verl

What it is and what it does

verl is an open-source reinforcement learning training library for large language models, initiated by ByteDance and maintained by the verl community. It provides a flexible programming model for building RL dataflows such as PPO and GRPO with modular APIs that decouple computation and data dependencies, enabling seamless integration with existing LLM infrastructure like FSDP, Megatron-LM, vLLM, and SGLang. The library supports flexible device mapping across different GPU configurations for efficient resource utilization and scalability.

The package is designed for production use in post-training LLMs with state-of-the-art throughput through efficient actor model resharding and integration with modern LLM training and inference engines. It includes ready integration with popular HuggingFace models and depends on a substantial ecosystem of ML libraries including transformers, ray, tensorboard, wandb, and PyTorch-related packages for distributed training and monitoring.

Use it for:

• Train reward-optimized language models using PPO or GRPO algorithms with flexible dataflow composition
• Integrate RL post-training into existing LLM infrastructure built on FSDP, Megatron-LM, or vLLM without rewriting core frameworks
• Distribute RL training across heterogeneous GPU clusters with flexible device mapping and model resharding
• Build custom RL algorithms by extending the hybrid-controller programming model for domain-specific post-training
• Monitor and observe distributed RL training workloads with integrated observability for training metrics and state traces

Worth the install?

AI-flagged interpretation of the facts on this page — verify before relying

verl is a reinforcement learning training library for large language models that implements flexible RL algorithms like PPO and GRPO with modular APIs for integration into existing LLM frameworks.

Yes, if you are building or fine-tuning large language models with reinforcement learning and have GPU infrastructure available. The library is actively maintained, has no known vulnerabilities, uses a permissive license, and integrates with standard LLM frameworks. Install friction is low, but practical use requires substantial ML dependencies and GPU resources. Not suitable for CPU-only environments or if you need RL training for non-LLM tasks.

Install

verl on PyPI

pip install verl

uv add verl

poetry add verl

Installing verl

Before you install

Low install friction with a pure-Python wheel. Active maintenance with recent releases and 22957 GitHub stars. Requires Python 3.10+. Depends on 18 runtime packages including heavy ML libraries (transformers, ray, tensorboard, wandb), so total environment setup will be substantial despite the package itself being lightweight.

License in practice

Apache-2.0 is permissive, allowing commercial use, modification, and distribution with minimal restrictions. Suitable for proprietary projects provided you include the license notice.

Quickstart

pip install verl

import verl
from verl import Actor, Critic
# Configure and run RL training with modular components

Requires Python 3.10 or later. Practical use requires GPU resources and familiarity with RL concepts; many runtime dependencies (transformers, ray, tensorboard) must be installed and properly configured.

Verify before relying

• Whether the 18 runtime dependencies are all strictly required or if some are optional for specific use cases
• Performance characteristics and throughput claims relative to other RL training frameworks
• Minimum GPU memory and cluster size requirements for practical training runs
• Whether all HuggingFace model architectures are supported or only a subset

Package facts

Evidence: verl-0.9.0-py3-none-any.whl

Tags

More Artificial Intelligence packages

Further reading