This might be a super basic thing but;
If I build lammps with GPU support, is it not possible to run it on a CPU-only machine?
It always looks for the libcuda.so from the CUDA driver.
Thanks you
It is possible.
You cannot get around this. If an executable depends on a shared library, it must be present to run the executable, even if it is not used.
If you want to have a more platform neutral approach, you can use the GPU package in OpenCL mode. This will load the accelerator specific shared libraries only dynamically when used at runtime. You still need the OpenCL library, but that is usually present.
Thank you @akohlmey for the response.
Do you think it should work if I point it to the stubs library file of libcuda.so? Right now it does not for me and gives an illegal instruction error but I am not sure if I am accidentally trying to used a cuda operation somewhere.
Illegal instruction means that something was compiled for a CPU architecture that has additional instructions to what your CPU provides. It has nothing to do with GPUs.
Probably not.