吞吐最高提升300%！港中文开源Libra，RL训练提速2.5倍
Ranking
Observed public metrics from 1 member.
Merged summary
TL;DR — Libra 是面向智能体强化学习后训练的开源资源管理系统，可在 rollout 与训练之间动态调配 GPU。在 48 张 NVIDIA A800 上，其吞吐量较 colocated verl 提升 196%–300%，达到目标奖励的速度最高提升 2.5 倍。
- 全局规划器联合优化训练并行策略、rollout GPU 分配及异构张量并行（TP）配置。
- C-MLFQ 利用工具返回的因果运行时信号，将不断演化的轨迹路由至合适的 TP 资源桶，无需预先预测轨迹长度；在 Search-R1 上路由准确率达 91.1%。
- 弹性混合 GPU 池可在 rollout 与训练间转移算力，无需重建核心训练通信拓扑或暂停训练。
- 在 Search-R1、DAPO-Math-17K 和 R2E-Gym 的搜索、数学与编程智能体任务中，Libra 在最终奖励相当的情况下取得了参评系统中的最高吞吐量。
注： PaperWeekly 更强调相对 colocated verl 的吞吐量提升，机器之心则突出 C-MLFQ 的路由准确率及跨基准表现。
Sources (2)
吞吐最高提升300%！港中文开源Libra，RL训练提速2.5倍
TL;DR - Libra is an open-source resource-management system for agentic RL post-training that dynamically balances GPUs between rollout and training. It improves throughput by up to 300% and reduces time-to-reward by up to 2.5×.
- A global planner jointly optimizes training parallelism, rollout GPU allocation, and heterogeneous TP configurations.
- C-MLFQ uses runtime tool-return signals to route evolving trajectories to suitable TP buckets instead of predicting lengths upfront.
- An elastic hybrid GPU pool shifts capacity without rebuilding the core training topology or pausing training.
- On 48 A800 GPUs, Libra achieved 196–300% higher throughput than colocated verl across search, math, and coding-agent workloads.
Agentic RL 后训练资源怎么分？港中文、恒生大学提出 Libra，吞吐最高提升 3 倍
TL;DR - Libra dynamically reallocates GPUs between rollout and training during agentic RL post-training, addressing workload drift and highly variable trajectory lengths. On 48 NVIDIA A800 GPUs, it delivered up to 3× baseline throughput and reached target rewards up to 2.5× faster.
- A global planner jointly optimizes training parallelism, heterogeneous rollout instances, and GPU allocation.
- An elastic hybrid pool shifts resources without rebuilding core training communication groups.
- C-MLFQ routes trajectories using causal signals from tool responses, achieving 91.1% routing accuracy on Search-R1.
- Tests across Search-R1, DAPO-Math-17K, and R2E-Gym showed the highest throughput among evaluated systems at comparable final rewards.