InfraRSI Applies Recursive Self-Improvement to Training Infra: 2.75x Speedup, MFU Lifted from 17.4% to 30%
TokenRhythm · hf · 2026-11-10
TokenRhythm/Infiniscence present InfraRSI, an agentic harness that turns open-ended human commands into verified distributed-training optimizations via a five-stage workflow (init/doctor/run/report/apply) with locked task contracts to prevent reward hacking, baseline profiling and cluster-noise calibration, and a cross-hardware experience bank.
- Fits global execution-time and censored peak-memory models across all trials (including OOMs), prices per-setting memory costs, revises miscalibrated LLM priors from live runs, and pairs memory-hungry optimizations with memory savers.
- Across four single/multi-node dense and MoE workloads: up to 2.75x speedup, mean MFU lifted from 17.4% to 30.0%, beating existing tuners and coding agents while transferring experience across GPUs.
- Forms a two-level RSI loop—revising priors online within runs and distilling experience across clusters—to push foundation model scaling.
More from coding & agent
- Skills vs linked docs: where should AI agents' knowledge live? — IllustriousFault6218 · 2026-11-10
- Fable 5.2 models a Bugatti Tourbillon in Blender from a single prompt in 2.5 hours — socialwithaayan · 2026-11-10
- Pake: Rust-based tool packs any webpage into a lightweight desktop app, 539K downloads — Aiden_Tech_Ai · 2026-11-10
- Frontier3DBench: 145 tasks test whether coding agents can replace 3D vision pipelines — Aiden_Tech_Ai · 2026-11-10
- LASIK open-sourced: strips anti-AI obfuscation, cutting 21M tokens to 17k — morphh7 · 2026-11-10
- Human-in-the-loop is easy; keeping the human attentive is hard — 4 metrics to test if your AI approvals mean anything — alex_verem · 2026-11-10