Closed lee212 closed 3 years ago
@mtitov , check the code block for gpu counting. On gpu
queue, you have to increase gpu counts by 4, a number of available gpu devices per node. For example, 1,2, and 3 will be failed on gpu
queue but okay with gpu-shared
queue.
@lee212 yeah, just I didn't see it explicitly defined at user guide, but if you checked that then ok. Also don't forget to consider option when gpu_arch
is not provided
This is to reflect slurm options for gpus, and addressing issue #813 according to https://portal.xsede.org/sdsc-comet#gpu.