NVIDIA / AMGX

Distributed multigrid linear solver library on GPU
497 stars 143 forks source link

Doesn't compile with CUDA 12.6.77 using conda environment #326

Open Maccy-Z opened 1 month ago

Maccy-Z commented 1 month ago

Describe the issue

Doesn't compile with CUDA 12.6.77

Environment information:

Configuration information

cmake ../

Compilation information

- Issue information

cmake ../ runs fine, until it returns an error:

CMake Error at CMakeLists.txt:272 (target_link_libraries): Target "amgx" links to:

CUDA::nvToolsExt

but the target was not found. Possible reasons include:

CMake Error at CMakeLists.txt:273 (target_link_libraries): Target "amgxsh" links to:

CUDA::nvToolsExt

but the target was not found. Possible reasons include:

Additional context Compiling in Conda environment. Compiles fine when I install older CUDA 11.7. Doesn't work with the cuda version 12.6 from: conda install cuda -c nvidia

Im not too familar with cuda, but nvToolsExt seems to be modified in CUDA 12 to be header only and not linkable and changed to nvtx3:

https://gitlab.kitware.com/cmake/cmake/-/issues/22893

https://github.com/pytorch/pytorch/issues/116242

marsaev commented 1 month ago

Hi @Maccy-Z I see linked issues are all referring Windows, just wanted to confirm if you experienced this on Ubuntu. I can't repro it on Linux + CMake 3.20 + CUDA 12.6, it seems on linux CUDA has both old nvtx + lib and nvtx3, and old nvtx is being used.

Maccy-Z commented 3 weeks ago

Yes this is on Ubuntu

However, Im using it in a conda env, is probably why then. The conda package only has the newer version and I don't have cuda installed on the main system.

marsaev commented 3 weeks ago

@Maccy-Z In general - we should support conda environment, so we should chase it down. Let me leave this issue open for future reference.