Input files
A simulation requires run.in, model.xyz, and a supported potential file in the working directory.
Run gpumd_replica from this directory.
One process uses the visible GPUs on a single node; select GPUs with CUDA_VISIBLE_DEVICES or the job scheduler.