Running lammps with kokkos

LAMMPS LAMMPS Installation
21

You could try building and running a raw CUDA “Hello World” example, I’m guessing it would fail in the same manner as Kokkos LAMMPS.

__global__ void cuda_hello(){
    printf("Hello World from GPU!\n");
}

int main() {
    cuda_hello<<<1,1>>>(); 
    return 0;
}
22

@stamoor: Thank you for the suggestion. I’ll try to do this hello-world test.

23

@akohlmey Well, the previous compilation did use GPU via cuda, and I watched GPU usage by using nvidia-smi. I guess cuda was involved through pytorch in my previous compilation. I’ll try to reach out the supercomputer supporting team as you suggested. Thank you for pointing out the direction.

24

The issue has been resolved by my colleague. His solution is quoted as follows:

“the problem was with the runtime linker path (rpath) included in the lmp executable. and the solution is very simple: don’t use the version you’ve just built! first you have to run `make install’, puts the final version of lmp in $HOME/.local/bin”

1 Like