molssi-seamm/lammps_step: Say plainly when a GPU job has too few tasks

Le résumé fourni par la source

A GPU command that drives an MDI engine runs the engine and LAMMPS as separate MPI ranks, and those counts are written into the command rather than derived from the allocation. A job given fewer tasks than the command asks for died inside mpirun with All nodes which are allocated for this job are already filled, which mentions neither MDI, nor tasks, nor what to change — and arrived wrapped in a conda error, obscuring it further. The step now checks before launching and says how many tasks the command needs, how many the job was given, and that an MDI run needs one task for the engine and one for the driver. This is easy to hit now that queues can offer a GPU count at submission: requesting one GPU and one task looks reasonable but cannot work.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.