Libra: Efficient Resource Management for Agentic RL Post-Training

arXiv:2606.03077v1 Announce Type: cross Abstract: Reinforcement learning (RL) has become a standard post-training paradigm for large language models (LLMs), extending beyond preference alignment to complex reasoning and multi-turn agentic behaviors. In agentic RL, the rollout stage generates trajectories while invoking tools, producing long-tailed and non-stationary workloads that challenge conventional resource-management assumptions. Three fundamental challenges arise. First, due to the long-tail distribution, a small fraction of trajectories dominates rollout makespan. Second, rollout and training exhibit strong asymmetry in compute patterns, memory demands, and sensitivity to sequence length. Third, as the RL policy evolves, the trajectory-length distribution drifts over time, rendering any static resource split progressively suboptimal. We present Libra, which introduces two core mechanisms. The first is a periodic global resource planner that jointly optimizes GPU allocation across rollout and training clusters. It leverages an elastic hybrid pool to enable lightweight, non-blocking worker reallocation between stages. The second is a causality-driven multi-level feedback queue (C-MLFQ) scheduler, which routes requests to heterogeneous rollout buckets based on causal signals derived from tool-return outcomes, rather than relying on fragile length predictions. Evaluated on 48 A800 GPUs, Libra achieves up to 3.0$\times$ higher throughput and converges up to 2.5$\times$ faster in reward compared to the baselines.

Abstract:Reinforcement learning (RL) has become a standard post-training paradigm for large language models (LLMs), extending beyond preference alignment to complex reasoning and multi-turn agentic behaviors. In agentic RL, the rollout stage generates trajectories while invoking tools, producing long-tailed and non-stationary workloads that challenge conventional resource-management assumptions. Three fundamental challenges arise. First, due to the long-tail distribution, a small fraction of trajectories dominates rollout makespan. Second, rollout and training exhibit strong asymmetry in compute patterns, memory demands, and sensitivity to sequence length. Third, as the RL policy evolves, the trajectory-length distribution drifts over time, rendering any static resource split progressively suboptimal. We present Libra, which introduces two core mechanisms. The first is a periodic global resource planner that jointly optimizes GPU allocation across rollout and training clusters. It leverages an elastic hybrid pool to enable lightweight, non-blocking worker reallocation between stages. The second is a causality-driven multi-level feedback queue (C-MLFQ) scheduler, which routes requests to heterogeneous rollout buckets based on causal signals derived from tool-return outcomes, rather than relying on fragile length predictions. Evaluated on 48 A800 GPUs, Libra achieves up to 3.0$\times$ higher throughput and converges up to 2.5$\times$ faster in reward compared to the baselines.

Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Distributed, Parallel, and Cluster Computing (cs.DC)

arXiv:2606.03077 [cs.LG]

(or arXiv:2606.03077v1 [cs.LG] for this version)

https://doi.org/10.48550/arXiv.2606.03077 Focus to learn more arXiv-issued DOI via DataCite (pending registration)

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.

analiza aiqbench

Libra: Efficient Resource Management for Agentic RL Post-Training

Subiecte conexe