These images are built from Nvidia Triton inference server base images, with updated PyTorch libraries included. Having these as Singularity containers helps users test fast ML inference code when they are on a machine without superuser permission (which is needed to use Docker).