@2297019091 Can you try running the same simulation but replace pair Allegro with pair LJ/cut? If the bug is in LAMMPS not allegro then the issue should persist and we can debug. One guess is that your simulation is exploding on timestep 0. This can lead to cudaErrorIllegalAddress when binning atoms for the neighbor list because the value is too large. For example, see LAMMPS Kokkos GPU: cudaErrorIllegalAddress during neighbor build (NBinKokkos::bin_atoms) - #3 by Qixuan.