veRL: Volcano Engine Reinforcement Learning for LLM

veRL is a flexible, efficient and production-ready RL training library for large language models (LLMs).

veRL is the open-source version of HybridFlow: A Flexible and Efficient RLHF Framework paper.

veRL is flexible and easy to use with:

• Easy extension of diverse RL algorithms: The Hybrid programming model combines the strengths of single-controller and multi-controller paradigms to enable flexible representation and efficient execution of complex Post-Training dataflows. Allowing users to build RL dataflows in a few lines of code.

• Seamless integration of existing LLM infra with modular APIs: Decouples computation and data dependencies, enabling seamless integration with existing LLM frameworks, such as PyTorch FSDP, Megatron-LM and vLLM. Moreover, users can easily extend to other LLM training and inference frameworks.

• Flexible device mapping: Supports various placement of models onto different sets of GPUs for efficient resource utilization and scalability across different cluster sizes.

• Readily integration with popular HuggingFace models

veRL is fast with:

• State-of-the-art throughput: By seamlessly integrating existing SOTA LLM training and inference frameworks, veRL achieves high generation and training throughput.

• Efficient actor model resharding with 3D-HybridEngine: Eliminates memory redundancy and significantly reduces communication overhead during transitions between training and generation phases.

| Documentation | Paper | Slack | Wechat | Twitter

• [2025/1] Doubao-1.5-pro is released with SOTA-level performance on LLM & VLM. The RL scaling preview model is trained using veRL, reaching OpenAI O1-level performance on math benchmarks (70.0 pass@1 on AIME).
• [2024/12] The team presented Post-training LLMs: From Algorithms to Infrastructure at NeurIPS 2024. Slides and video available.
• [2024/10] veRL is presented at Ray Summit. Youtube video available.
• [2024/08] HybridFlow (verl) is accepted to EuroSys 2025.

• FSDP and Megatron-LM for training.
• vLLM and TGI for rollout generation, SGLang support coming soon.
• huggingface models support
• Supervised fine-tuning
• Reinforcement learning from human feedback with PPO and GRPO
  • Support model-based reward and function-based reward (verifiable reward)
• flash-attention integration, sequence packing, and long context support via DeepSpeed Ulysses
• scales up to 70B models and hundreds of GPUs
• experiment tracking with wandb and mlflow

• Reward model training
• DPO training

Checkout this Jupyter Notebook to get started with PPO training with a single 24GB L4 GPU (FREE GPU quota provided by Lighting Studio)!

Quickstart:

• Installation
• Quickstart

Running a PPO example step-by-step:

• Data and Reward Preparation
  • Prepare Data (Parquet) for Post-Training
  • Implement Reward Function for Dataset
• Understanding the PPO Example
  • PPO Example Architecture
  • Config Explanation
  • Run GSM8K Example

Reproducible algorithm baselines:

• PPO

For code explanation and advance usage (extension):

• PPO Trainer and Workers
  • PPO Ray Trainer
  • PyTorch FSDP Backend
  • Megatron-LM Backend
• Advance Usage and Extension
  • Ray API design tutorial
  • Extend to Other RL(HF) algorithms
  • Add Models with the FSDP Backend
  • Add Models with the Megatron-LM Backend
  • Deployment using Separate GPU Resources

The performance is essential for on-policy RL algorithm. We write a detailed performance tuning guide to allow people tune the performance. See here for more details.

Contributions from the community are welcome!

We use yapf (Google style) to enforce strict code formatting when reviewing PRs. To reformat you code locally, make sure you installed latest yapf

pip3 install yapf --upgrade

Then, make sure you are at top level of verl repo and run

bash scripts/format.sh

If you find the project helpful, please cite:

• HybridFlow: A Flexible and Efficient RLHF Framework
• A Framework for Training Large Language Models for Code Generation via Proximal Policy Optimization

@article{sheng2024hybridflow,
  title   = {HybridFlow: A Flexible and Efficient RLHF Framework},
  author  = {Guangming Sheng and Chi Zhang and Zilingfeng Ye and Xibin Wu and Wang Zhang and Ru Zhang and Yanghua Peng and Haibin Lin and Chuan Wu},
  year    = {2024},
  journal = {arXiv preprint arXiv: 2409.19256}
}

verl is inspired by the design of Nemo-Aligner, Deepspeed-chat and OpenRLHF. The project is adopted and supported by Anyscale, Bytedance, LMSys.org, Shanghai AI Lab, Tsinghua University, UC Berkeley, UCLA, UIUC, and University of Hong Kong.

• Enhancing Multi-Step Reasoning Abilities of Language Models through Direct Q-Function Optimization
• Flaming-hot Initiation with Regular Execution Sampling for Large Language Models
• Process Reinforcement Through Implicit Rewards
• TinyZero: a reproduction of DeepSeek R1 Zero recipe for reasoning tasks
• RAGEN: a general-purpose reasoning agent training framework

We are HIRING! Send us an email if you are interested in internship/FTE opportunities in MLSys/LLM reasoning/multimodal alignment.