[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M6KQS92VEC68HR31NWBAXCXQ":3,"skill-same:01M6KQS92VEC68HR31NWBAXCXQ":85},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":15,"categories":16,"keywords":17,"publishedAt":39,"linkOnly":40,"markdownContent":41,"markdownBody":42,"frontMatter":43,"refName":55,"commitSha":56,"skillDirectory":57,"forksCount":13,"repositoryHtmlUrl":58,"repositoryHomepageUrl":6,"ownerHtmlUrl":59,"ownerBlogUrl":6,"ownerDescription":6,"updatedAt":60,"install":61,"resources":74},"01M6KQS92VEC68HR31NWBAXCXQ","verl-rl-training",null,"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.","endman100\u002Fskill-router-nexus","endman100","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F31997019?v=4","01M6KPDX17BYDWW4RRB0TNGGMX","01M6KPDX64Y2QBEWVFG5W5PWPF",0,"MIT","https:\u002F\u002Fgithub.com\u002Fendman100\u002Fskill-router-nexus\u002Fblob\u002Fmain\u002FAI-Training\u002Fverl\u002FSKILL.md",[],[18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38],"training","verl rl training","verl","rl","use","when","for","implementing","orchestra","at","volcano","algorithms","llms","learning","Transformers","Ray","PPO","RLHF","Post-Training","Distributed Training","infrastructure","2026-11-10 10:12:30",false,"---\nname: verl-rl-training\ndescription: 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.\nversion: 1.0.0\nauthor: Orchestra Research\nlicense: MIT\ntags: [Reinforcement Learning, RLHF, GRPO, PPO, Post-Training, Distributed Training]\ndependencies: [verl>=0.3.0, torch>=2.0.0, ray>=2.41.0, vllm>=0.8.2, transformers>=4.40.0]\n---\n\n# verl: Volcano Engine Reinforcement Learning for LLMs\n\nverl 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.\n\n## When to Use verl\n\n**Choose verl when you need:**\n- Production-ready RL training at scale (tested up to 671B parameters)\n- Flexibility to swap backends (FSDP ↔ Megatron-LM ↔ vLLM ↔ SGLang)\n- Support for multiple RL algorithms (PPO, GRPO, RLOO, REINFORCE++, DAPO)\n- Multi-turn rollout with tool calling for agentic workflows\n- Vision-language model RL training\n\n**Consider alternatives when:**\n- You need Megatron-native training → use **slime** or **miles**\n- You want PyTorch-native abstractions with Monarch → use **torchforge**\n- You only need simple SFT\u002FDPO → use **TRL** or **Axolotl**\n\n## Key Features\n\n- **Training backends**: FSDP, FSDP2, Megatron-LM\n- **Rollout engines**: vLLM, SGLang, HuggingFace Transformers\n- **Algorithms**: PPO, GRPO, DAPO, RLOO, ReMax, REINFORCE++, SPIN, SPPO\n- **Models**: Qwen-3, Llama-3.1, DeepSeek, Gemma-2 (0.5B to 671B)\n- **Advanced**: LoRA RL, sequence parallelism, expert parallelism, multi-turn tools\n\n## Installation\n\n```bash\n# Option 1: pip install\npip install verl[vllm]  # or verl[sglang] for SGLang backend\n\n# Option 2: Docker (recommended for production)\ndocker pull verlai\u002Fverl:vllm011.latest\n\n# Option 3: From source\ngit clone https:\u002F\u002Fgithub.com\u002Fvolcengine\u002Fverl.git\ncd verl && pip install -e .[vllm,math]\n```\n\n## Quick Start: GRPO Training\n\n```bash\npython3 -m verl.trainer.main_ppo \\\n    algorithm.adv_estimator=grpo \\\n    data.train_files=~\u002Fdata\u002Fgsm8k\u002Ftrain.parquet \\\n    actor_rollout_ref.model.path=Qwen\u002FQwen2.5-7B \\\n    actor_rollout_ref.rollout.n=8 \\\n    actor_rollout_ref.actor.use_kl_loss=True \\\n    trainer.n_gpus_per_node=8\n```\n\n## Core Architecture\n\nverl uses a **HybridFlow** programming model separating control flow from computation:\n\n```\n┌─────────────────────────────────────────────────────────┐\n│ Single-Process Controller (Ray)                         │\n│ - Orchestrates: rollout → reward → train → sync        │\n└─────────────────────┬───────────────────────────────────┘\n                      │\n┌─────────────────────▼───────────────────────────────────┐\n│ Multi-Process Workers                                   │\n│ ├── ActorRolloutRefWorker (policy + generation)        │\n│ ├── CriticWorker (value estimation, PPO only)          │\n│ └── RewardManager (model-based or rule-based rewards)  │\n└─────────────────────────────────────────────────────────┘\n```\n\n---\n\n## Workflow 1: Math Reasoning with GRPO\n\nUse this workflow for training reasoning models on math tasks like GSM8K or MATH.\n\n### Prerequisites Checklist\n- [ ] GPU cluster with 8+ GPUs (H100 recommended)\n- [ ] Dataset in parquet format with `prompt` and `reward_model` columns\n- [ ] Base model from HuggingFace Hub\n\n### Step 1: Prepare Dataset\n\n```python\nimport pandas as pd\n\ndata = [\n    {\n        \"prompt\": [{\"role\": \"user\", \"content\": \"What is 15 + 27?\"}],\n        \"reward_model\": {\"ground_truth\": \"42\"}\n    },\n    # ... more examples\n]\ndf = pd.DataFrame(data)\ndf.to_parquet(\"train.parquet\")\n```\n\n### Step 2: Define Reward Function\n\n```python\n# reward_function.py\nimport re\n\ndef compute_reward(responses, ground_truths):\n    rewards = []\n    for response, gt in zip(responses, ground_truths):\n        # Extract answer from response\n        match = re.search(r'\\\\boxed{([^}]+)}', response)\n        if match and match.group(1).strip() == gt.strip():\n            rewards.append(1.0)\n        else:\n            rewards.append(0.0)\n    return rewards\n```\n\n### Step 3: Create Training Config\n\n```yaml\n# config\u002Fgrpo_math.yaml\nalgorithm:\n  adv_estimator: grpo\n  gamma: 1.0\n  lam: 1.0\n\ndata:\n  train_files: \u002Fpath\u002Fto\u002Ftrain.parquet\n  val_files: \u002Fpath\u002Fto\u002Fval.parquet\n  train_batch_size: 256\n  max_prompt_length: 512\n  max_response_length: 2048\n\nactor_rollout_ref:\n  model:\n    path: Qwen\u002FQwen2.5-7B-Instruct\n  actor:\n    use_kl_loss: true\n    kl_loss_coef: 0.001\n    ppo_mini_batch_size: 64\n  rollout:\n    name: vllm\n    n: 8  # samples per prompt\n    temperature: 0.7\n    top_p: 0.95\n\ntrainer:\n  total_epochs: 3\n  n_gpus_per_node: 8\n  save_freq: 100\n```\n\n### Step 4: Launch Training\n\n```bash\npython3 -m verl.trainer.main_ppo \\\n    --config-path config \\\n    --config-name grpo_math \\\n    trainer.experiment_name=grpo_math_qwen7b\n```\n\n### Step 5: Monitor and Validate\n- [ ] Check WandB\u002FTensorBoard for loss curves\n- [ ] Verify reward is increasing over steps\n- [ ] Run evaluation on held-out test set\n\n---\n\n## Workflow 2: PPO with Critic Model\n\nUse this workflow when you need value-based advantage estimation (GAE).\n\n### Key Differences from GRPO\n- Requires separate critic model\n- Uses Generalized Advantage Estimation (GAE)\n- Better for tasks with dense rewards\n\n### Configuration\n\n```yaml\nalgorithm:\n  adv_estimator: gae  # Use GAE instead of GRPO\n  gamma: 0.99\n  lam: 0.95\n\ncritic:\n  model:\n    path: Qwen\u002FQwen2.5-7B-Instruct  # Can be same or different from actor\n  ppo_mini_batch_size: 64\n\nactor_rollout_ref:\n  actor:\n    use_kl_loss: true\n    kl_loss_coef: 0.02\n    clip_ratio: 0.2  # PPO clipping\n```\n\n### Launch with Critic\n\n```bash\npython3 -m verl.trainer.main_ppo \\\n    algorithm.adv_estimator=gae \\\n    critic.model.path=Qwen\u002FQwen2.5-7B-Instruct \\\n    trainer.n_gpus_per_node=8\n```\n\n---\n\n## Workflow 3: Large-Scale Training with Megatron\n\nUse this workflow for models >70B parameters or when you need expert parallelism.\n\n### Prerequisites\n- [ ] Install Megatron-LM bridge: `pip install mbridge`\n- [ ] Convert model to Megatron format\n- [ ] Multi-node cluster with NVLink\u002FInfiniBand\n\n### Configuration for 70B+ Models\n\n```yaml\nactor_rollout_ref:\n  model:\n    path: \u002Fpath\u002Fto\u002Fmegatron\u002Fcheckpoint\n    backend: megatron\n  actor:\n    strategy: megatron\n    tensor_model_parallel_size: 8\n    pipeline_model_parallel_size: 2\n  rollout:\n    name: vllm\n    tensor_parallel_size: 8\n```\n\n### Launch Multi-Node\n\n```bash\n# On head node\nray start --head --port=6379\n\n# On worker nodes\nray start --address='head_ip:6379'\n\n# Launch training\npython3 -m verl.trainer.main_ppo \\\n    trainer.nnodes=4 \\\n    trainer.n_gpus_per_node=8\n```\n\n---\n\n## Configuration Reference\n\n### Algorithm Selection\n\n| Algorithm | `adv_estimator` | Use Case |\n|-----------|-----------------|----------|\n| GRPO | `grpo` | Critic-free, math\u002Freasoning |\n| PPO\u002FGAE | `gae` | Dense rewards, value estimation |\n| REINFORCE++ | `reinforce_plus_plus` | Variance reduction |\n| RLOO | `rloo` | Leave-one-out baseline |\n| ReMax | `remax` | Maximum reward baseline |\n| OPO | `opo` | Optimal policy optimization |\n\n### Key Parameters\n\n```yaml\n# Rollout parameters\nactor_rollout_ref.rollout.n: 8              # Samples per prompt\nactor_rollout_ref.rollout.temperature: 0.7  # Sampling temperature\nactor_rollout_ref.rollout.top_p: 0.95       # Nucleus sampling\n\n# Training parameters\nactor_rollout_ref.actor.lr: 1e-6            # Learning rate\nactor_rollout_ref.actor.ppo_mini_batch_size: 64\nactor_rollout_ref.actor.clip_ratio: 0.2     # PPO clip range\n\n# KL control\nactor_rollout_ref.actor.use_kl_loss: true\nactor_rollout_ref.actor.kl_loss_coef: 0.001\nalgorithm.kl_ctrl.target_kl: 0.1            # For adaptive KL control\n```\n\n---\n\n## Common Issues and Solutions\n\n### Issue: OOM During Rollout\n\n**Symptoms**: CUDA out of memory during generation phase\n\n**Solutions**:\n```yaml\n# Reduce batch size\nactor_rollout_ref.rollout.log_prob_micro_batch_size: 4\n\n# Enable gradient checkpointing\nactor_rollout_ref.model.enable_gradient_checkpointing: true\n\n# Use FSDP2 with CPU offloading\nactor_rollout_ref.actor.strategy: fsdp2\nactor_rollout_ref.actor.fsdp_config.offload_policy: true\n```\n\n### Issue: Training Instability\n\n**Symptoms**: Loss spikes, reward collapse\n\n**Solutions**:\n```yaml\n# Reduce learning rate\nactor_rollout_ref.actor.lr: 5e-7\n\n# Increase KL penalty\nactor_rollout_ref.actor.kl_loss_coef: 0.01\n\n# Enable gradient clipping\nactor_rollout_ref.actor.max_grad_norm: 1.0\n```\n\n### Issue: Slow Weight Sync\n\n**Symptoms**: Long pauses between rollout and training\n\n**Solutions**:\n```bash\n# Use FSDP2 for faster resharding\nactor_rollout_ref.actor.strategy=fsdp2\n\n# Enable async weight transfer\ntrainer.async_weight_update=true\n```\n\n### Issue: vLLM Version Mismatch\n\n**Symptoms**: Import errors or generation failures\n\n**Solution**: Use compatible versions:\n```bash\npip install vllm>=0.8.5,\u003C=0.12.0\n# Avoid vLLM 0.7.x (known bugs)\n```\n\n---\n\n## Advanced Topics\n\n### Multi-Turn Tool Calling\n\nSee [references\u002Fmulti-turn.md](references\u002Fmulti-turn.md) for agentic workflows with tool use.\n\n### Vision-Language Models\n\n```yaml\nactor_rollout_ref:\n  model:\n    path: Qwen\u002FQwen2.5-VL-7B-Instruct\n  rollout:\n    name: vllm\n    enable_vision: true\n```\n\n### LoRA Training\n\n```yaml\nactor_rollout_ref:\n  actor:\n    lora:\n      enabled: true\n      r: 16\n      alpha: 32\n      target_modules: [\"q_proj\", \"v_proj\"]\n```\n\n---\n\n## Resources\n\n- **Documentation**: https:\u002F\u002Fverl.readthedocs.io\u002F\n- **Paper**: https:\u002F\u002Farxiv.org\u002Fabs\u002F2409.19256\n- **GitHub**: https:\u002F\u002Fgithub.com\u002Fvolcengine\u002Fverl\n- **Recipes**: https:\u002F\u002Fgithub.com\u002Fverl-project\u002Fverl-recipe (DAPO, GSPO, etc.)\n- **Community**: Slack at verl-project\n\n","# verl: Volcano Engine Reinforcement Learning for LLMs\n\nverl 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.\n\n## When to Use verl\n\n**Choose verl when you need:**\n- Production-ready RL training at scale (tested up to 671B parameters)\n- Flexibility to swap backends (FSDP ↔ Megatron-LM ↔ vLLM ↔ SGLang)\n- Support for multiple RL algorithms (PPO, GRPO, RLOO, REINFORCE++, DAPO)\n- Multi-turn rollout with tool calling for agentic workflows\n- Vision-language model RL training\n\n**Consider alternatives when:**\n- You need Megatron-native training → use **slime** or **miles**\n- You want PyTorch-native abstractions with Monarch → use **torchforge**\n- You only need simple SFT\u002FDPO → use **TRL** or **Axolotl**\n\n## Key Features\n\n- **Training backends**: FSDP, FSDP2, Megatron-LM\n- **Rollout engines**: vLLM, SGLang, HuggingFace Transformers\n- **Algorithms**: PPO, GRPO, DAPO, RLOO, ReMax, REINFORCE++, SPIN, SPPO\n- **Models**: Qwen-3, Llama-3.1, DeepSeek, Gemma-2 (0.5B to 671B)\n- **Advanced**: LoRA RL, sequence parallelism, expert parallelism, multi-turn tools\n\n## Installation\n\n```bash\n# Option 1: pip install\npip install verl[vllm]  # or verl[sglang] for SGLang backend\n\n# Option 2: Docker (recommended for production)\ndocker pull verlai\u002Fverl:vllm011.latest\n\n# Option 3: From source\ngit clone https:\u002F\u002Fgithub.com\u002Fvolcengine\u002Fverl.git\ncd verl && pip install -e .[vllm,math]\n```\n\n## Quick Start: GRPO Training\n\n```bash\npython3 -m verl.trainer.main_ppo \\\n    algorithm.adv_estimator=grpo \\\n    data.train_files=~\u002Fdata\u002Fgsm8k\u002Ftrain.parquet \\\n    actor_rollout_ref.model.path=Qwen\u002FQwen2.5-7B \\\n    actor_rollout_ref.rollout.n=8 \\\n    actor_rollout_ref.actor.use_kl_loss=True \\\n    trainer.n_gpus_per_node=8\n```\n\n## Core Architecture\n\nverl uses a **HybridFlow** programming model separating control flow from computation:\n\n```\n┌─────────────────────────────────────────────────────────┐\n│ Single-Process Controller (Ray)                         │\n│ - Orchestrates: rollout → reward → train → sync        │\n└─────────────────────┬───────────────────────────────────┘\n                      │\n┌─────────────────────▼───────────────────────────────────┐\n│ Multi-Process Workers                                   │\n│ ├── ActorRolloutRefWorker (policy + generation)        │\n│ ├── CriticWorker (value estimation, PPO only)          │\n│ └── RewardManager (model-based or rule-based rewards)  │\n└─────────────────────────────────────────────────────────┘\n```\n\n---\n\n## Workflow 1: Math Reasoning with GRPO\n\nUse this workflow for training reasoning models on math tasks like GSM8K or MATH.\n\n### Prerequisites Checklist\n- [ ] GPU cluster with 8+ GPUs (H100 recommended)\n- [ ] Dataset in parquet format with `prompt` and `reward_model` columns\n- [ ] Base model from HuggingFace Hub\n\n### Step 1: Prepare Dataset\n\n```python\nimport pandas as pd\n\ndata = [\n    {\n        \"prompt\": [{\"role\": \"user\", \"content\": \"What is 15 + 27?\"}],\n        \"reward_model\": {\"ground_truth\": \"42\"}\n    },\n    # ... more examples\n]\ndf = pd.DataFrame(data)\ndf.to_parquet(\"train.parquet\")\n```\n\n### Step 2: Define Reward Function\n\n```python\n# reward_function.py\nimport re\n\ndef compute_reward(responses, ground_truths):\n    rewards = []\n    for response, gt in zip(responses, ground_truths):\n        # Extract answer from response\n        match = re.search(r'\\\\boxed{([^}]+)}', response)\n        if match and match.group(1).strip() == gt.strip():\n            rewards.append(1.0)\n        else:\n            rewards.append(0.0)\n    return rewards\n```\n\n### Step 3: Create Training Config\n\n```yaml\n# config\u002Fgrpo_math.yaml\nalgorithm:\n  adv_estimator: grpo\n  gamma: 1.0\n  lam: 1.0\n\ndata:\n  train_files: \u002Fpath\u002Fto\u002Ftrain.parquet\n  val_files: \u002Fpath\u002Fto\u002Fval.parquet\n  train_batch_size: 256\n  max_prompt_length: 512\n  max_response_length: 2048\n\nactor_rollout_ref:\n  model:\n    path: Qwen\u002FQwen2.5-7B-Instruct\n  actor:\n    use_kl_loss: true\n    kl_loss_coef: 0.001\n    ppo_mini_batch_size: 64\n  rollout:\n    name: vllm\n    n: 8  # samples per prompt\n    temperature: 0.7\n    top_p: 0.95\n\ntrainer:\n  total_epochs: 3\n  n_gpus_per_node: 8\n  save_freq: 100\n```\n\n### Step 4: Launch Training\n\n```bash\npython3 -m verl.trainer.main_ppo \\\n    --config-path config \\\n    --config-name grpo_math \\\n    trainer.experiment_name=grpo_math_qwen7b\n```\n\n### Step 5: Monitor and Validate\n- [ ] Check WandB\u002FTensorBoard for loss curves\n- [ ] Verify reward is increasing over steps\n- [ ] Run evaluation on held-out test set\n\n---\n\n## Workflow 2: PPO with Critic Model\n\nUse this workflow when you need value-based advantage estimation (GAE).\n\n### Key Differences from GRPO\n- Requires separate critic model\n- Uses Generalized Advantage Estimation (GAE)\n- Better for tasks with dense rewards\n\n### Configuration\n\n```yaml\nalgorithm:\n  adv_estimator: gae  # Use GAE instead of GRPO\n  gamma: 0.99\n  lam: 0.95\n\ncritic:\n  model:\n    path: Qwen\u002FQwen2.5-7B-Instruct  # Can be same or different from actor\n  ppo_mini_batch_size: 64\n\nactor_rollout_ref:\n  actor:\n    use_kl_loss: true\n    kl_loss_coef: 0.02\n    clip_ratio: 0.2  # PPO clipping\n```\n\n### Launch with Critic\n\n```bash\npython3 -m verl.trainer.main_ppo \\\n    algorithm.adv_estimator=gae \\\n    critic.model.path=Qwen\u002FQwen2.5-7B-Instruct \\\n    trainer.n_gpus_per_node=8\n```\n\n---\n\n## Workflow 3: Large-Scale Training with Megatron\n\nUse this workflow for models >70B parameters or when you need expert parallelism.\n\n### Prerequisites\n- [ ] Install Megatron-LM bridge: `pip install mbridge`\n- [ ] Convert model to Megatron format\n- [ ] Multi-node cluster with NVLink\u002FInfiniBand\n\n### Configuration for 70B+ Models\n\n```yaml\nactor_rollout_ref:\n  model:\n    path: \u002Fpath\u002Fto\u002Fmegatron\u002Fcheckpoint\n    backend: megatron\n  actor:\n    strategy: megatron\n    tensor_model_parallel_size: 8\n    pipeline_model_parallel_size: 2\n  rollout:\n    name: vllm\n    tensor_parallel_size: 8\n```\n\n### Launch Multi-Node\n\n```bash\n# On head node\nray start --head --port=6379\n\n# On worker nodes\nray start --address='head_ip:6379'\n\n# Launch training\npython3 -m verl.trainer.main_ppo \\\n    trainer.nnodes=4 \\\n    trainer.n_gpus_per_node=8\n```\n\n---\n\n## Configuration Reference\n\n### Algorithm Selection\n\n| Algorithm | `adv_estimator` | Use Case |\n|-----------|-----------------|----------|\n| GRPO | `grpo` | Critic-free, math\u002Freasoning |\n| PPO\u002FGAE | `gae` | Dense rewards, value estimation |\n| REINFORCE++ | `reinforce_plus_plus` | Variance reduction |\n| RLOO | `rloo` | Leave-one-out baseline |\n| ReMax | `remax` | Maximum reward baseline |\n| OPO | `opo` | Optimal policy optimization |\n\n### Key Parameters\n\n```yaml\n# Rollout parameters\nactor_rollout_ref.rollout.n: 8              # Samples per prompt\nactor_rollout_ref.rollout.temperature: 0.7  # Sampling temperature\nactor_rollout_ref.rollout.top_p: 0.95       # Nucleus sampling\n\n# Training parameters\nactor_rollout_ref.actor.lr: 1e-6            # Learning rate\nactor_rollout_ref.actor.ppo_mini_batch_size: 64\nactor_rollout_ref.actor.clip_ratio: 0.2     # PPO clip range\n\n# KL control\nactor_rollout_ref.actor.use_kl_loss: true\nactor_rollout_ref.actor.kl_loss_coef: 0.001\nalgorithm.kl_ctrl.target_kl: 0.1            # For adaptive KL control\n```\n\n---\n\n## Common Issues and Solutions\n\n### Issue: OOM During Rollout\n\n**Symptoms**: CUDA out of memory during generation phase\n\n**Solutions**:\n```yaml\n# Reduce batch size\nactor_rollout_ref.rollout.log_prob_micro_batch_size: 4\n\n# Enable gradient checkpointing\nactor_rollout_ref.model.enable_gradient_checkpointing: true\n\n# Use FSDP2 with CPU offloading\nactor_rollout_ref.actor.strategy: fsdp2\nactor_rollout_ref.actor.fsdp_config.offload_policy: true\n```\n\n### Issue: Training Instability\n\n**Symptoms**: Loss spikes, reward collapse\n\n**Solutions**:\n```yaml\n# Reduce learning rate\nactor_rollout_ref.actor.lr: 5e-7\n\n# Increase KL penalty\nactor_rollout_ref.actor.kl_loss_coef: 0.01\n\n# Enable gradient clipping\nactor_rollout_ref.actor.max_grad_norm: 1.0\n```\n\n### Issue: Slow Weight Sync\n\n**Symptoms**: Long pauses between rollout and training\n\n**Solutions**:\n```bash\n# Use FSDP2 for faster resharding\nactor_rollout_ref.actor.strategy=fsdp2\n\n# Enable async weight transfer\ntrainer.async_weight_update=true\n```\n\n### Issue: vLLM Version Mismatch\n\n**Symptoms**: Import errors or generation failures\n\n**Solution**: Use compatible versions:\n```bash\npip install vllm>=0.8.5,\u003C=0.12.0\n# Avoid vLLM 0.7.x (known bugs)\n```\n\n---\n\n## Advanced Topics\n\n### Multi-Turn Tool Calling\n\nSee [references\u002Fmulti-turn.md](references\u002Fmulti-turn.md) for agentic workflows with tool use.\n\n### Vision-Language Models\n\n```yaml\nactor_rollout_ref:\n  model:\n    path: Qwen\u002FQwen2.5-VL-7B-Instruct\n  rollout:\n    name: vllm\n    enable_vision: true\n```\n\n### LoRA Training\n\n```yaml\nactor_rollout_ref:\n  actor:\n    lora:\n      enabled: true\n      r: 16\n      alpha: 32\n      target_modules: [\"q_proj\", \"v_proj\"]\n```\n\n---\n\n## Resources\n\n- **Documentation**: https:\u002F\u002Fverl.readthedocs.io\u002F\n- **Paper**: https:\u002F\u002Farxiv.org\u002Fabs\u002F2409.19256\n- **GitHub**: https:\u002F\u002Fgithub.com\u002Fvolcengine\u002Fverl\n- **Recipes**: https:\u002F\u002Fgithub.com\u002Fverl-project\u002Fverl-recipe (DAPO, GSPO, etc.)\n- **Community**: Slack at verl-project\n\n",{"name":5,"tags":44,"author":47,"license":14,"version":48,"description":7,"dependencies":49},[45,35,46,34,36,37],"Reinforcement Learning","GRPO","Orchestra Research","1.0.0",[50,51,52,53,54],"verl>=0.3.0","torch>=2.0.0","ray>=2.41.0","vllm>=0.8.2","transformers>=4.40.0","main","f19b254e8975d0bfddb394686706106930d7964d","AI-Training\u002Fverl","https:\u002F\u002Fgithub.com\u002Fendman100\u002Fskill-router-nexus","https:\u002F\u002Fgithub.com\u002Fendman100","2026-11-12 18:13:37",{"source":8,"skillName":5,"commands":62,"prompt":72,"zipAvailable":73},[63,66,69],{"tool":64,"command":65},"npx","npx skills add endman100\u002Fskill-router-nexus --skill verl-rl-training",{"tool":67,"command":68},"bunx","bunx skills add endman100\u002Fskill-router-nexus --skill verl-rl-training",{"tool":70,"command":71},"pnpm","pnpm dlx skills add endman100\u002Fskill-router-nexus --skill verl-rl-training","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fendman100\u002Fskill-router-nexus\u002Fblob\u002Fmain\u002FAI-Training\u002Fverl\u002FSKILL.md",true,[75,81],{"path":76,"resourceType":77,"resourceTypeDesc":78,"sizeBytes":79,"sourceHtmlUrl":80},"references\u002Fapi-reference.md","REFERENCE","参考文档",6941,"https:\u002F\u002Fgithub.com\u002Fendman100\u002Fskill-router-nexus\u002Fblob\u002Fmain\u002FAI-Training\u002Fverl\u002Freferences\u002Fapi-reference.md",{"path":82,"resourceType":77,"resourceTypeDesc":78,"sizeBytes":83,"sourceHtmlUrl":84},"references\u002Ftroubleshooting.md",6792,"https:\u002F\u002Fgithub.com\u002Fendman100\u002Fskill-router-nexus\u002Fblob\u002Fmain\u002FAI-Training\u002Fverl\u002Freferences\u002Ftroubleshooting.md",[86,110,145,167,176,200],{"skillNo":87,"name":88,"nameZh":6,"description":89,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":90,"sourceHtmlUrl":91,"categories":92,"keywords":98,"publishedAt":39},"01M6MPCT5BXVER8HQZTW3QE9AS","create-master","基于佛教经典文献，生成特定高僧大德的 AI 教学角色","https:\u002F\u002Fgithub.com\u002Fendman100\u002Fskill-router-nexus\u002Fblob\u002Fmain\u002FMetaphysics-and-Spirituality\u002FMaster-skill\u002FLICENSE","https:\u002F\u002Fgithub.com\u002Fendman100\u002Fskill-router-nexus\u002Fblob\u002Fmain\u002FMetaphysics-and-Spirituality\u002FMaster-skill\u002FSKILL.md",[93],{"categoryNo":94,"categoryName":95,"taxonomyType":96,"taxonomyTypeDesc":97},"01JBSEEDCAT000000000000039","教师与教育","PROFESSION","适用职业",[99,100,101,102,103,104,105,106,107,108,109],"create master","master","create","生成","true","webfetch","edit","write","read","grep","Bash",{"skillNo":111,"name":112,"nameZh":6,"description":113,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":114,"sourceHtmlUrl":115,"categories":116,"keywords":128,"publishedAt":39},"01M6M43AKE95Y64E1JRXEF5QQ3","writing-dna-skill","从至少 20 篇完整文章中蒸馏可复用的写作 DNA，分析语言、文章结构、选题逻辑、素材策略、认知框架和视觉风格，并生成 Writing-DNA.md；按该 DNA 写作时会读取全部蒸馏产物和 5 篇相关原文。用于中英文作者、账号、品牌或出版物的风格分析与一致性写作。 Distill reusable Writing DNA from at least 20 complete articles for Chinese or English authors, publications, brands, and accounts; use for language, structure, topic logic, source strategy, cognitive-frame, and visual-style analysis.","https:\u002F\u002Fgithub.com\u002Fendman100\u002Fskill-router-nexus\u002Fblob\u002Fmain\u002FWriting-and-Content\u002Fwriting-dna-skill\u002FLICENSE","https:\u002F\u002Fgithub.com\u002Fendman100\u002Fskill-router-nexus\u002Fblob\u002Fmain\u002FWriting-and-Content\u002Fwriting-dna-skill\u002FSKILL.md",[117,122,125],{"categoryNo":118,"categoryName":119,"taxonomyType":120,"taxonomyTypeDesc":121},"01JBSEEDCAT000000000000018","写作与文案","SCENARIO","使用场景",{"categoryNo":123,"categoryName":124,"taxonomyType":120,"taxonomyTypeDesc":121},"01JBSEEDCAT000000000000023","Skill与智能体开发",{"categoryNo":126,"categoryName":127,"taxonomyType":96,"taxonomyTypeDesc":97},"01JBSEEDCAT000000000000036","内容创作者",[129,130,131,132,133,134,135,136,137,138,22,139,140,102,141,142,143,24,27,144],"writing dna skill","DNA","writing","strategy","english","logic","visual","structure","chinese","reusable","style","distill","analysis","frame","source","complete",{"skillNo":146,"name":147,"nameZh":6,"description":148,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":149,"sourceHtmlUrl":150,"categories":151,"keywords":154,"publishedAt":39},"01M6M428TTCGQ4ME3KY99VN037","stop-slop","Remove AI writing patterns from prose. Use when drafting, editing, or reviewing text to eliminate predictable AI tells.","https:\u002F\u002Fgithub.com\u002Fendman100\u002Fskill-router-nexus\u002Fblob\u002Fmain\u002FWriting-and-Content\u002Fstop-slop\u002FLICENSE","https:\u002F\u002Fgithub.com\u002Fendman100\u002Fskill-router-nexus\u002Fblob\u002Fmain\u002FWriting-and-Content\u002Fstop-slop\u002FSKILL.md",[152,153],{"categoryNo":118,"categoryName":119,"taxonomyType":120,"taxonomyTypeDesc":121},{"categoryNo":126,"categoryName":127,"taxonomyType":96,"taxonomyTypeDesc":97},[155,156,157,158,23,159,160,161,162,22,163,164,165,24,166,131],"slop","stop slop","stop","drafting","drafts","https","patterns","prose","remove","editing","text","content",{"skillNo":168,"name":169,"nameZh":6,"description":170,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":171,"sourceHtmlUrl":172,"categories":173,"keywords":174,"publishedAt":39},"01M6M41X0B7EBX42TK435DMGGR","shuorenhua","检查和清理中英文文本里的 AI 套路，适用于“去 AI 味”“说人话”“自然一点”“别像模板”“先标问题”这类改写和审稿需求；按场景控制力度，同时保留事实、术语、语域和责任主体。","https:\u002F\u002Fgithub.com\u002Fendman100\u002Fskill-router-nexus\u002Fblob\u002Fmain\u002FWriting-and-Content\u002Fshuorenhua\u002FLICENSE","https:\u002F\u002Fgithub.com\u002Fendman100\u002Fskill-router-nexus\u002Fblob\u002Fmain\u002FWriting-and-Content\u002Fshuorenhua\u002FSKILL.md",[],[169,175],"文本",{"skillNo":177,"name":178,"nameZh":6,"description":179,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":180,"sourceHtmlUrl":181,"categories":182,"keywords":183,"publishedAt":39},"01M6M3ZD12WBEW5KXKDM4N6DFG","naming","Name products, SaaS, brands, open source projects, bots, and apps. Use when the user needs to name something, find a brand name, or pick a product name. Metaphor-driven process that produces memorable, meaningful names and avoids AI slop.","https:\u002F\u002Fgithub.com\u002Fendman100\u002Fskill-router-nexus\u002Fblob\u002Fmain\u002FWriting-and-Content\u002Fnaming-main\u002FLICENSE","https:\u002F\u002Fgithub.com\u002Fendman100\u002Fskill-router-nexus\u002Fblob\u002Fmain\u002FWriting-and-Content\u002Fnaming-main\u002FSKILL.md",[],[178,184,185,186,187,188,22,189,107,143,190,191,192,193,109,194,195,196,197,23,198,199,155],"repo","websearch","product name","apps","bots","view","brand","pick","gh","projects","open","saas","open source","product","describe","ai slop",{"skillNo":201,"name":202,"nameZh":6,"description":203,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":204,"sourceHtmlUrl":205,"categories":206,"keywords":208,"publishedAt":39},"01M6M3YVXAFAAWJ7F6JMPFREYE","humanizer-zh","去除文本中的 AI 生成痕迹。适用于编辑或审阅文本，使其听起来更自然、更像人类书写。\n基于维基百科的\"AI 写作特征\"综合指南。检测并修复以下模式：夸大的象征意义、\n宣传性语言、以 -ing 结尾的肤浅分析、模糊的归因、破折号过度使用、三段式法则、\nAI 词汇、否定式排比、过多的连接性短语。","https:\u002F\u002Fgithub.com\u002Fendman100\u002Fskill-router-nexus\u002Fblob\u002Fmain\u002FWriting-and-Content\u002Fhumanizer-zh\u002FLICENSE","https:\u002F\u002Fgithub.com\u002Fendman100\u002Fskill-router-nexus\u002Fblob\u002Fmain\u002FWriting-and-Content\u002Fhumanizer-zh\u002FSKILL.md",[207],{"categoryNo":118,"categoryName":119,"taxonomyType":120,"taxonomyTypeDesc":121},[209,210,211,107,106,105,131,102,175,212,155,157],"humanizer","zh","humanizer zh","翻译"]