Windows Builds

[alihassanijr]

Opening this discussion for Windows users that are trying to build NATTEN from source.
I'm hoping users that have experience in doing that successfully could help out.

I remember there was once a HuggingFace page with NATTEN wheels for windows.
I think @guonetnet51 may also be running on Windows.
If you could please share some information with other Windows users on how to get around various build issues, that would be greatly appreciated.

Users in #171, #212, #214, and any other Windows users, could you please start threads on build issues you've been having?

[DON6679]

comfyui , natten import failed , python3.11.9 pytorch 2.7+cu128 , Install natten-0.17.5+torch270cu128-cp311-cp311-win_amd64 ，How to repair and install

[DON6679]

FileNotFoundError: [Errno 2] No such file or directory: 'E:\01SD-AI\ComfyUI-aki-v1.6\ComfyUI\custom_nodes\NATTEN\init.py'

Cannot import E:\01SD-AI\ComfyUI-aki-v1.6\ComfyUI\custom_nodes\NATTEN module for custom nodes: [Errno 2] No such file or directory: 'E:\01SD-AI\ComfyUI-aki-v1.6\ComfyUI\custom_nodes\NATTEN\init.py'

[twoicewu]

error_log.txt

os: Windows 11
python: Python 3.12.9
torch: 2.7.1+cu128
cl: 19.44.35211

[alihassanijr]

Some good exchange in this issue: #241

[1803233552]

I tried using AI to fix it, but it didn't work. Here are the conclusions from the AI analysis.
Hi team,
I'm experiencing a persistent build failure when trying to compile the latest version of NATTEN from source on a modern Windows machine. The build consistently fails during the CMake compiler identification step.
Environment:
OS: Windows 11
GPU: NVIDIA RTX 5090 (Blackwell Architecture)
CUDA Toolkit: 12.8
PyTorch: 2.7.0 (installed via pip with +cu128)
Python: 3.12
Compiler: Visual Studio 2022 Build Tools (with "Desktop development with C++" workload installed)
Problem Description:
The build process fails during CMake configuration while trying to verify the CUDA compiler (nvcc.exe). The detailed log shows that nvcc invokes the MSVC compiler (cl.exe), which then exits with a "The system cannot find the path specified" error (the original error in Chinese is 系统找不到指定的路径。).
This occurs even when the environment is correctly set up inside the Visual Studio x64 Native Tools Command Prompt. This strongly suggests a deep toolchain issue, as hinted at by the warning in the setup.py script regarding CUDA 12.x on Windows.
Troubleshooting Steps Attempted:
I have already tried numerous workarounds, none of which have resolved the issue:
Ensured all commands were run from the x64 Native Tools Command Prompt for VS 2022.
Explicitly set the CUDA architecture using set NATTEN_CUDA_ARCH=90.
Attempted to force the compiler via environment variables (set CC=..., set CXX=...).
Directly modified the setup.py file to hardcode the CMAKE_C_COMPILER, CMAKE_CXX_COMPILER, and CMAKE_CUDA_HOST_COMPILER variables passed to the CMake command.
Even with the most explicit compiler path hardcoded into the build script, the build still fails with the exact same error, indicating that the problem lies within the interaction between CMake, nvcc, and cl.exe.
Full Error Log:
Here is the complete output from the final build attempt. This log was generated using a modified setup.py that hardcodes the compiler paths, demonstrating that the failure persists even with all local workarounds applied.

log.txt

It seems this is a toolchain issue that is difficult to resolve with local configuration changes. Any help or guidance would be greatly appreciated. I'm happy to provide more information or run further diagnostics if needed.
Thank you for your time and for your work on this library.
Best regards.