Stability Matrix's stability was lost in the f!@#ing Matrix — Wan2GP + AMD RX 6900 XT / ROCm install failure

[zerotheox]

I really wanted Stability Matrix to work because the entire appeal is avoiding the usual Python/AI dependency nightmare. Unfortunately, trying to install Wan2GP on an AMD RX 6900 XT under Windows turned into exactly that nightmare.

System:

Windows 11
AMD Radeon RX 6900 XT, 16 GB VRAM
RDNA2 / gfx1030
Stability Matrix 2.16.4
Wan2GP, main branch
PyTorch Index: ROCm

On the first attempt, Stability Matrix created a Python 3.11.16 environment and correctly identified the GPU-specific gfx1030 ROCm path. However, the ROCm installation failed during dependency resolution because the required NumPy 2.5.2 package had no compatible CPython 3.11 wheel. The resolver reported wheels for CPython 3.12+ instead.

I completely removed the Wan2GP package and reinstalled it using Python 3.12.14, while leaving the PyTorch index set to ROCm.

That allowed the installation to complete and Wan2GP appeared normally in Stability Matrix with a Launch button.

Unfortunately, launching it immediately crashed during Python imports. The installed SciPy reported that it required:

NumPy >=2.0.0 and <2.8.0

while the environment contained:

NumPy 1.26.4

It then terminated with:

AttributeError: module 'numpy' has no attribute 'long'

So the automated installer ultimately produced an environment whose own NumPy/SciPy dependencies were incompatible.

Importantly, neither failure occurred while actually running a model or testing GPU inference. The installation/environment failed before I could even determine whether ROCm inference works correctly on the RX 6900 XT.

I realize Windows AMD/ROCm support is marked experimental, and I don't expect every RDNA2 configuration to work perfectly. But if Stability Matrix offers a ROCm option for Wan2GP and automatically builds the environment, I think it should either resolve a compatible dependency set or fail before presenting the package as successfully installed.

Hopefully the details above help reproduce it.

TL;DR: Stability Matrix correctly found my RX 6900 XT. It correctly found gfx1030. It correctly attempted Windows ROCm. Then Python dependency hell swallowed the whole thing.

Stability Matrix's stability was, unfortunately, lost in the Matrix. 😁

[sueun-dev]

There's a concrete pin worth checking in the installer: v2.16.4's Wan2GP Windows ROCm profile includes numpy==1.26.4 in PostTorchInstallPipArgs. The ROCm helper applies those arguments after installing Torch, subject to user pip overrides. Selecting Python 3.12 doesn't itself remove that pin.

On Python 3.12.14 in an isolated Linux container, SciPy 1.18.1 with NumPy 1.26.4 produced the same numpy.long exception when importing scipy.stats; python -m pip check also reported the incompatible requirement. With NumPy 2.5.2, the import and a numerical control passed. That tests the dependency interaction, not the Windows installer or ROCm inference, and your exact SciPy version isn't shown here.

Could you check the post-Torch install log for that NumPy pin and share the installed SciPy version? Running python -m pip check with the Wan2GP environment's Python would capture the final dependency conflicts. I'd establish those before treating a NumPy upgrade alone as a fix for the full stack.