InfraRSI: Harnessing Recursive Self-Improvement for Training Infrastructure
InfraRSI reports up to 2.75× faster LLM training across four tested workloads.
The harness jointly tunes launch settings and code patches, checking candidates against fixed workload, loss and timing requirements. It learns from every trial, including out-of-memory failures, to pair memory-hungry optimizations with memory-saving changes. Across single- and multi-node dense and Mixture-of-Experts workloads, mean Model FLOPs Utilization rose from 17.4% to 30.0%. The authors report outperforming existing tuners and coding agents while transferring experience across GPUs.
HF Daily Papers' note
The harness jointly tunes launch settings and code patches, checking candidates against fixed workload, loss and timing requirements. It learns from every trial, including out-of-memory failures, to pair memory-hungry optimizations with memory-saving changes. Across single- and multi-node dense and Mixture-of-Experts workloads, mean Model FLOPs Utilization rose from 17.4% to 30.0%. The authors report outperforming existing tuners and coding agents while transferring experience across GPUs.
HF Daily Papers' note
score 5