I am trying to build this extension GitHub - daerduoCarey/PyTorchEMD: PyTorch Wrapper for Earth-Mover-Distance (EMD) for 3D point cloud regression but I am getting an errors with CUDA 11.1 and Pytorch cuda 11.3. and Pytorch cuda10.2
The detected CUDA version (11.1) mismatches the version that was used to compile PyTorch (10.2). Please make sure to use the same CUDA versions.
I think I had this issue because I was using Python 3.6 . Then I got
cuda/emd_kernel.cu(385): error: identifier "TORCH_CHECK" is undefined cuda/emd_kernel.cu(391): error: no suitable constructor exists to convert from "c10::ScalarType" to "at::Type" 6 errors detected in the compilation of "cuda/emd_kernel.cu". error: command '/usr/bin/nvcc' failed with exit status 1
Moving to Python 3.8, and Installing Pytorch cuda 11.3 I run into these messages:
line 1694, in _get_cuda_arch_flags arch_list[-1] += '+PTX' IndexError: list index out of range
So I used
TORCH_CUDA_ARCH_LIST="7.5" python setup.py install but I got this
error: THC/THC.h: No such file or directory 14 | #include <THC/THC.h> | ^~~~~~~~~~~ compilation terminated.
I am not sure how could I install Pytorch extension properly, compiling them correctly, or where can I learn more about it outside of this page Custom C++ and CUDA Extensions — PyTorch Tutorials 1.12.1+cu102 documentation