TL;DR of the paper, generated automatically.

A harness that tunes LLM training setups, reporting up to 2.75x speedup.

• Results: mean Model FLOPs Utilization rises from 17.4% to 30.0%.
• Method: locks a fixed task contract, then tunes launch settings and code patches.
• Coverage: four dense and MoE workloads, single and multi-node.