ATTUNE: Agentic Throughput Tuning for LLM Training

Key points

• Training recipes are tuned by hand for one model, cluster and workload, and every new model, accelerator, framework version or dataset means tuning again.1
• Automated tuners search a fixed space of knobs and check only that a job runs, yet a configuration can run to completion while it silently changes training.1
• On four held-out models at 64-128 GPUs, ATTUNE raises mean throughput over the strongest public recipe run on our stack by 6.7% on Llama-3.1-405B (best campaign 11.9%) and 33.9% on GLM-4.5 (54.7%), whose public recipe was written for H100.1
• Under a common GPU-hour budget it beats OptiCo, Bayesian optimization and random search on GLM-4.5 and MiniMax-M2; on GLM-4.5 its mean gain is more than four times the best alternative's at less than half the GPU-hours, and on both models its tuning pays for itself within a day of training.1

Sources (1)

• [1]ATTUNE: Agentic Throughput Tuning for LLM Training
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 13, 12:49 AM

  ATTUNE is a single coding agent that tunes any working recipe for the workload and stack at hand, unattended, within a GPU-hour budget the user sets.

  Training recipes are tuned by hand for one model, cluster and workload, and every new model, accelerator, framework version or dataset means tuning again.

Extractive summary: sentences quoted from the sources.