Amazon's ATTUNE coding agent lifts LLM training throughput 1.7x on H100/H200/B200 clusters

kalyan_kpl · x · 2026-10-25

• Amazon releases ATTUNE, an unattended coding agent that tunes any working LLM training recipe within a GPU-hour budget; open-sourced at github.com/amazon-science/attune with the OpenRecipes leaderboard (openrecipes.ai).
• Motivation: training throughput depends on the recipe (GPU sharding, memory-compute tradeoffs, kernel and communication-overlap choices), and the best recipe shifts with every model, accelerator, and framework version. Existing tuners search a fixed knob space and only check that jobs run, missing levers and sometimes silently changing training.
• How it works: guided by a skill of search rules, scripts, and lessons from past tuning campaigns, it diagnoses each run and changes one lever at a time; a correctness gate outside the agent rejects any argument known to change the training objective.
• Results: 1.7x average throughput gain tuning Amazon's production post-training recipes on H100, H200, and B200; beats published throughput of four NVIDIA AutoModel benchmark recipes by 6–17%; up to 2.6x on held-out MiniMax-M2; outperforms Bayesian optimization, random search, and OptiCo at equal GPU-hour budget; on GLM-4.5 gains are 4x+ the best alternative. The same agent without ATTUNE's skill gains less and fails half its launches.

Original post →

More from Infra

• DeepInfer is deeply tailored to DeepSeek 4.1F: KV cache of just 890 bytes per token — teortaxesTex · 2026-10-25
• Reasoning costs plummet as startup Talaas matches System 1 speeds — where's the alpha? — Square-Listen-7989 · 2026-10-25
• NextDC plunges 40% in a day: the real AI bottleneck is the power line, not contracts — ccerrato147 · 2026-10-25
• A neocloud is a mortgage on a data center: rent comes later, interest is due now — ccerrato147 · 2026-10-25
• Korean AI data center firm Elice falls 32% on debut despite 400x oversubscription — ccerrato147 · 2026-10-25
• DeepSeek 552B model runs on a single AMD MI325X: 4 agents at 100 tok/s — realmrfakename · 2026-10-25

Infra channel →