AGENT SKILL

verl-rl-training

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.

License 许可
MIT
最近更新
2026-11-10
Stars
★ 1
Forks
0

01 前言 Front Matter

name
verl-rl-training
tags
[ "Reinforcement Learning", "RLHF", "GRPO", "PPO", "Post-Training", "Distributed Training" ]
author
Orchestra Research
license
MIT
version
1.0.0
description
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.
dependencies
[ "verl>=0.3.0", "torch>=2.0.0", "ray>=2.41.0", "vllm>=0.8.2", "transformers>=4.40.0" ]

02 资源文件 2

03 SKILL.md 正文

verl: Volcano Engine Reinforcement Learning for LLMs

verl is a flexible, efficient, and production-ready RL training library for large language models from ByteDance's Seed team. It implements the HybridFlow framework (EuroSys 2025) and powers models like Doubao-1.5-pro achieving O1-level performance on math benchmarks.

When to Use verl

Choose verl when you need:

  • 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

Consider alternatives when:

  • 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

Key Features

  • Training backends: FSDP, FSDP2, Megatron-LM
  • Rollout engines: vLLM, SGLang, HuggingFace Transformers
  • Algorithms: PPO, GRPO, DAPO, RLOO, ReMax, REINFORCE++, SPIN, SPPO
  • Models: Qwen-3, Llama-3.1, DeepSeek, Gemma-2 (0.5B to 671B)
  • Advanced: LoRA RL, sequence parallelism, expert parallelism, multi-turn tools

Installation

# Option 1: pip install
pip install verl[vllm]  # or verl[sglang] for SGLang backend

# Option 2: Docker (recommended for production)
docker pull verlai/verl:vllm011.latest

# Option 3: From source
git clone https://github.com/volcengine/verl.git
cd verl && pip install -e .[vllm,math]

Quick Start: GRPO Training

python3 -m verl.trainer.main_ppo \
    algorithm.adv_estimator=grpo \
    data.train_files=~/data/gsm8k/train.parquet \
    actor_rollout_ref.model.path=Qwen/Qwen2.5-7B \
    actor_rollout_ref.rollout.n=8 \
    actor_rollout_ref.actor.use_kl_loss=True \
    trainer.n_gpus_per_node=8

Core Architecture

verl uses a HybridFlow programming model separating control flow from computation:

┌─────────────────────────────────────────────────────────┐
│ Single-Process Controller (Ray)                         │
│ - Orchestrates: rollout → reward → train → sync        │
└─────────────────────┬───────────────────────────────────┘
                      │
┌─────────────────────▼───────────────────────────────────┐
│ Multi-Process Workers                                   │
│ ├── ActorRolloutRefWorker (policy + generation)        │
│ ├── CriticWorker (value estimation, PPO only)          │
│ └── RewardManager (model-based or rule-based rewards)  │
└─────────────────────────────────────────────────────────┘

Workflow 1: Math Reasoning with GRPO

Use this workflow for training reasoning models on math tasks like GSM8K or MATH.

Prerequisites Checklist

  • [ ] GPU cluster with 8+ GPUs (H100 recommended)
  • [ ] Dataset in parquet format with prompt and reward_model columns
  • [ ] Base model from HuggingFace Hub

Step 1: Prepare Dataset

import pandas as pd

data = [
    {
        "prompt": [{"role": "user", "content": "What is 15 + 27?"}],
        "reward_model": {"ground_truth": "42"}
    },
    # ... more examples
]
df = pd.DataFrame(data)
df.to_parquet("train.parquet")

Step 2: Define Reward Function

# reward_function.py
import re

def compute_reward(responses, ground_truths):
    rewards = []
    for response, gt in zip(responses, ground_truths):
        # Extract answer from response
        match = re.search(r'\\boxed{([^}]+)}', response)
        if match and match.group(1).strip() == gt.strip():
            rewards.append(1.0)
        else:
            rewards.append(0.0)
    return rewards

Step 3: Create Training Config

# config/grpo_math.yaml
algorithm:
  adv_estimator: grpo
  gamma: 1.0
  lam: 1.0

data:
  train_files: /path/to/train.parquet
  val_files: /path/to/val.parquet
  train_batch_size: 256
  max_prompt_length: 512
  max_response_length: 2048

actor_rollout_ref:
  model:
    path: Qwen/Qwen2.5-7B-Instruct
  actor:
    use_kl_loss: true
    kl_loss_coef: 0.001
    ppo_mini_batch_size: 64
  rollout:
    name: vllm
    n: 8  # samples per prompt
    temperature: 0.7
    top_p: 0.95

trainer:
  total_epochs: 3
  n_gpus_per_node: 8
  save_freq: 100

Step 4: Launch Training

python3 -m verl.trainer.main_ppo \
    --config-path config \
    --config-name grpo_math \
    trainer.experiment_name=grpo_math_qwen7b

Step 5: Monitor and Validate

  • [ ] Check WandB/TensorBoard for loss curves
  • [ ] Verify reward is increasing over steps
  • [ ] Run evaluation on held-out test set

Workflow 2: PPO with Critic Model

Use this workflow when you need value-based advantage estimation (GAE).

Key Differences from GRPO

  • Requires separate critic model
  • Uses Generalized Advantage Estimation (GAE)
  • Better for tasks with dense rewards

Configuration

algorithm:
  adv_estimator: gae  # Use GAE instead of GRPO
  gamma: 0.99
  lam: 0.95

critic:
  model:
    path: Qwen/Qwen2.5-7B-Instruct  # Can be same or different from actor
  ppo_mini_batch_size: 64

actor_rollout_ref:
  actor:
    use_kl_loss: true
    kl_loss_coef: 0.02
    clip_ratio: 0.2  # PPO clipping

Launch with Critic

python3 -m verl.trainer.main_ppo \
    algorithm.adv_estimator=gae \
    critic.model.path=Qwen/Qwen2.5-7B-Instruct \
    trainer.n_gpus_per_node=8

Workflow 3: Large-Scale Training with Megatron

Use this workflow for models >70B parameters or when you need expert parallelism.

Prerequisites

  • [ ] Install Megatron-LM bridge: pip install mbridge
  • [ ] Convert model to Megatron format
  • [ ] Multi-node cluster with NVLink/InfiniBand

Configuration for 70B+ Models

actor_rollout_ref:
  model:
    path: /path/to/megatron/checkpoint
    backend: megatron
  actor:
    strategy: megatron
    tensor_model_parallel_size: 8
    pipeline_model_parallel_size: 2
  rollout:
    name: vllm
    tensor_parallel_size: 8

Launch Multi-Node

# On head node
ray start --head --port=6379

# On worker nodes
ray start --address='head_ip:6379'

# Launch training
python3 -m verl.trainer.main_ppo \
    trainer.nnodes=4 \
    trainer.n_gpus_per_node=8

Configuration Reference

Algorithm Selection

Algorithm adv_estimator Use Case
GRPO grpo Critic-free, math/reasoning
PPO/GAE gae Dense rewards, value estimation
REINFORCE++ reinforce_plus_plus Variance reduction
RLOO rloo Leave-one-out baseline
ReMax remax Maximum reward baseline
OPO opo Optimal policy optimization

Key Parameters

# Rollout parameters
actor_rollout_ref.rollout.n: 8              # Samples per prompt
actor_rollout_ref.rollout.temperature: 0.7  # Sampling temperature
actor_rollout_ref.rollout.top_p: 0.95       # Nucleus sampling

# Training parameters
actor_rollout_ref.actor.lr: 1e-6            # Learning rate
actor_rollout_ref.actor.ppo_mini_batch_size: 64
actor_rollout_ref.actor.clip_ratio: 0.2     # PPO clip range

# KL control
actor_rollout_ref.actor.use_kl_loss: true
actor_rollout_ref.actor.kl_loss_coef: 0.001
algorithm.kl_ctrl.target_kl: 0.1            # For adaptive KL control

Common Issues and Solutions

Issue: OOM During Rollout

Symptoms: CUDA out of memory during generation phase

Solutions:

# Reduce batch size
actor_rollout_ref.rollout.log_prob_micro_batch_size: 4

# Enable gradient checkpointing
actor_rollout_ref.model.enable_gradient_checkpointing: true

# Use FSDP2 with CPU offloading
actor_rollout_ref.actor.strategy: fsdp2
actor_rollout_ref.actor.fsdp_config.offload_policy: true

Issue: Training Instability

Symptoms: Loss spikes, reward collapse

Solutions:

# Reduce learning rate
actor_rollout_ref.actor.lr: 5e-7

# Increase KL penalty
actor_rollout_ref.actor.kl_loss_coef: 0.01

# Enable gradient clipping
actor_rollout_ref.actor.max_grad_norm: 1.0

Issue: Slow Weight Sync

Symptoms: Long pauses between rollout and training

Solutions:

# Use FSDP2 for faster resharding
actor_rollout_ref.actor.strategy=fsdp2

# Enable async weight transfer
trainer.async_weight_update=true

Issue: vLLM Version Mismatch

Symptoms: Import errors or generation failures

Solution: Use compatible versions:

pip install vllm>=0.8.5,<=0.12.0
# Avoid vLLM 0.7.x (known bugs)

Advanced Topics

Multi-Turn Tool Calling

See references/multi-turn.md for agentic workflows with tool use.

Vision-Language Models

actor_rollout_ref:
  model:
    path: Qwen/Qwen2.5-VL-7B-Instruct
  rollout:
    name: vllm
    enable_vision: true

LoRA Training

actor_rollout_ref:
  actor:
    lora:
      enabled: true
      r: 16
      alpha: 32
      target_modules: ["q_proj", "v_proj"]

Resources

04 同仓库更多 Skills

YYC-Cube
YYC-Cube/YYC3-Knowledge-Base

vercel-sandbox

Run agent-browser + Chrome inside Vercel Sandbox microVMs for browser automation from any Vercel-deployed app. Use when the user needs browser automation in a Vercel app (Next.js, SvelteKit, Nuxt, Remix, Astro, etc.), wants to run headless Chrome without binary size limits, needs persistent browser sessions across commands, or wants ephemeral isolated browser environments. Triggers include "Vercel Sandbox browser", "microVM Chrome", "agent-browser in sandbox", "browser automation on Vercel", or any task requiring Chrome in a Vercel Sandbox.

★ 1
YYC-Cube
YYC-Cube/YYC3-Knowledge-Base

slack

Interact with Slack workspaces using browser automation. Use when the user needs to check unread channels, navigate Slack, send messages, extract data, find information, search conversations, or automate any Slack task. Triggers include "check my Slack", "what channels have unreads", "send a message to", "search Slack for", "extract from Slack", "find who said", or any task requiring programmatic Slack interaction.

★ 1
YYC-Cube
YYC-Cube/YYC3-Knowledge-Base

electron

Automate Electron desktop apps (VS Code, Slack, Discord, Figma, Notion, Spotify, etc.) using agent-browser via Chrome DevTools Protocol. Use when the user needs to interact with an Electron app, automate a desktop app, connect to a running app, control a native app, or test an Electron application. Triggers include "automate Slack app", "control VS Code", "interact with Discord app", "test this Electron app", "connect to desktop app", or any task requiring automation of a native Electron application.

★ 1
YYC-Cube
YYC-Cube/YYC3-Knowledge-Base

dogfood

Systematically explore and test a web application to find bugs, UX issues, and other problems. Use when asked to "dogfood", "QA", "exploratory test", "find issues", "bug hunt", "test this app/site/platform", or review the quality of a web application. Produces a structured report with full reproduction evidence -- step-by-step screenshots, repro videos, and detailed repro steps for every issue -- so findings can be handed directly to the responsible teams.

★ 1
YYC-Cube
YYC-Cube/YYC3-Knowledge-Base

agent-browser

Browser automation CLI for AI agents. Use when the user needs to interact with websites, including navigating pages, filling forms, clicking buttons, taking screenshots, extracting data, testing web apps, or automating any browser task. Triggers include requests to "open a website", "fill out a form", "click a button", "take a screenshot", "scrape data from a page", "test this web app", "login to a site", "automate browser actions", or any task requiring programmatic web interaction.

★ 1
YYC-Cube
YYC-Cube/YYC3-Knowledge-Base

shadcn

Manages shadcn components and projects — adding, searching, fixing, debugging, styling, and composing UI, including chat interfaces. Provides project context, component docs, and usage examples. Applies when working with shadcn/ui, component registries, presets, --preset codes, or any project with a components.json file. Also triggers for "shadcn init", "create an app with --preset", "switch to --preset", and for building a design system or applying a DESIGN.md or brand spec to shadcn/ui.

★ 1