verl-FL

verl-FL is a fork of verl designed to support diverse AI accelerators. It is built on top of FlagOS, a unified open-source AI system software stack, and integrates key components including the training engines Megatron-LM-FL and Transformer-Engine-FL, as well as the inference engine vllm-plugin-FL.

👋 Hi, everyone! verl is a RL training library initiated by ByteDance Seed team and maintained by the verl community.

seed logo

verl: Volcano Engine Reinforcement Learning for LLMs

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-controller programming model enables flexible representation and efficient execution of complex post-training dataflows. Build RL dataflows such as GRPO, PPO 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 FSDP, Megatron-LM, vLLM, SGLang, etc
Flexible device mapping: Supports various placement of models onto different sets of GPUs for efficient resource utilization and scalability across different cluster sizes.
Ready integration with popular HuggingFace models

verl is fast with:

State-of-the-art throughput: SOTA LLM training and inference engine integrations and SOTA RL throughput.
Efficient actor model resharding with 3D-HybridEngine: Eliminates memory redundancy and significantly reduces communication overhead during transitions between training and generation phases.

News

[2025/12] Mind Lab successfully used verl and Megatron-bridge to train GRPO Lora for Trillion-parameter model on 64 H800 - See their techblog.
[2025/11] recipes have been moved to a new repository: verl-recipe
[2025/10] verl is presented in the PyTorch Conference 2025.
[2025/08] verl is presented in the PyTorch Expert Exchange Webinar. Slides available.
[2025/07] The ReTool recipe is fully open sourced. Blog
[2025/07] The first verl meetup will be held at ICML Vancouver on July 16th! Please join us if you are at ICML! (onsite only)
[2025/06] verl with Megatron backend enables large MoE models such as DeepSeek-671B and Qwen3-235B.
[2025/03] DAPO is the open-sourced SOTA RL algorithm that achieves 50 points on AIME 2024 based on the Qwen2.5-32B pre-trained model, surpassing the previous SOTA achieved by DeepSeek’s GRPO (DeepSeek-R1-Zero-Qwen-32B). DAPO’s training is fully powered by verl and the reproduction code is available in recipe/dapo now.
more...
- [2025/04] [Seed-Thinking-v1.5](https://github.com/ByteDance-Seed/Seed-Thinking-v1.5/blob/main/seed-thinking-v1.5.pdf) tech report is released! Trained with verl, Seed-Thinking-v1.5 achieves 86.7 on AIME 2024, 55.0 on Codeforces and 77.3 on GPQA, demonstrating excellent reasoning abilities in STEM and coding. Beyond reasoning tasks, the method demonstrates notable generalization across diverse domains.
- [2025/07] verl keynote at [AWS AI Hours Singapore](https://pages.awscloud.com/aws-ai-hours-sg.html#agenda) on 7/8, verl & verl-agent project updates at [Agent for SWE meetup](https://lu.ma/e498qhsi) by LF AI & Data Singapore on 7/11.
- [2025/06] verl team will provide latest project updates at [PyTorch Day China](https://www.lfasiallc.com/pytorch-day-china/) on June 7th. Meet our dev team in Beijing!
- [2025/04] [VAPO](https://arxiv.org/pdf/2504.05118) (value-based augmented PPO) paper covers our latest RL method for reasoning models. Trained from Qwen-32B-base model, VAPO achieves 60.4 on AIME 2024, outperforming DAPO-32B.
- [2025/05] [PF-PPO](https://arxiv.org/abs/2409.06957), accepted to ICML 2025, is now supported in verl! PF-PPO enhances policy learning efficiency and robustness by filtering potentially noisy reward signals and reusing high-quality experiences via a replay buffer.
- [2025/04] We will give a tutorial about latest post-training techniques and programming guide for verl at [ICLR 2025 Expo](https://iclr.cc/virtual/2025/calendar?filter_events=Expo+Talk+Panel&filter_rooms=), [SCI-FM workshop](https://open-foundation-model.github.io/) and [LMSys afterparty](https://lu.ma/d23nyynm). Talk materials available [here](https://github.com/eric-haibin-lin/verl-community/tree/main/iclr25).
- [2025/03] verl v0.3.0.post1 is released! See [release note](https://github.com/volcengine/verl/releases/) for details. It achieves [~1.4x speedup](https://tongyx361.github.io/blogs/posts/verl-intro/#/verl-flexible-and-efficient-rl-for-llms) compared to prev versions.
- [2025/05] verl will be presented at [A2M Shanghai](https://a2m.msup.com.cn/home/?aid=4488&city=shanghai) on 5/16 - 5/17.
- [2025/05] verl will be presented at [GOSIM x PyTorch Day 2025](https://paris2025.gosim.org/). See you in Paris!
- [2025/03] We introduced the programming model of verl at the [vLLM Beijing Meetup](https://mp.weixin.qq.com/s/n77GibL2corAtQHtVEAzfg) and [verl intro and updates](https://github.com/eric-haibin-lin/verl-community/blob/main/slides/verl-lmsys-meetup.pdf) at the [SGLang-LMSYS Org Meetup](https://lu.ma/ntjrr7ig) in Sunnyvale mid-March.
- [2025/03] We will present verl(HybridFlow) at EuroSys 2025. See you in Rotterdam!
- [2025/02] verl v0.2.0.post2 is released!
- [2025/02] We presented verl in the Bytedance/NVIDIA/Anyscale Ray Meetup. See you in San Jose!
- [2025/01] [Doubao-1.5-pro](https://team.doubao.com/zh/special/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] verl is presented at Ray Forward 2024. Slides available here
- [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.

Key Features

FSDP, FSDP2 and Megatron-LM for training.
vLLM, SGLang and HF Transformers for rollout generation.
Compatible with Hugging Face Transformers and Modelscope Hub: Qwen-3, Qwen-2.5, Llama3.1, Gemma2, DeepSeek-LLM, etc
Supervised fine-tuning.
Reinforcement learning with PPO, GRPO, GSPO, ReMax, REINFORCE++, RLOO, PRIME, DAPO, DrGRPO, KL_Cov & Clip_Cov etc.
- Support model-based reward and function-based reward (verifiable reward) for math, coding, etc
- Support vision-language models (VLMs) and multi-modal RL with Qwen2.5-vl, Kimi-VL
- Multi-turn with tool calling
LLM alignment recipes such as Self-play preference optimization (SPPO)
Flash attention 2, sequence packing, sequence parallelism support via DeepSpeed Ulysses, LoRA, Liger-kernel.
Scales up to 671B models and hundreds of GPUs with expert parallelism
Multi-gpu LoRA RL support to save memory.
Experiment tracking with wandb, swanlab, mlflow and tensorboard.
Hardware Support: Supports NVIDIA, AMD, Ascend

Upcoming Features and Changes

Q3 Roadmap https://github.com/volcengine/verl/issues/2388
DeepSeek 671b optimizations with Megatron https://github.com/volcengine/verl/issues/1033
Multi-turn rollout and tools using optimizations https://github.com/volcengine/verl/issues/1882
Agent integration
Async and off-policy architecture https://github.com/volcengine/verl/pull/2231
List of breaking changes since v0.4 https://github.com/volcengine/verl/discussions/2270

Getting Started

Documentation

Quickstart:

Installation
Quickstart
Programming Guide & Tech Talk (in Chinese)
PPO in verl
GRPO in verl

Running a PPO example step-by-step:

Reproducible algorithm baselines:

RL performance on coding, math

For code explanation and advance usage (extension):

PPO Trainer and Workers
Advanced Usage and Extension

Blogs from the community

Performance Tuning Guide

The performance is essential for on-policy RL algorithm. We have written a detailed performance tuning guide to help you optimize performance.

Upgrade to vLLM >= v0.8.2

verl now supports vLLM>=0.8.2 when using FSDP as the training backend. Please refer to this document for the installation guide and more information. Please avoid vllm 0.7.x, which contains bugs that may lead to OOMs and unexpected errors.

Use Latest SGLang

SGLang is fully supported with verl, and SGLang RL Group is working extensively on building unique features, including multi-turn agentic RL, VLM RLHF, server-based RL, and partial rollout. Please refer to this document for the installation guide and more information.

Upgrade to FSDP2

verl is fully embracing FSDP2! FSDP2 is recommended by torch distributed team, providing better throughput and memory usage, and is composible with other features (e.g. torch.compile). To enable FSDP2, simply use verl main and set the following options:

actor_rollout_ref.ref.strategy=fsdp2
actor_rollout_ref.actor.strategy=fsdp2
critic.strategy=fsdp2
reward_model.strategy=fsdp2

Furthermore, FSDP2 cpu offloading is compatible with gradient accumulation. You can turn it on to save memory with actor_rollout_ref.actor.fsdp_config.offload_policy=True. For more details, see https://github.com/volcengine/verl/pull/1026

AMD Support (ROCm Kernel)

verl now supports FSDP as the training engine (Megatron support coming soon) and both integrates with vLLM and SGLang as inference engines. Please refer to this document for the installation guide and more information, and this document for the vLLM performance tuning for ROCm.

Citation and acknowledgement

If you find the project helpful, please cite:

@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 contributed by Bytedance, Anyscale, LMSys.org, Alibaba Qwen team, Shanghai AI Lab, Tsinghua University, UC Berkeley, UCLA, UIUC, University of Hong Kong, ke.com, All Hands AI, ModelBest, JD AI Lab, Microsoft Research, StepFun, Amazon, LinkedIn, Meituan, Camel-AI, OpenManus, Xiaomi, NVIDIA research, Baichuan, RedNote, SwissAI, Moonshot AI (Kimi), Baidu, Snowflake, Skywork.ai, JetBrains, IceSword Lab, and many more.

Awesome work using verl

TinyZero: a reproduction of DeepSeek R1 Zero recipe for reasoning tasks
SkyThought: RL training for Sky-T1-7B by NovaSky AI team.
simpleRL-reason: SimpleRL-Zoo: Investigating and Taming Zero Reinforcement Learning for Open Base Models in the Wild
Easy-R1: Multi-modal RL training framework
OpenManus-RL: LLM Agents RL tuning framework for multiple agent environments.
rllm: async RL training with verl-pipeline
RAGEN: a general-purpose reasoning agent training framework
Search-R1: RL with reasoning and searching (tool-call) interleaved LLMs
ReSearch: Learning to Reason with Search for LLMs via Reinforcement Learning
Skywork-OR1: Skywork open reaonser series
ToRL: Scaling tool-integrated RL
Absolute Zero Reasoner: A no human curated data self-play framework for reasoning
verl-agent: A scalable training framework for long-horizon LLM/VLM agents, along with a new algorithm GiGPO
RL-Factory: An easy and efficient RL post-training framework for Agentic Learning
ReTool: ReTool: reinforcement learning for strategic tool use in LLMs. Code release is in progress…
verl-tool: An unified and easy-to-extend tool-agent training framework based on verl
PRIME: Process reinforcement through implicit rewards
MemAgent: MemAgent: Reshaping Long-Context LLM with Multi-Conv RL based Memory Agent
POLARIS: A Post-training recipe for scaling RL on Advanced Reasoning models
GUI-R1: GUI-R1: A Generalist R1-style Vision-Language Action Model For GUI Agents
DeepRetrieval: RL Training of Search Agent with Search/Retrieval Outcome
Code-R1: Reproducing R1 for Code with Reliable Rewards
DeepResearcher: Scaling deep research via reinforcement learning in real-world environments
VAGEN: Training VLM agents with multi-turn reinforcement learning
RM-R1: RL training of reasoning reward models
LUFFY: Learning to Reason under Off-Policy Guidance
DeepMath: DeepMath-103K data and series models for math reasoning
PACS: Implicit Actor Critic Coupling via a Supervised Learning Framework for RLVR
Entropy Mechanism of RL: The Entropy Mechanism of Reinforcement Learning for Large Language Model Reasoning
LLaSA-TTS-GRPO: TTS fine-tuning with GRPO optimization based on LLASA models
PF-PPO: Policy Filtration for PPO based on the reliability of reward signals for more efficient and robust RLHF.
RACRO: Build multi-modal reasoning models via decoupling it into query-conditioned captioning and text-only reasoning
Agent Lightning: A flexible and extensible framework that enables seamless agent optimization for any existing agent framework.
VTool-R1: VLMs Learn to Think with Images via Reinforcement Learning on Multimodal Tool Use.
Kimina-Prover-RL: Training pipeline for formal theorem proving, based on a paradigm inspired by DeepSeek-R1.
RL-PLUS: Countering Capability Boundary Collapse of LLMs in Reinforcement Learning with Hybrid-policy Optimization.
rStar2-Agent: Using reinforcement learning with multi-step tool-calling for math tasks, rStar2-Agent-14B reaches frontier-level math reasoning in just 510 RL training steps
Vision-SR1: Self-Rewarding Vision-Language Model via Reasoning Decomposition
SimpleVLA-RL: SimpleVLA-RL: A Simple yet Effective Vision-Language Action Model for Reinforcement Learning
Table-R1: Table-R1: Inference-Time Scaling for Table Reasoning
Revisual-R1: Revisual-R1: Advancing Multimodal Reasoning From Optimized Cold Start to Staged Reinforcement Learning
ARES: ARES: Multimodal Adaptive Reasoning via Difficulty-Aware Token-Level Entropy Shaping
Meta-Bandit-LLM: Meta-Bandit-LLM: Long-horizon multiturn interactive training for meta-bandit agents
PokeeResearch: PokeeResearch: State-of-the-art 7B DeepResearch Agent that leverages web search and content reading capabilities to answer complex questions using the most up-to-date information available online.
Search Self-play: Pushing the Frontier of Agent Capability without Supervision
OneThinker: All-in-one Reasoning Model for Image and Video
OpenTinker: Democratizing Agentic Reinforcement Learning as a Service

and many more awesome work listed in recipe.

Contribution Guide

See contributions guide

About ByteDance Seed Team

Founded in 2023, ByteDance Seed Team is dedicated to crafting the industry’s most advanced AI foundation models. The team aspires to become a world-class research team and make significant contributions to the advancement of science and society. You can get to know Bytedance Seed better through the following channels👇