lazyllm-verl 0.3.2.dev1

An Efficient, Scalable, Multi-Modality RL Training Framework based on veRL

EasyR1: An Efficient, Scalable, Multi-Modality RL Training Framework

Used by Amazon Web Services

This project is a clean fork of the original veRL project to support vision language models, we thank all the authors for providing such a high-performance RL training framework.

EasyR1 is efficient and scalable due to the design of HybirdEngine and the latest release of vLLM's SPMD mode.

Features

• Supported models

  • Llama3/Qwen2/Qwen2.5/Qwen3 language models
  • Qwen2/Qwen2.5-VL vision language models
  • DeepSeek-R1 distill models

• Supported algorithms

  • GRPO
  • DAPO
  • Reinforce++
  • ReMax
  • RLOO

• Supported datasets

  • Any text, vision-text dataset in a specific format

• Supported tricks

  • Padding-free training
  • Resuming from the latest/best checkpoint
  • Wandb & SwanLab & Mlflow & Tensorboard tracking

Requirements

Software Requirements

• Python 3.9+
• transformers>=4.54.0
• flash-attn>=2.4.3
• vllm>=0.8.3

We provide a Dockerfile to easily build environments.

We recommend using the pre-built docker image in EasyR1.

docker pull hiyouga/verl:ngc-th2.7.1-cu12.6-vllm0.10.0
docker run -it --ipc=host --gpus=all hiyouga/verl:ngc-th2.7.1-cu12.6-vllm0.10.0

If your environment does not support Docker, you can consider using Apptainer:

apptainer pull easyr1.sif docker://hiyouga/verl:ngc-th2.7.1-cu12.6-vllm0.10.0
apptainer shell --nv --cleanenv --bind /mnt/your_dir:/mnt/your_dir easyr1.sif

Use USE_MODELSCOPE_HUB=1 to download models from the ModelScope hub.

Hardware Requirements

* estimated

Method	Bits	1.5B	3B	7B	32B	72B
GRPO Full Fine-Tuning	AMP	2*24GB	4*40GB	8*40GB	16*80GB	32*80GB
GRPO Full Fine-Tuning	BF16	1*24GB	1*40GB	4*40GB	8*80GB	16*80GB

[!NOTE] Use worker.actor.fsdp.torch_dtype=bf16 and worker.actor.optim.strategy=adamw_bf16 to enable bf16 training.

We are working hard to reduce the VRAM in RL training, LoRA support will be integrated in next updates.

Tutorial: Run Qwen2.5-VL GRPO on Geometry3K Dataset in Just 3 Steps

Installation

git clone https://github.com/hiyouga/EasyR1.git
cd EasyR1
pip install -e .

GRPO Training

bash examples/qwen2_5_vl_7b_geo3k_grpo.sh

Merge Checkpoint in Hugging Face Format

python3 scripts/model_merger.py --local_dir checkpoints/easy_r1/exp_name/global_step_1/actor

[!TIP] If you encounter issues with connecting to Hugging Face, consider using export HF_ENDPOINT=https://hf-mirror.com.

If you want to use SwanLab logger, consider using bash examples/qwen2_5_vl_7b_geo3k_swanlab.sh.

Custom Dataset

Please refer to the example datasets to prepare your own dataset.

• Text dataset: https://huggingface.co/datasets/hiyouga/math12k
• Image-text dataset: https://huggingface.co/datasets/hiyouga/geometry3k
• Multi-image-text dataset: https://huggingface.co/datasets/hiyouga/journeybench-multi-image-vqa
• Text-image mixed dataset: https://huggingface.co/datasets/hiyouga/rl-mixed-dataset

How to Understand GRPO in EasyR1

• To learn about the GRPO algorithm, you can refer to Hugging Face's blog.

How to Run 70B+ Model in Multi-node Environment

1. Start the Ray head node.

ray start --head --port=6379 --dashboard-host=0.0.0.0

2. Start the Ray worker node and connect to the head node.

ray start --address=<head_node_ip>:6379

3. Check the Ray resource pool.

ray status

4. Run training script on the Ray head node only.

bash examples/qwen2_5_vl_7b_geo3k_grpo.sh

See the veRL's official doc for more details about multi-node training and Ray debugger.

Other Baselines

We also reproduced the following two baselines of the R1-V project.

• CLEVR-70k-Counting: Train the Qwen2.5-VL-3B-Instruct model on counting problem.
• GeoQA-8k: Train the Qwen2.5-VL-3B-Instruct model on GeoQA problem.

Performance Baselines

See baselines.md.

Awesome Work using EasyR1

• MMR1: Advancing the Frontiers of Multimodal Reasoning.
• Vision-R1: Incentivizing Reasoning Capability in Multimodal Large Language Models.
• Seg-Zero: Reasoning-Chain Guided Segmentation via Cognitive Reinforcement.
• MetaSpatial: Reinforcing 3D Spatial Reasoning in VLMs for the Metaverse.
• Temporal-R1: Envolving Temporal Reasoning Capability into LMMs via Temporal Consistent Reward.
• NoisyRollout: Reinforcing Visual Reasoning with Data Augmentation.
• GUI-R1: A Generalist R1-Style Vision-Language Action Model For GUI Agents.
• R1-Track: Direct Application of MLLMs to Visual Object Tracking via Reinforcement Learning.
• VisionReasoner: Unified Visual Perception and Reasoning via Reinforcement Learning.
• MM-UPT: Unsupervised Post-Training for Multi-Modal LLM Reasoning via GRPO.
• RL-with-Cold-Start: Advancing Multimodal Reasoning via Reinforcement Learning with Cold Start.
• ViGoRL: Grounded Reinforcement Learning for Visual Reasoning.
• Revisual-R1: Advancing Multimodal Reasoning: From Optimized Cold Start to Staged Reinforcement Learning.
• SophiaVL-R1: Reinforcing MLLMs Reasoning with Thinking Reward.
• Vision-Matters: Simple Visual Perturbations Can Boost Multimodal Math Reasoning.
• VTool-R1: VLMs Learn to Think with Images via Reinforcement Learning on Multimodal Tool Use.
• Long-RL: Scaling RL to Long Sequences.

TODO

• Support LoRA (high priority).
• Support ulysses parallelism for VLMs (middle priority).
• Support more VLM architectures.

[!NOTE] We will not provide scripts for supervised fine-tuning and inference in this project. If you have such requirements, we recommend using LLaMA-Factory.

Known bugs

These features are temporarily disabled for now, we plan to fix them one-by-one in the future updates.

• Vision language models are not compatible with ulysses parallelism yet.

Discussion Group

👋 Join our WeChat group.

FAQs

ValueError: Image features and image tokens do not match: tokens: 8192, features 9800

Increase the data.max_prompt_length or reduce the data.max_pixels.

RuntimeError: CUDA Error: out of memory at /workspace/csrc/cumem_allocator.cpp:62

Reduce the worker.rollout.gpu_memory_utilization and enable worker.actor.offload.offload_params.

RuntimeError: 0 active drivers ([]). There should only be one.

Uninstall deepspeed from the current python environment.

Citation

Core contributors: Yaowei Zheng, Junting Lu, Shenzhi Wang, Zhangchi Feng, Dongdong Kuang and Yuwen Xiong

We also thank Guangming Sheng and Chi Zhang for helpful discussions.

@misc{zheng2025easyr1,
  title        = {EasyR1: An Efficient, Scalable, Multi-Modality RL Training Framework},
  author       = {Yaowei Zheng, Junting Lu, Shenzhi Wang, Zhangchi Feng, Dongdong Kuang, Yuwen Xiong},
  howpublished = {\url{https://github.com/hiyouga/EasyR1}},
  year         = {2025}
}

We recommend to also cite the original work.

@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}
}