SeT-Diff: Towards Semantic Foundation Models for HPC Telemetry and Time-Series

Recent research introduces SeT-Diff, the first foundational model for compute node telemetry and time-series, utilizing a diffusion-based approach to model complex HPC workloads and physical metrics. Simultaneously, a new framework employs LLM as Forecasting Planner to integrate natural-language context with Time-Series Foundation Models. This training-free text conditioning allows forecasts to account for external events without distorting the temporal structure, overcoming the challenges of combining numerical forecasting with the reasoning capabilities of large language models.