Scheduling Mixed RL Rollouts Beyond Prefix Locality
2026-08-11 • Distributed, Parallel, and Cluster Computing
Distributed, Parallel, and Cluster ComputingMachine Learning
AI summaryⓘ
The authors address problems in efficiently running different types of large language model (LLM) reinforcement learning tasks that share resources but have different needs. They introduce MISA-T, a method that smartly manages memory and task admission to improve processing speed without mixing up the proportions of tasks as intended. Their tests show MISA-T boosts throughput and reduces wait times compared to previous methods while keeping task quality stable. This helps better balance multiple learning jobs running together.
reinforcement learninglarge language modelsrollout workloadprefix-aware routingKV-cacheRLHFRLVRinference servingthroughputcache hit rate
Authors
Zetao Hong, Song Yuan, Yuanhao Ding, Yibo Zhu, Daxin Jiang, Zhibin Wang, Chen Tian
Abstract
Modern reinforcement learning (RL) post-training pipelines for large language models (LLMs) increasingly combine rollout workloads across multiple domains and feedback paradigms. Prefix-aware routing improves inference efficiency through cache reuse and load balancing, but it does not control how heterogeneous rollout sessions compete for KV-cache capacity. When reinforcement learning with verifiable rewards (RLVR), reinforcement learning from human feedback (RLHF), and agentic rollouts share an asynchronous inference service, their distinct sequence structures, interaction patterns, and KV-residency times create substantially different serving demands. Rollout scheduling must account for this heterogeneity without distorting the workload mixture specified by the trainer. We present MISA-T, a routing-layer admission policy for mixed rollout serving. MISA-T combines adaptive session admission, workload-aware KV-capacity allocation, and residency-time-aware KV accounting. In rollout-only ablations on Step3.7 and Qwen3.6-35B-A3B, MISA-T improves rollout throughput over a sweep-tuned cache-aware vLLM Router by 53.3% and 43.6%, respectively, while maintaining high prefix-cache hit rates. In a matched 50-iteration Step3.7 experiment, it increases rollout throughput by 35.6% and reduces mean iteration time by 22.8%, while keeping the consumed workload mixture close to the trainer target and achieving comparable task scores.