verl-rl-training
Category: ml-training · License: MIT · Version: 1.0.0
Tags: Reinforcement Learning, RLHF, GRPO, PPO, Post-Training, Distributed Training
When to use it
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.Example requests
Ask your agent in plain language, for example:- “How do I set up GRPO training on GSM8K with verl?”
- “I’m running verl on 2 nodes of 8xH100 to RL-train Qwen3-32B on my math dataset at /data/math_train.parquet. During the rollout phase I get ‘CUDA out of memory’ from vLLM, and the pauses between rollout and training are very long. Config is FSDP2 with vLLM rollout and gpu_memory_utilization=0.9. What should I change?”
- “I want to post-train an open LLM to solve competition math using only a programmatic right/wrong answer checker, with no value model, on a multi-node GPU cluster where I can swap the inference engine and the training backend (FSDP vs Megatron) without rewriting my code. What framework and setup would you recommend?”
Install
/plugin marketplace add KalarisLabs/research-agent-skills and install the whole ml-training category with /plugin install ml-training@research-agent-skills.