Search for: All records

Fine-tuning large language models (LLMs) often exceeds GPU memory limits, prompting systems to offload model states to CPU memory. However, existing offloaded training frameworks like ZeRO-Offload treat all parameters equally and update the full model on the CPU, causing severe GPU stalls, where fast, expensive GPUs sit idle, waiting for slow CPU updates and limited-bandwidth PCIe transfers. We present ZenFlow, a new offloaded training framework that decouples updates between GPU and CPU by prioritizing important gradients. ZenFlow performs in-place updates of important gradients on GPU while offloading and accumulating less important ones on CPU without blocking training iterations, fully overlapping CPU work with GPU computation. To make this importance-aware design scalable, ZenFlow leverages a novel spatio-temporal locality of important gradients—concentrated in specific positions that remain stable across iterations—to make important gradient identification efficient and lightweight while preserving fidelity. Evaluation shows that ZenFlow achieves up to 5×end-to-end speedup, 2×lower PCIe traffic, 85% less GPU stalls, and up to 72% lower cloud GPU cost, all while preserving accuracy. 

Long-running scientific workflows, such as tomographic data analysis pipelines, are prone to a variety of failures, including hardware and network disruptions, as well as software errors. These failures can substantially degrade performance and increase turnaround times, particularly in large-scale, geographically distributed, and time-sensitive environments like synchrotron radiation facilities. In this work, we propose and evaluate resilience strategies aimed at mitigating the impact of failures in tomographic reconstruction workflows. Specifically, we introduce an asynchronous, non-blocking checkpointing mechanism and a dynamic load redistribution technique with lazy recovery, designed to enhance workflow reliability and minimize failure-induced overheads. These approaches facilitate progress preservation, balanced load distribution, and efficient recovery in error-prone environments. To evaluate their effectiveness, we implement a 3D tomographic reconstruction pipeline and deploy it across Argonne's leadership computing infrastructure and synchrotron facilities. Our results demonstrate that the proposed resilience techniques significantly reduce failure impact—by up to 500× —while maintaining negligible overhead (<3%). 

The landmark detection of both gravitational waves (GWs) and electromagnetic (EM) radiation from the binary neutron star merger GW170817 has spurred efforts to streamline the follow-up of GW alerts in current and future observing runs of ground-based GW detectors. Within this context, the radio band of the EM spectrum presents unique challenges. Sensitive radio facilities capable of detecting the faint radio afterglow seen in GW170817, and with sufficient angular resolution, have small fields of view compared to typical GW localization areas. Additionally, theoretical models predict that the radio emission from binary neutron star mergers can evolve over weeks to years, necessitating long-term monitoring to probe the physics of the various postmerger ejecta components. These constraints, combined with limited radio observing resources, make the development of more coordinated follow-up strategies essential—especially as the next generation of GW detectors promises a dramatic increase in detection rates. Here, we presentRADAR, a framework designed to address these challenges by promoting community-driven information sharing, federated data analysis, and system resilience, while integrating AI methods for both GW signal identification and radio data aggregation. We show that it is possible to preserve data rights while sharing models that can help design and/or update follow-up strategies. We demonstrate our approach through a case study of GW170817, and discuss future directions for refinement and broader application.