JigsawRL: Assembling RL Pipelines for Efficient LLM Post-Training

TL;DR

JigsawRL is a framework that improves RL pipeline efficiency by dynamic resource allocation and multiplexing, achieving significant throughput gains on GPU clusters.

Contribution

It introduces Pipeline Multiplexing and a Sub-Stage Graph abstraction to optimize resource use and coordination in RL pipelines.

Findings

Up to 1.85x throughput improvement over Verl on synchronous RL.

Up to 1.54x throughput improvement over StreamRL and AReaL on asynchronous RL.

Supports heterogeneous pipelines with moderate latency trade-offs.

Abstract

We present JigsawRL, a cost-efficient framework that explores Pipeline Multiplexing as a new dimension of RL parallelism. JigsawRL decomposes each pipeline into a Sub-Stage Graph that exposes the intra-stage and inter-worker imbalance hidden by stage-level systems. On this abstraction, JigsawRL resolves multiplexing interference through dynamic resource allocation, eliminates fragmented utilization by migrating long-tail rollouts across workers, and formulates their coordination as a graph scheduling problem solved with a look-ahead heuristic. On 4-64 H100/A100 GPUs across different agentic RL pipelines and models, JigsawRL achieves up to 1.85x throughput over Verl on synchronous RL, 1.54x over StreamRL and AReaL on asynchronous RL, and supports heterogeneous pipelines with moderate latency trade-off.