# CUDA setup not found with update-alternatives symlinks on ubuntu
**URL:** https://discourse.cmake.org/t/cuda-setup-not-found-with-update-alternatives-symlinks-on-ubuntu/11745
**Category:** Usage
**Tags:** os:linux, gen:ninja, lang:cuda
**Created:** [September 22, 2024, 2:29pm UTC](https://discourse.cmake.org/t/cuda-setup-not-found-with-update-alternatives-symlinks-on-ubuntu/11745 "2024-09-22T14:29:12Z")
**Posts on this page:** 3
**Page:** 1
### Author:  [@maichmueller](https://discourse.cmake.org/u/maichmueller)
#### Post date: [September 22, 2024, 2:29pm UTC](https://discourse.cmake.org/t/cuda-setup-not-found-with-update-alternatives-symlinks-on-ubuntu/11745/1 "2024-09-22T14:29:12Z")
Hi,
cmake (26.5 and also 30.3) seem to fail searching for the cuda toolkit when it is symlinked by ubuntu’s `update-alternatives`. More precisely, I have this sample project cmakelists.txt from CLion:
```auto
cmake_minimum_required(VERSION 3.24)
project(xmimir CXX CUDA)
set(CMAKE_CUDA_STANDARD 20)
message(STATUS "CMAKE_CUDA_COMPILER: ${CMAKE_CUDA_COMPILER}")
message(STATUS "CUDAToolkit_ROOT: ${CUDAToolkit_ROOT}")
find_package(Threads REQUIRED)
find_package(CUDAToolkit REQUIRED)
add_library(xmimir STATIC library.cu)
set_target_properties(
xmimir
PROPERTIES
CUDA_SEPARABLE_COMPILATION ON
)
```
At work we have an IT setup in which /usr/local is mounted with nfs4 from another server. The CUDA toolkit is then found in the default /usr/local directory:
```auto
l /usr/local/
drwxr-xr-x@ - root 20 Aug 11:21 admin/
drwxr-xr-x@ - root 5 Jun 2023 bib/
drwxr-xr-x@ - root 19 Sep 16:41 bin/
lrwxrwxrwx - root 19 Sep 16:38 cuda -> /etc/alternatives/cuda/
drwxr-xr-x@ - root 9 Aug 2022 cuda-10.1/
drwxr-xr-x@ - root 9 Aug 2022 cuda-11.2/
drwxr-xr-x@ - root 4 Aug 2022 cuda-11.7/
lrwxrwxrwx - root 19 Sep 16:38 cuda-12 -> /etc/alternatives/cuda-12/
drwxr-xr-x@ - root 19 Sep 16:38 cuda-12.6/
drwxr-xr-x@ - root 9 Aug 2022 cudnn-10.1-v7.6/
drwxr-xr-x@ - root 9 Aug 2022 cudnn-11.X-v8.4/
```
and
```auto
l /usr/bin/nvcc
lrwxrwxrwx - root 21 Sep 19:12 /usr/bin/nvcc -> /etc/alternatives/nvcc*
```
as well as the config in update-alternatives:
```auto
update-alternatives --display nvcc
nvcc - auto mode
link best version is /usr/local/cuda-12.6/bin/nvcc
link currently points to /usr/local/cuda-12.6/bin/nvcc
link nvcc is /usr/bin/nvcc
/usr/local/cuda-12.6/bin/nvcc - priority 100
```
```auto
update-alternatives --display cuda
cuda - auto mode
link best version is /usr/local/cuda-12
link currently points to /usr/local/cuda-12
link cuda is /usr/local/cuda
/usr/local/cuda-12 - priority 100
```
```auto
update-alternatives --display cuda-12
cuda-12 - auto mode
link best version is /usr/local/cuda-12.6
link currently points to /usr/local/cuda-12.6
link cuda-12 is /usr/local/cuda-12
/usr/local/cuda-12.6 - priority 100
```
If I configure this project with the symlinked versions of the cuda toolkit, ie.
nvcc → /usr/bin/nvcc
cuda → /usr/local/cuda,
then cmake cannot find the cuda toolkit / compiler:
```auto
cmake -DCMAKE_BUILD_TYPE=Debug -DCMAKE_C_COMPILER=/usr/bin/gcc-12 -DCMAKE_CXX_COMPILER=/usr/bin/g++-12 -G Ninja -DCUDAToolkit_ROOT=/usr/local/cuda -DCMAKE_CUDA_ARCHITECTURES=80 -DCMAKE_CUDA_COMPILER=/usr/bin/nvcc --debug-output -S /work/rleap1/michael.aichmueller/github/xmimir -B /work/rleap1/michael.aichmueller/github/xmimir/cmake-build-debug-gcc12-nvcc-polonium-broken
Running with debug output on.
-- The CXX compiler identification is GNU 12.3.0
Called from: [3] /u/michael.aichmueller/cmake-3.26.5-linux-x86_64/share/cmake-3.26/Modules/CMakeDetermineCompilerId.cmake
[2] /u/michael.aichmueller/cmake-3.26.5-linux-x86_64/share/cmake-3.26/Modules/CMakeDetermineCXXCompiler.cmake
[1] /work/rleap1/michael.aichmueller/github/xmimir/CMakeLists.txt
CMake Error at /u/michael.aichmueller/cmake-3.26.5-linux-x86_64/share/cmake-3.26/Modules/CMakeDetermineCUDACompiler.cmake:227 (message):
Couldn't find CUDA library root.
Call Stack (most recent call first):
CMakeLists.txt:2 (project)
Called from: [2] /u/michael.aichmueller/cmake-3.26.5-linux-x86_64/share/cmake-3.26/Modules/CMakeDetermineCUDACompiler.cmake
[1] /work/rleap1/michael.aichmueller/github/xmimir/CMakeLists.txt
-- Configuring incomplete, errors occurred!
```
But If I give the true paths, i.e.
nvcc → /usr/local/cuda-12.6/bin/nvcc
cuda → /usr/local/cuda-12.6,
then cmake is able to find everything:
```auto
cmake -DCMAKE_BUILD_TYPE=Debug -DCMAKE_C_COMPILER=/usr/bin/gcc-12 -DCMAKE_CXX_COMPILER=/usr/bin/g++-12 -G Ninja -DCUDAToolkit_ROOT=/usr/local/cuda-12.6 -DCMAKE_CUDA_ARCHITECTURES=80 -DCMAKE_CUDA_COMPILER=/usr/local/cuda-12.6/bin/nvcc --debug-output -S /work/rleap1/michael.aichmueller/github/xmimir -B /work/rleap1/michael.aichmueller/github/xmimir/cmake-build-debug-gcc12-nvcc-polonium
Running with debug output on.
-- CMAKE_CUDA_COMPILER: /usr/local/cuda-12.6/bin/nvcc
Called from: [1] /work/rleap1/michael.aichmueller/github/xmimir/CMakeLists.txt
-- CUDAToolkit_ROOT: /usr/local/cuda-12.6
Called from: [1] /work/rleap1/michael.aichmueller/github/xmimir/CMakeLists.txt
-- Configuring done (0.3s)
-- Generating /work/rleap1/michael.aichmueller/github/xmimir/cmake-build-debug-gcc12-nvcc-polonium
Called from: [1] /work/rleap1/michael.aichmueller/github/xmimir/CMakeLists.txt
-- Generating done (0.0s)
-- Build files have been written to: /work/rleap1/michael.aichmueller/github/xmimir/cmake-build-debug-gcc12-nvcc-polonium
```
Is this behaviour expected?
Is it due to symlinking with update-alternatives or does cmake simply see multiple cuda-xx folders and decide not to choose, despite being given an explicit path?
If the latter, shouldn’t then at least nvcc be found since there is no alternative in /usr/bin ?
---
### Author:  [@craig.scott](https://discourse.cmake.org/u/craig.scott)
#### Post date: [September 22, 2024, 10:01pm UTC](https://discourse.cmake.org/t/cuda-setup-not-found-with-update-alternatives-symlinks-on-ubuntu/11745/2 "2024-09-22T22:01:04Z")
Might be something @robert.maynard knows more about? If I had to guess, I’d assume CMake probably looks for some things in locations relative to where it finds `nvcc`, but it doesn’t follow symlinks and therefore doesn’t see the real locations, but that’s all just a guess.
---
### Author:  [@robert.maynard](https://discourse.cmake.org/u/robert.maynard)
#### Post date: [September 30, 2024, 6:47pm UTC](https://discourse.cmake.org/t/cuda-setup-not-found-with-update-alternatives-symlinks-on-ubuntu/11745/3 "2024-09-30T18:47:36Z")
You don’t want to symlink `nvcc` directly as that breaks the internal compilers detection logic and makes it unusable:
```auto
~ $ ln -s /usr/local/cuda-12.4/bin/nvcc nvcc
~ $ ./nvcc -v -c test.cu
#$ _NVVM_BRANCH_=nvvm
#$ _NVVM_BRANCH_SUFFIX_=
#$ _SPACE_=
#$ _CUDART_=cudart
#$ _HERE_=.
#$ _THERE_=.
#$ _TARGET_SIZE_=
#$ _TARGET_DIR_=
#$ _TARGET_SIZE_=64
#$ gcc -D __CUDA_ARCH_LIST__ =520 -D __NV_LEGACY_LAUNCH -E -x c++ -D__ CUDACC __-D__ NVCC __-D__ CUDACC_VER_MAJOR __=12 -D__ CUDACC_VER_MINOR __=4 -D__ CUDACC_VER_BUILD __=131 -D__ CUDA_API_VER_MAJOR __=12 -D__ CUDA_API_VER_MINOR __=4 -D__ NVCC_DIAG_PRAGMA_SUPPORT__=1 -include "cuda_runtime.h" -m64 "test.cu" -o "/tmp/tmpxft_00008507_00000000-5_test.cpp4.ii"
cc1plus: fatal error: cuda_runtime.h: No such file or directory
compilation terminated.
# --error 0x1 --
~ $ nvcc -v -c test.cu
#$ _NVVM_BRANCH_=nvvm
#$ _NVVM_BRANCH_SUFFIX_=
#$ _SPACE_=
#$ _CUDART_=cudart
#$ _HERE_=/usr/local/cuda/bin
#$ _THERE_=/usr/local/cuda/bin
#$ _TARGET_SIZE_=
#$ _TARGET_DIR_=
#$ _TARGET_DIR_=targets/x86_64-linux
#$ TOP=/usr/local/cuda/bin/..
#$ NVVMIR_LIBRARY_DIR=/usr/local/cuda/bin/../nvvm/libdevice
#$ LD_LIBRARY_PATH=/usr/local/cuda/bin/../lib:/usr/local/cuda/lib64
#$ PATH=/usr/local/cuda/bin/../nvvm/bin:/usr/local/cuda/bin:/home/rmaynard/.local/bin:/usr/local/cuda/bin:/home/rmaynard/.cargo/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin:/snap/bin:/home/rmaynard/.cargo/bin
#$ INCLUDES="-I/usr/local/cuda/bin/../targets/x86_64-linux/include"
#$ LIBRARIES= "-L/usr/local/cuda/bin/../targets/x86_64-linux/lib/stubs" "-L/usr/local/cuda/bin/../targets/x86_64-linux/lib"
#$ CUDAFE_FLAGS=
#$ PTXAS_FLAGS=
#$ gcc -D __CUDA_ARCH_LIST__ =520 -D __NV_LEGACY_LAUNCH -E -x c++ -D__ CUDACC __-D__ NVCC __"-I/usr/local/cuda/bin/../targets/x86_64-linux/include" -D__ CUDACC_VER_MAJOR __=12 -D__ CUDACC_VER_MINOR __=4 -D__ CUDACC_VER_BUILD __=131 -D__ CUDA_API_VER_MAJOR __=12 -D__ CUDA_API_VER_MINOR __=4 -D__ NVCC_DIAG_PRAGMA_SUPPORT__=1 -include "cuda_runtime.h" -m64 "test.cu" -o "/tmp/tmpxft_00008518_00000000-5_test.cpp4.ii"
...
#$ gcc -D __CUDA_ARCH__ =520 -D __CUDA_ARCH_LIST__ =520 -D__NV_LEGACY_LAUNCH -c -x c++ -DCUDA_DOUBLE_MATH_FUNCTIONS -Wno-psabi "-I/usr/local/cuda/bin/../targets/x86_64-linux/include" -m64 "/tmp/tmpxft_00008518_00000000-6_test.cudafe1.cpp" -o "test.o"
```
What you want to do is use your `update-alternatives` setup to switch the cuda symlink directory ( which it looks like already have setup ).