Speaker
Description
Deploying GPU-accelerated AI workflows from photon-source beamlines to remote HPC facilities requires navigating incompatible job schedulers, GPU architectures, CUDA versions, and container runtimes. Users and beamline staff must maintain separate deployment procedures for each facility, limiting portability and slowing adoption of new methods.
We present a facility-agnostic pipeline framework and dashboard, developed at SSRL, that enables single-action submission of AI workflows to heterogeneous compute systems. The framework currently orchestrates jobs across five compute platforms spanning national laboratory HPC, an on-premises Kubernetes cluster, and commercial cloud, unified through a common REST API that abstracts away differences in job schedulers and GPU architectures. Globus handles data transfer between the beamline and remote compute. We have demonstrated full container portability between the Kubernetes and cloud targets, with containerization of the remaining HPC targets underway. Beamline staff select a target facility and launch a workflow without modifying the underlying job scripts.
We demonstrate the framework with a GPU-accelerated tomography pipeline that performs reconstruction and zero-shot 3D segmentation using SAM3, a vision foundation model, processing 38 GB of raw synchrotron data in under 30 minutes. We discuss container build strategies for facilities without system CUDA installations, dtype portability challenges in foundation model inference, and our approach to building a catalog of reusable, facility-portable AI workflows for photon science.
