Skip to content

GPUs and NVIDIA MPS

Request accelerators as allocation resources. Your target program owns GPU use.

GPU allocation

from molq import Accelerator, Resources

gpu_resources = Resources(
    nodes=1,
    tasks=1,
    cpus_per_task=4,
    accelerators=(Accelerator(kind="gpu", quantity=1, scope="per_node"),),
)

Slurm supports per-node or total GPU requests. OpenPBS supports per-node GPUs. A model selector requires an exact native mapping; see scheduler limits.

MPS allocation

mps_resources = Resources(
    nodes=1,
    tasks=1,
    cpus_per_task=2,
    accelerators=(Accelerator(kind="nvidia_mps", quantity=20, scope="per_node"),),
)

Slurm renders this as --gres=mps:20. The quantity uses the site's configured MPS resource units. The scheduler must advertise that resource.

molq does not start an MPS daemon or replace a rejected MPS request with a whole GPU allocation. Put any site-required MPS setup and cleanup in your payload. The MPS example shows an explicit payload.