verl-rl-training - RL Training Library for LLMs
Provides guidance for training LLMs with RL using verl library
Tags
Updated: 2026-06-30Capabilities
Typical Inputs
Typical Outputs
What this skill does
- train models with RLHF
- train models with GRPO
- train models with PPO
- train models with RLOO
- train models with REINFORCE++
- train models with DAPO
- train models with ReMax
- train models with SPIN
- train models with SPPO
- switch training backends
- perform rollout generation
- compute rewards
- train vision-language models
- execute multi-turn rollout
- call tools during rollout
- use FSDP backend
- use Megatron-LM backend
- use vLLM backend
- use SGLang backend
- configure LoRA training
Inputs
- training dataset
- base model
- reward function
- training configuration
- GPU cluster
- model checkpoint
Outputs
- trained model checkpoint
- training metrics
- evaluation results
Requirements
- Python environment
- verl>=0.3.0
- torch>=2.0.0
- ray>=2.41.0
- vllm>=0.8.2
- transformers>=4.40.0
- GPU cluster with 8+ GPUs