Hi all!

I'm trying to get the ML Train GSplats TOPs network working on Windows and running into a CUDA compilation error. Hoping someone has a fix or workaround (using the ML Train Gsplats from Karma recipe )

My Setup:

Houdini 22.0.448 (Python 3.11)
Windows 10
NVIDIA Quadro RTX 5000 (16GB, Turing/Compute 7.5)
CUDA 12.4
MSVC 14.39 (VS 2022 Build Tools)
gsplat 1.5.3
torch 2.6.0+cu124

What's happening:
The TOPs network starts training, initializes the model correctly (200k Gaussians), but fails when gsplat tries to JIT compile its CUDA extensions. The specific error is:

cl : Command line error D8021 : invalid numeric argument '/Wno-attributes'

This happens at step 12/30 when compiling Adam.cpp. The flag /Wno-attributes is a GCC/Linux flag that MSVC doesn't support.

I've tried installing MSVC v14.39 toolset to match CUDA 12.4 requirements
Setting CUDA and MSVC paths explicitly before launching Houdini via Developer Command Prompt
Downgrading torch from 2.9.1 to 2.6.0

The CUDA .cu files all compile successfully, but it's specifically the .cpp files being compiled with cl.exe that fail due to the GCC flag.



Is there a Windows compatible build of the gsplat CUDA extensions available, or a workaround to patch the compile flags for MSVC? Unless maybe this a known issue on Windows?

Thanks for any help!

Hi,

By default the node should install its own virtual environment with torch2.9.1/cuda12.8 and prebuilt gsplat wheels we ship with Houdini (linux&windows). Cuda 12.8 supports compute architectures >sm_60 so you should be well within the supported range.

Did this fail for you? What lead to you trying to build from source/switching to torch2.6+cuda12.4?

Which driver version are you using?

Best,

Jakob