PSRL Documentation#

PSRL

PSRL is an efficient asynchronous RL framework for LLM post-training. Developed by a joint team from PKU & SJTU and built on top of veRL, PSRL features efficient RDMA weight synchronization via parameter servers, fine-grained staleness control, and flexible rollout coordination to achieve up to 2.68x throughput improvement.


Overview

Project introduction, key features, and performance highlights.

Overview
Quick Start

Get PSRL running in minutes with a minimal DAPO training example.

Quick Start
Tutorial

Step-by-step guides for installation, configuration, and first training run.

Tutorial
Examples

Production-ready recipes for RLVR, Agentic RL, and more.

Examples
Architecture

Understand the decoupled train/gen design and module interactions.

Overall System Architecture
Designs & Features

Deep-dive into staleness control, flexible rollout, and parameter server.

Designs & Features

Key Features#

Efficient RDMA-based Weight Transfer

Push/Pull weights to/from CPU-side Parameter Server and P2P RDMA transfers.

Fine-grained Staleness Control

Trajectory-level version binding with Reserve/Occupy/Consume protocol ensures data freshness without sacrificing throughput.

Flexible Rollout Coordination

Partial rollout, redundant rollout, intelligent routing, and load-balanced migration work together to minimize idle time.

Easy-to-use Agentic RL Support

Native environment loops and SessionRouter/TITO support both integrated and black-box agents.

Hierarchical KV Cache Management

SMG cache-aware routing, vLLM GPU prefix cache, LMCache offload, and P2P transfer reduce re-prefill.

Multiple Backends and Algorithms Support

FSDP2 and Megatron training integrate with PPO, GRPO, DAPO, and generative reward workflows.