Libra
A systems framework that coordinates GPU and NPU clusters for agentic reinforcement-learning post-training, routing requests across heterogeneous vLLM workers and adjusting resource allocation as workload changes.
Libra splits agentic RL post-training into a core training pool, a core rollout pool and an elastic hybrid pool. A Global Resource Planner decides how many GPUs go to training versus rollout and picks training parallelism and rollout TP bucket layouts, while online replanning periodically reevaluates allocations against runtime history and queue pressure.
A Causality-Guided Bucket Scheduler tracks tool-return states to estimate how much sequence length remains and routes requests to the right buckets, and elastic execution moves complete data-parallel replicas between rollout and training without disrupting the core training topology. Across three agentic RL benchmarks on GPU and NPU clusters, it reports up to 4.2x higher throughput and 2.7x faster reward convergence. Written in Python, MIT licensed, accompanies a research paper.