InfraRSI harness applies recursive self-improvement to LLM training
Original title: InfraRSI: Harnessing Recursive Self-Improvement for Training Infrastructure
On four dense and MoE workloads, the authors report up to 2.75x speedup and mean MFU rising from 17.4% to 30.0%.
InfraRSI is an agentic harness that turns open-ended commands into verified training optimizations through a five-stage workflow. It locks a task contract, profiles baselines, co-optimizes launch settings and code patches, and audits promoted changes against loss and timing invariants. The authors say it seeds diverse configurations, models execution time and peak memory including out-of-memory trials, and transfers experience across GPUs.
Unconfirmed: Speedup and MFU gains are the authors' results across four workloads.
- The workflow stages are init, doctor, run, report, and apply.
- It revises LLM priors from live runs and pairs memory-hungry changes with savers.
- Authors say it outperforms existing tuners and coding agents.
- A cross-hardware experience bank stores findings for later clusters.
From the original article
1 min read at huggingface.co
While Recursive Self-Improvement (RSI) accelerates isolated algorithm and kernel synthesis, distributed LLM training depends on tightly coupled launch settings, operator kernels, and communication schedules, where single edits often hit out-of-memory (OOM) boundaries, compiler conflicts, measurement jitter, or silent numerical drift.
By InfraRSI Team, Yuchuan Tian, Haocheng Mei · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.