Reproduce “ATTUNE: Agentic Throughput Tuning for LLM Training” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Leo Fan, Hongyang Wang, Wenbo Chen, Jingguo Liu, Xian Li, Xuan Wang, Ruochen Wen, Yiming Yan and 5 more

ATTUNE uses a coding agent to diagnose language-model training runs and test recipe changes, while a deterministic gate checks known objective changes. On B200, its Qwen3.5-397B recipe used 64 GPUs instead of 128, raising throughput per GPU 3.9× and total throughput about 2×. For held-out MiniMax-M2, it reached 2,365.1 tokens/s/GPU, 164.2% above a ported public-recipe baseline, for the tested stack and cluster.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.