install --pre torch torchaudio torchvision rocm[devel] --index-url https://rocm.nightlies.amd.com/v2-staging/gfx1150/

For gfx103X-dgpu (RDNA 2):

pip install --pre torch torchaudio torchvision rocm[devel] --index-url https://rocm.nightlies.amd.com/v2-staging/gfx103X-dgpu/

This will automatically install the latest PyTorch, torchaudio, and torchvision wheels with ROCm support.

Step 3: Install Wan2GP Dependencies

:: Install core dependencies
pip install -r requirements.txt

Step 4: Verify Installation

python -c "import torch; print('PyTorch:', torch.__version__); print('ROCm available:', torch.cuda.is_available()); print('Device:', torch.cuda.get_device_name(0) if torch.cuda.is_available() else 'No GPU')"

Expected output example:

PyTorch: 2.11.0+rocm7.12.0
ROCm available: True
Device: AMD Radeon RX 9070 XT

Attention Modes

WanGP supports multiple attention implementations via triton-windows.

First, install triton-windows in your virtual environment. If you have an older version of Triton installed, uninstall it first. ROCm SDK needs to be initialized. Visual Studio environment should also be activated.

pip uninstall triton
pip install triton-windows
rocm-sdk init
"C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Auxiliary\Build\vcvars64.bat" >nul 2>&1

Supported attention implementations

  • SageAttention V1 (Requires the .post26 wheel or newer to fix Triton compilation issues without needing unofficial patches. Download it from this URL)
pip install "sageattention <2"
  • FlashAttention-2 (Only the Triton backend is supported):
git clone https://github.com/Dao-AILab/flash-attention.git
cd flash-attention
pip install ninja
pip install packaging
set FLASH_ATTENTION_TRITON_AMD_ENABLE=TRUE && python setup.py install
  • SDPA Flash: Available by default in PyTorch on post-RDNA2 GPUs via AOTriton.

Running Wan2GP

For future sessions, activate the environment every time if it isn't already activated, then run python wgp.py:

cd \path-to\Wan2GP
wan2gp-env\Scripts\activate
:: Add the AMD-specific environment variables mentioned below here
python wgp.py

It is advised to set the following environment variables at the start of every new session (you can create a .bat file that activates your venv, sets these, then launches wgp.py):

set ROCM_HOME=%ROCM_ROOT%
set PATH=%ROCM_ROOT%\lib\llvm\bin;%ROCM_BIN%;%PATH%
set CC=clang-cl
set CXX=clang-cl
set DISTUTILS_USE_SDK=1
set FLASH_ATTENTION_TRITON_AMD_ENABLE=TRUE
set TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1

MIOpen (AMD’s equivalent of NVIDIA’s cuDNN) is not yet fully stable on several architectures; it can cause out-of-memory errors (OOMs), crash the display driver, or significantly increase generation times. Currently, it is recommended to use fast mode by setting:

set MIOPEN_FIND_MODE=FAST

Alternatively, you can disable MIOpen entirely by editing wgp.py and adding the following line below import torch (around line 51):

...
:: /lines already in the file/
:: ...
:: import torch
torch.backends.cudnn.enabled = False # <-- Add this here
:: import gc
:: ...
...

To verify that it is disabled, or to enable verbose logging, you can set:

set MIOPEN_ENABLE_LOGGING=1
set MIOPEN_ENABLE_LOGGING_CMD=1
set MIOPEN_LOG_LEVEL=5

Troubleshooting

GPU Not Detected

If torch.cuda.is_available() returns False:

  1. Verify your GPU is supported - Check the Supported GPUs list above
  2. Check AMD drivers - Ensure you have the latest AMD Adrenalin drivers installed
  3. Verify correct index URL - Make sure you used the right GPU family index URL

Installation Errors

"Could not find a version that satisfies the requirement":

  • Double-check that you're using the correct --index-url for your GPU family. You can also try adding the --pre flag or replacing /v2/ in the URL with /v2/staging/
  • Ensure you're using Python 3.11, and not 3.10

"No matching distribution found":

  • Your GPU architecture may not be supported
  • Check that you've activated your virtual environment

Performance Issues

  • Monitor VRAM usage - Reduce batch size or resolution if running out of memory
  • Close GPU-intensive apps - Apps with hardware acceleration enabled (browsers, Discord etc.).

Known Issues

Windows packages are new and may be unstable.

Known issues are tracked at: ROCm/TheRock#808

Additional Resources

For additional troubleshooting guidance for Wan2GP, see TROUBLESHOOTING.md.