mir-group
LAMMPS-Allegro compile failed
Hi,
I'm trying to use pair allegro, but I'm having problems compiling it and would like to seek help from the developers.
When I try to compile lammps following the tutorial in Github, it goes smoothly at first and no errors appear. But eventually it will prompt the following error message.
[ 98%] Linking CXX static library liblammps.a
[ 98%] Built target lammps
[ 98%] Building CXX object CMakeFiles/lmp.dir/public/home/cqian/00-APP/LAMMPS-NequIP/lammps/src/main.cpp.o
[100%] Linking CXX executable lmp
/public/home/cqian/00-APP/LAMMPS-NequIP/libtorch/lib/libtorch_cuda_cpp.so: undefined reference to `[email protected]'
/public/home/cqian/00-APP/LAMMPS-NequIP/libtorch/lib/libtorch_cuda_cpp.so: undefined reference to `[email protected]'
/public/home/cqian/00-APP/LAMMPS-NequIP/libtorch/lib/libtorch_cuda_cu.so: undefined reference to `[email protected]'
/public/home/cqian/00-APP/LAMMPS-NequIP/libtorch/lib/libtorch_cuda_cu.so: undefined reference to `[email protected]'
/public/home/cqian/00-APP/LAMMPS-NequIP/libtorch/lib/libtorch_cuda_cpp.so: undefined reference to `[email protected]'
/public/home/cqian/00-APP/LAMMPS-NequIP/libtorch/lib/libtorch_cuda_cpp.so: undefined reference to `[email protected]'
collect2: error: ld returned 1 exit status
make[2]: *** [lmp] Error 1
make[1]: *** [CMakeFiles/lmp.dir/all] Error 2
make: *** [all] Error 2
this is the cmake command
cmake ../cmake -DCMAKE_PREFIX_PATH=/public/home/cqian/00-APP/LAMMPS-NequIP/libtorch -DMKL_INCLUDE_DIR="$CONDA_PREFIX/include"
Do anyone know how to solve this problem ?
Thanks.
Looks like you need to add the folders containing libcudart to $LD_LIBRARY_PATH before compiling. On NERSC Perlmutter for example, this is done by loading the cudatoolkit/12.2 or similar module. cudatoolkit/12.2 adds /opt/nvidia/hpc_sdk/Linux_x86_64/23.9/cuda/12.2/lib64, which contains libcudart.so.
# Best way is to load a CUDA module if it is available on your machine
module load cudatoolkit/12.2
# ...or add the library path yourself it there is no module, e.g.
export LD_LIBRARY_PATH=/opt/nvidia/hpc_sdk/Linux_x86_64/23.9/cuda/12.2/lib64:$LD_LIBRARY_PATH
# ...or if your system has no module but has `$CUDA_HOME` defined,
export LD_LIBRARY_PATH=$CUDA_HOME/lib64:$LD_LIBRARY_PATH
If the linker also can't find libcuda.so, you may also need to add the stubs directory to $LD_LIBRARY_PATH. On NERSC Perlmutter, this directory is at /opt/nvidia/hpc_sdk/Linux_x86_64/23.9/cuda/12.2/lib64/stubs and contains things like libcuda.so, libnppc.so, etc.
# Add stubs folder directly.
export LD_LIBRARY_PATH=/opt/nvidia/hpc_sdk/Linux_x86_64/23.9/cuda/12.2/lib64/stubs:$LD_LIBRARY_PATH
# ...or if your system has no module but has `$CUDA_HOME` defined,
export LD_LIBRARY_PATH=$CUDA_HOME/lib64/stubs:$LD_LIBRARY_PATH
