verl-project/verl

Health v2 · maintenance only

Adoption not part of the score

23828 stars · 4735 forks observed · 2026-10-12

What it is AI-extracted, prompt v1, taxonomy v1, 2026-10-12, confidence 0.97

Verl is a flexible, efficient, and production‑ready reinforcement learning library for large language models. It provides a modular API to integrate with existing LLM infrastructures and supports advanced RL algorithms like PPO and GRPO.

Use cases

• fine‑tune large language models with reinforcement learning from human feedback
• run scalable RL training pipelines for LLMs across multiple GPUs
• experiment with custom RL algorithms on transformer models
• integrate RL training into existing LLM frameworks such as Megatron‑LM or vLLM
• perform high‑throughput RL dataflow execution for RLHF tasks

When to choose

• You need a production‑grade RLHF framework that works with popular LLM toolkits.
• Your training workload requires flexible device mapping and multi‑GPU scaling.
• You want to extend or customize RL algorithms without rewriting low‑level code.

When to avoid

• You are looking for a lightweight, single‑node RL library with minimal dependencies.
• Your project does not involve large language models or RLHF.
• You need support for languages other than Python.

Facets

library · maturity active

reinforcement-learning agent-framework llm-training machine-learning artificial-intelligence machine-learning large-language-models reinforcement-learning python windows llm rlhf hybridflow distributed-training torch federated scalable distributed linux macos gpu

Member repositories

For agents

markdown · JSON · MCP: product_card(name="verl-project/verl")