domains / families / research / verl

verl

separate

Provides guidance for training LLMs with reinforcement learning using verl (Volcano Engine RL). Use when implementing RLHF, GRPO, PPO, or other RL algorithms for LLM post-training at scale with flexible infrastructure backends.

When to use

• Production-ready RL training at scale (tested up to 671B parameters)
• Flexibility to swap backends (FSDP ↔ Megatron-LM ↔ vLLM ↔ SGLang)
• Support for multiple RL algorithms (PPO, GRPO, RLOO, REINFORCE++, DAPO)
• Multi-turn rollout with tool calling for agentic workflows
• Vision-language model RL training
• You need Megatron-native training → use **slime** or **miles**
• You want PyTorch-native abstractions with Monarch → use **torchforge**
• You only need simple SFT/DPO → use **TRL** or **Axolotl**

When not to use

• Megatron-native training
• PyTorch-native abstractions with Monarch
• Simple SFT/DPO

Skill page

Source