Libra System Boosts Agentic RL Post-Training Throughput Up to 3x Over Baselines
机器之心 · wechat · 2026-08-12
To address workload instability and dynamic GPU demands during Agentic RL post-training, a research team from CUHK and HSUHK proposed Libra, a unified resource management system.
- Core Approach: Treats training and rollout as a coupled system rather than a fixed bottleneck. A Global Resource Planner jointly optimizes GPU allocation and parallelism strategies, periodically adjusting based on real-time workloads.
- Elastic Scheduling: Introduces a hybrid pool and a causality-driven scheduler (C-MLFQ). By leveraging causal signals from tool returns to predict trajectory length, it achieves a 91.1% single-route accuracy.
- Results: On 48 NVIDIA A800 GPUs across three tasks (Search-R1, DAPO-Math, R2E-Gym), Libra achieved the highest throughput (up to 3.0x over baselines) and reduced wall-clock time to target reward by up to 2.5x.
Both the paper and code are open-source.
More from Infra
- AI Agents Drive Up Vercel Build Costs, Developer Seks Optimization — jonathan_wilke · 2026-08-12
- Linux Cloud GPU Blocked from Video Upscaling: NVIDIA RTX VSR is Windows-Only — emacrema · 2026-08-12
- GPU Demand Surges: H100 Rental Prices Jump 40% in Six Months — jessi_cata · 2026-08-12
- AI Data Center Firm DayOne Confidentially Files for US IPO — davidyin44 · 2026-08-12
- Anthropic Reportedly Inks $9.1B Compute Deal; Microsoft to Massively Boost AI Chip Production — 创业邦 · 2026-08-12
- AI Chip Yield Anxiety: Why Wafer-Level Testing is Becoming Critical — demian_ai · 2026-08-12