Hi, NVIDIA CUDA version >= 11.4 does not seem to install successfully.

[XiaoqiWang]

I tried to install the package in Tesla A100 and GeForce RTX 3090 with CUDA version 11.4 both failed. Can you provide some help please? Thank you very much!

[XiaoqiWang]

E:\conda\lib\site-packages\torch\include\pybind11\cast.h(1429): error: too few arguments for template template parameter "Tuple"
detected during instantiation of class "pybind11::detail::tuple_caster<Tuple, Ts...> [with Tuple=std::pair, Ts=<T1, T2>]"
(1507): here

  E:\conda\lib\site-packages\torch\include\pybind11\cast.h(1503): error: too few arguments for template template parameter "Tuple"
            detected during instantiation of class "pybind11::detail::tuple_caster<Tuple, Ts...> [with Tuple=std::pair, Ts=<T1, T2>]"
  (1507): here

  2 errors detected in the compilation of "torchsort/isotonic_cuda.cu".
  isotonic_cuda.cu
  error: command 'C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v11.4\\bin\\nvcc.exe' failed with exit code 1

[teddykoker]

Hi @XiaoqiWang, what version of PyTorch do you have installed? I have been able to build torchsort just fine in 11.4 on my machine, but I wonder if a different pytorch version is causing the issue.

[XiaoqiWang]

Hi，thank you for your reply. My pytorch version is 1.11.0 + cu113.

[teddykoker]

I see you are on a windows system - what version is your c++ compiler? It seems there may be an issue with pybind and certain versions of visual studio and CUDA, see: facebookresearch/pytorch3d#1024

[XiaoqiWang]

Thanks. I will try these solutions and get back to you as soon as possible.

[XiaoqiWang]

I successfully installed the package on a linux centos7 system, the system's g++ version requires higher than 5.0, otherwise the following error is reported: unrecognized command line option '-std=c++14'

[teddykoker]

Great! Seems like there are some issues with windows and pytorch cuda extensions that may have caused the issue. Glad you were able to install it!