Problems running PyTorch bidings

[nvlabs]

Problems running PyTorch bidings

[Tom94]

These errors look like a mismatch between the installed visual studio and CUDA version -- unfortunately, there are frequent breaking changes between the two.

I recommend updating Visual Studio 2019 through the VS installer, afterwards installing the latest CUDA 11.6, rebooting, and then trying again. Fingers crossed!

[danperazzo]

Did not work :/

[danperazzo]

To provide an update, updating the Visual Studio 2019 broke my installation of tiny-cuda-nn :/ I don't know if there is a problem with the incompatibility.

[danperazzo]

I managed to install the Python bindings! Here are the configurations I used:

Windows 10
GTX 1060
CUDA 11.6.1
Visual Studio 2019 (using the CXX compiler MSVC 19.29.30140.0)
Cmake 3.22.1
PyTorch 1.10.2 (installed with conda with cudatoolkit 11.3)

Initially, when I installed the python bindings (running python setup.py install , I got the following error:

fully_fused_mlp.cu
C:\Users\Temp\Desktop\3d-object-stationary-camera-implicit\code_c\tiny-cuda-nn\src\fully_fused_mlp.cu(415): error: name followed by "::" must be a class or namespace name

C:\Users\Temp\Desktop\3d-object-stationary-camera-implicit\code_c\tiny-cuda-nn\src\fully_fused_mlp.cu(493): error: name followed by "::" must be a class or namespace name

C:\Users\Temp\Desktop\3d-object-stationary-camera-implicit\code_c\tiny-cuda-nn\src\fully_fused_mlp.cu(493): error: name followed by "::" must be a class or namespace name

3 errors detected in the compilation of "../../src/fully_fused_mlp.cu".
error: command 'C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v11.6\\bin\\nvcc.exe' failed with exit status 1

However, since my GPU does not have so much memory, I removed the FullyFused from my installation (it would not work anyway). Doing this ensured the instalation weny smoothly and I managed to run the sample python example

[Tom94]

This is a really good catch, thank you! I pushed a fix to setup.py that won't include fully fused MLPs on older architectures where compilation would fail.

(CMake already does this, but I had forgotten to replicate the behavior in Python.)