Workshop: AgenticAI4HPC’26: The First International Workshop on Agentic AI for HPC
Authors: Yukito Tajima (Institute of Science Tokyo; CoeFont Co., Ltd.); Kazuki Fujii (Institute of Science Tokyo, National Institute of Advanced Industrial Science and Technology); and Rio Yokota (Institute of Science Tokyo)
Abstract: Automated planners for distributed training must evaluate configurations that have not yet run. ScaleTether transforms version-pinned Megatron executions into workload graphs for tensor, data, and pipeline parallelism. It exports graphs as Chakra traces for ASTRA-sim, while graph support and measured H100 timings determine whether it returns an estimate, requests measurements, or rejects an unsupported configuration. Six graphs created before their target runs matched later captures on every checked field, and all 72 injected corruptions were rejected. Across cases, median timing error ranged from 0.73% to 3.59%. In 80 four-GPU searches, all 35 estimates chose the measured optimum, while measurement intercepted all 19 ungated choices above 1% regret. Selective measurement cut acquisition by 30.3% to 7.58 H100 GPU-hours including calibration and by 44.0% to 6.09 with calibration reuse. Agents followed guidance in 120/120 offline sessions versus 52/120 with configuration numbers alone.
Back to AgenticAI4HPC’26: The First International Workshop on Agentic AI for HPC Archive Listing
Back to Full Workshop Archive Listing