[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7D34N03368TMS48S12ANCD1":3,"skill-same:01M7D34N03368TMS48S12ANCD1":83},{"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":18,"linkOnly":19,"markdownContent":20,"markdownBody":21,"frontMatter":22,"refName":45,"commitSha":46,"skillDirectory":47,"forksCount":13,"repositoryHtmlUrl":48,"repositoryHomepageUrl":6,"ownerHtmlUrl":49,"ownerBlogUrl":6,"ownerDescription":6,"updatedAt":50,"install":51,"resources":64},"01M7D34N03368TMS48S12ANCD1","openrlhf-training",null,"High-performance RLHF framework with Ray+vLLM acceleration. Use for PPO, GRPO, RLOO, DPO training of large models (7B-70B+). Built on Ray, vLLM, ZeRO-3. 2× faster than DeepSpeedChat with distributed architecture and GPU resource sharing.","beyondai\u002Fsandbox","beyondai","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F2845958?v=4","01M7D24SBNM47BJGJ9J2HB13BP","01M7D24SCDC9DYE005XH6R553S",0,"MIT","https:\u002F\u002Fgithub.com\u002Fbeyondai\u002Fsandbox\u002Fblob\u002Fmain\u002Flabs\u002F.cursor\u002Fskills\u002Fopenrlhf\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: openrlhf-training\ndescription: High-performance RLHF framework with Ray+vLLM acceleration. Use for PPO, GRPO, RLOO, DPO training of large models (7B-70B+). Built on Ray, vLLM, ZeRO-3. 2× faster than DeepSpeedChat with distributed architecture and GPU resource sharing.\nversion: 1.0.0\nauthor: Orchestra Research\nlicense: MIT\ntags: [Post-Training, OpenRLHF, RLHF, PPO, GRPO, RLOO, DPO, Ray, vLLM, Distributed Training, Large Models, ZeRO-3]\ndependencies: [openrlhf, ray, vllm, torch, transformers, deepspeed]\n---\n\n# OpenRLHF - High-Performance RLHF Training\n\n## Quick start\n\nOpenRLHF is a Ray-based RLHF framework optimized for distributed training with vLLM inference acceleration.\n\n**Installation**:\n```bash\n# Launch Docker container\ndocker run --runtime=nvidia -it --rm --shm-size=\"10g\" --cap-add=SYS_ADMIN \\\n  -v $PWD:\u002Fopenrlhf nvcr.io\u002Fnvidia\u002Fpytorch:25.02-py3 bash\n\n# Uninstall conflicts\nsudo pip uninstall xgboost transformer_engine flash_attn pynvml -y\n\n# Install OpenRLHF with vLLM\npip install openrlhf[vllm]\n```\n\n**PPO Training** (Hybrid Engine):\n```bash\nray start --head --node-ip-address 0.0.0.0 --num-gpus 8\n\nray job submit --address=\"http:\u002F\u002F127.0.0.1:8265\" \\\n  --runtime-env-json='{\"working_dir\": \"\u002Fopenrlhf\"}' \\\n  -- python3 -m openrlhf.cli.train_ppo_ray \\\n  --ref_num_nodes 1 --ref_num_gpus_per_node 8 \\\n  --reward_num_nodes 1 --reward_num_gpus_per_node 8 \\\n  --critic_num_nodes 1 --critic_num_gpus_per_node 8 \\\n  --actor_num_nodes 1 --actor_num_gpus_per_node 8 \\\n  --vllm_num_engines 4 --vllm_tensor_parallel_size 2 \\\n  --colocate_all_models \\\n  --vllm_gpu_memory_utilization 0.5 \\\n  --pretrain OpenRLHF\u002FLlama-3-8b-sft-mixture \\\n  --reward_pretrain OpenRLHF\u002FLlama-3-8b-rm-700k \\\n  --save_path .\u002Foutput\u002Fllama3-8b-rlhf \\\n  --micro_train_batch_size 8 --train_batch_size 128 \\\n  --micro_rollout_batch_size 16 --rollout_batch_size 1024 \\\n  --max_epochs 1 --prompt_max_len 1024 --generate_max_len 1024 \\\n  --zero_stage 3 --bf16 \\\n  --actor_learning_rate 5e-7 --critic_learning_rate 9e-6 \\\n  --init_kl_coef 0.01 --normalize_reward \\\n  --gradient_checkpointing --packing_samples \\\n  --vllm_enable_sleep --deepspeed_enable_sleep\n```\n\n**GRPO Training** (Group Normalized Policy Optimization):\n```bash\n# Same command as PPO, but add:\n--advantage_estimator group_norm\n```\n\n## Common workflows\n\n### Workflow 1: Full RLHF pipeline (SFT → Reward Model → PPO)\n\n**Step 1: Train reward model** (DPO):\n```bash\ndeepspeed --module openrlhf.cli.train_rm \\\n  --save_path .\u002Foutput\u002Fllama3-8b-rm \\\n  --save_steps -1 --logging_steps 1 \\\n  --eval_steps -1 --train_batch_size 256 \\\n  --micro_train_batch_size 1 --pretrain meta-llama\u002FMeta-Llama-3-8B \\\n  --bf16 --max_epochs 1 --max_len 8192 \\\n  --zero_stage 3 --learning_rate 9e-6 \\\n  --dataset OpenRLHF\u002Fpreference_dataset_mixture2_and_safe_pku \\\n  --apply_chat_template --chosen_key chosen \\\n  --rejected_key rejected --flash_attn --gradient_checkpointing\n```\n\n**Step 2: PPO training**:\n```bash\nray start --head --node-ip-address 0.0.0.0 --num-gpus 8\n\nray job submit --address=\"http:\u002F\u002F127.0.0.1:8265\" \\\n  -- python3 -m openrlhf.cli.train_ppo_ray \\\n  --ref_num_nodes 1 --ref_num_gpus_per_node 8 \\\n  --reward_num_nodes 1 --reward_num_gpus_per_node 8 \\\n  --critic_num_nodes 1 --critic_num_gpus_per_node 8 \\\n  --actor_num_nodes 1 --actor_num_gpus_per_node 8 \\\n  --vllm_num_engines 4 --vllm_tensor_parallel_size 2 \\\n  --colocate_all_models \\\n  --pretrain OpenRLHF\u002FLlama-3-8b-sft-mixture \\\n  --reward_pretrain .\u002Foutput\u002Fllama3-8b-rm \\\n  --save_path .\u002Foutput\u002Fllama3-8b-ppo \\\n  --micro_train_batch_size 8 --train_batch_size 128 \\\n  --micro_rollout_batch_size 16 --rollout_batch_size 1024 \\\n  --max_epochs 1 --prompt_max_len 1024 --generate_max_len 1024 \\\n  --zero_stage 3 --bf16 \\\n  --actor_learning_rate 5e-7 --critic_learning_rate 9e-6 \\\n  --init_kl_coef 0.01 --normalize_reward \\\n  --vllm_enable_sleep --deepspeed_enable_sleep\n```\n\n### Workflow 2: GRPO training (no critic model needed)\n\nMemory-efficient alternative to PPO:\n\n```bash\nray job submit --address=\"http:\u002F\u002F127.0.0.1:8265\" \\\n  -- python3 -m openrlhf.cli.train_ppo_ray \\\n  --advantage_estimator group_norm \\\n  --ref_num_nodes 1 --ref_num_gpus_per_node 8 \\\n  --reward_num_nodes 1 --reward_num_gpus_per_node 8 \\\n  --actor_num_nodes 1 --actor_num_gpus_per_node 8 \\\n  --vllm_num_engines 4 --vllm_tensor_parallel_size 2 \\\n  --colocate_all_models \\\n  --pretrain OpenRLHF\u002FLlama-3-8b-sft-mixture \\\n  --reward_pretrain OpenRLHF\u002FLlama-3-8b-rm-700k \\\n  --save_path .\u002Foutput\u002Fllama3-8b-grpo \\\n  --micro_train_batch_size 8 --train_batch_size 128 \\\n  --micro_rollout_batch_size 16 --rollout_batch_size 1024 \\\n  --max_epochs 1 --bf16 \\\n  --actor_learning_rate 5e-7 \\\n  --init_kl_coef 0.01 --use_kl_loss --kl_estimator k3 \\\n  --normalize_reward --no_advantage_std_norm\n```\n\n**Key GRPO parameters**:\n- `--advantage_estimator group_norm` - Enables GRPO\n- `--use_kl_loss` - KL loss from GRPO paper\n- `--kl_estimator k3` - Loss function (k2 ≈ k1)\n- `--no_advantage_std_norm` - Disables std normalization\n\n### Workflow 3: DPO training (preference optimization)\n\nSimpler alternative without reward model:\n\n```bash\ndeepspeed --module openrlhf.cli.train_dpo \\\n  --save_path .\u002Foutput\u002Fllama3-8b-dpo \\\n  --save_steps -1 --logging_steps 1 \\\n  --eval_steps -1 --train_batch_size 256 \\\n  --micro_train_batch_size 2 --pretrain meta-llama\u002FMeta-Llama-3-8B \\\n  --bf16 --max_epochs 1 --max_len 8192 \\\n  --zero_stage 3 --learning_rate 5e-7 --beta 0.1 \\\n  --dataset OpenRLHF\u002Fpreference_dataset_mixture2_and_safe_pku \\\n  --apply_chat_template --chosen_key chosen \\\n  --rejected_key rejected --flash_attn --gradient_checkpointing\n```\n\n## When to use vs alternatives\n\n**Use OpenRLHF when**:\n- Training large models (7B-70B+) with RL\n- Need vLLM inference acceleration\n- Want distributed architecture with Ray\n- Have multi-node GPU cluster\n- Need PPO\u002FGRPO\u002FRLOO\u002FDPO in one framework\n\n**Algorithm selection**:\n- **PPO**: Maximum control, best for complex rewards\n- **GRPO**: Memory-efficient, no critic needed\n- **RLOO**: Modified PPO with per-token KL\n- **REINFORCE++**: More stable than GRPO, faster than PPO\n- **DPO**: Simplest, no reward model needed\n\n**Use alternatives instead**:\n- **TRL**: Single-node training, simpler API\n- **veRL**: ByteDance's framework for 671B models\n- **DeepSpeedChat**: Integrated with DeepSpeed ecosystem\n\n## Common issues\n\n**Issue: GPU OOM with large models**\n\nDisable model colocation:\n```bash\n# Remove --colocate_all_models flag\n# Allocate separate GPUs for each model\n--actor_num_gpus_per_node 8 \\\n--critic_num_gpus_per_node 8 \\\n--reward_num_gpus_per_node 8 \\\n--ref_num_gpus_per_node 8\n```\n\n**Issue: DeepSpeed GPU index out of range**\n\nSet environment variable:\n```bash\nexport RAY_EXPERIMENTAL_NOSET_CUDA_VISIBLE_DEVICES=1\n```\n\n**Issue: Training instability**\n\nUse Hybrid Engine instead of async:\n```bash\n--colocate_all_models \\\n--vllm_enable_sleep \\\n--deepspeed_enable_sleep\n```\n\nAdjust KL coefficient:\n```bash\n--init_kl_coef 0.05  # Increase from 0.01\n```\n\n**Issue: Slow generation during PPO**\n\nEnable vLLM acceleration:\n```bash\n--vllm_num_engines 4 \\\n--vllm_tensor_parallel_size 2 \\\n--vllm_gpu_memory_utilization 0.5\n```\n\n## Advanced topics\n\n**Hybrid Engine GPU sharing**: See [references\u002Fhybrid-engine.md](references\u002Fhybrid-engine.md) for vLLM sleep mode, DeepSpeed sleep mode, and optimal node allocation.\n\n**Algorithm comparison**: See [references\u002Falgorithm-comparison.md](references\u002Falgorithm-comparison.md) for PPO vs GRPO vs RLOO vs REINFORCE++ benchmarks and hyperparameters.\n\n**Multi-node setup**: See [references\u002Fmulti-node-training.md](references\u002Fmulti-node-training.md) for Ray cluster configuration and fault tolerance.\n\n**Custom reward functions**: See [references\u002Fcustom-rewards.md](references\u002Fcustom-rewards.md) for reinforced fine-tuning and agent RLHF.\n\n## Hardware requirements\n\n- **GPU**: NVIDIA A100\u002FH100 recommended\n- **VRAM**:\n  - 7B model: 8× A100 40GB (Hybrid Engine)\n  - 70B model: 48× A100 80GB (vLLM:Actor:Critic = 1:1:1)\n- **Multi-node**: Ray cluster with InfiniBand recommended\n- **Docker**: NVIDIA PyTorch container 25.02+\n\n**Performance**:\n- 2× faster than DeepSpeedChat\n- vLLM inference acceleration\n- Hybrid Engine minimizes GPU idle time\n\n## Resources\n\n- Docs: https:\u002F\u002Fgithub.com\u002FOpenRLHF\u002FOpenRLHF\n- Paper: https:\u002F\u002Farxiv.org\u002Fabs\u002F2405.11143\n- Examples: https:\u002F\u002Fgithub.com\u002FOpenRLHF\u002FOpenRLHF\u002Ftree\u002Fmain\u002Fexamples\n- Discord: Community support\n\n\n\n","# OpenRLHF - High-Performance RLHF Training\n\n## Quick start\n\nOpenRLHF is a Ray-based RLHF framework optimized for distributed training with vLLM inference acceleration.\n\n**Installation**:\n```bash\n# Launch Docker container\ndocker run --runtime=nvidia -it --rm --shm-size=\"10g\" --cap-add=SYS_ADMIN \\\n  -v $PWD:\u002Fopenrlhf nvcr.io\u002Fnvidia\u002Fpytorch:25.02-py3 bash\n\n# Uninstall conflicts\nsudo pip uninstall xgboost transformer_engine flash_attn pynvml -y\n\n# Install OpenRLHF with vLLM\npip install openrlhf[vllm]\n```\n\n**PPO Training** (Hybrid Engine):\n```bash\nray start --head --node-ip-address 0.0.0.0 --num-gpus 8\n\nray job submit --address=\"http:\u002F\u002F127.0.0.1:8265\" \\\n  --runtime-env-json='{\"working_dir\": \"\u002Fopenrlhf\"}' \\\n  -- python3 -m openrlhf.cli.train_ppo_ray \\\n  --ref_num_nodes 1 --ref_num_gpus_per_node 8 \\\n  --reward_num_nodes 1 --reward_num_gpus_per_node 8 \\\n  --critic_num_nodes 1 --critic_num_gpus_per_node 8 \\\n  --actor_num_nodes 1 --actor_num_gpus_per_node 8 \\\n  --vllm_num_engines 4 --vllm_tensor_parallel_size 2 \\\n  --colocate_all_models \\\n  --vllm_gpu_memory_utilization 0.5 \\\n  --pretrain OpenRLHF\u002FLlama-3-8b-sft-mixture \\\n  --reward_pretrain OpenRLHF\u002FLlama-3-8b-rm-700k \\\n  --save_path .\u002Foutput\u002Fllama3-8b-rlhf \\\n  --micro_train_batch_size 8 --train_batch_size 128 \\\n  --micro_rollout_batch_size 16 --rollout_batch_size 1024 \\\n  --max_epochs 1 --prompt_max_len 1024 --generate_max_len 1024 \\\n  --zero_stage 3 --bf16 \\\n  --actor_learning_rate 5e-7 --critic_learning_rate 9e-6 \\\n  --init_kl_coef 0.01 --normalize_reward \\\n  --gradient_checkpointing --packing_samples \\\n  --vllm_enable_sleep --deepspeed_enable_sleep\n```\n\n**GRPO Training** (Group Normalized Policy Optimization):\n```bash\n# Same command as PPO, but add:\n--advantage_estimator group_norm\n```\n\n## Common workflows\n\n### Workflow 1: Full RLHF pipeline (SFT → Reward Model → PPO)\n\n**Step 1: Train reward model** (DPO):\n```bash\ndeepspeed --module openrlhf.cli.train_rm \\\n  --save_path .\u002Foutput\u002Fllama3-8b-rm \\\n  --save_steps -1 --logging_steps 1 \\\n  --eval_steps -1 --train_batch_size 256 \\\n  --micro_train_batch_size 1 --pretrain meta-llama\u002FMeta-Llama-3-8B \\\n  --bf16 --max_epochs 1 --max_len 8192 \\\n  --zero_stage 3 --learning_rate 9e-6 \\\n  --dataset OpenRLHF\u002Fpreference_dataset_mixture2_and_safe_pku \\\n  --apply_chat_template --chosen_key chosen \\\n  --rejected_key rejected --flash_attn --gradient_checkpointing\n```\n\n**Step 2: PPO training**:\n```bash\nray start --head --node-ip-address 0.0.0.0 --num-gpus 8\n\nray job submit --address=\"http:\u002F\u002F127.0.0.1:8265\" \\\n  -- python3 -m openrlhf.cli.train_ppo_ray \\\n  --ref_num_nodes 1 --ref_num_gpus_per_node 8 \\\n  --reward_num_nodes 1 --reward_num_gpus_per_node 8 \\\n  --critic_num_nodes 1 --critic_num_gpus_per_node 8 \\\n  --actor_num_nodes 1 --actor_num_gpus_per_node 8 \\\n  --vllm_num_engines 4 --vllm_tensor_parallel_size 2 \\\n  --colocate_all_models \\\n  --pretrain OpenRLHF\u002FLlama-3-8b-sft-mixture \\\n  --reward_pretrain .\u002Foutput\u002Fllama3-8b-rm \\\n  --save_path .\u002Foutput\u002Fllama3-8b-ppo \\\n  --micro_train_batch_size 8 --train_batch_size 128 \\\n  --micro_rollout_batch_size 16 --rollout_batch_size 1024 \\\n  --max_epochs 1 --prompt_max_len 1024 --generate_max_len 1024 \\\n  --zero_stage 3 --bf16 \\\n  --actor_learning_rate 5e-7 --critic_learning_rate 9e-6 \\\n  --init_kl_coef 0.01 --normalize_reward \\\n  --vllm_enable_sleep --deepspeed_enable_sleep\n```\n\n### Workflow 2: GRPO training (no critic model needed)\n\nMemory-efficient alternative to PPO:\n\n```bash\nray job submit --address=\"http:\u002F\u002F127.0.0.1:8265\" \\\n  -- python3 -m openrlhf.cli.train_ppo_ray \\\n  --advantage_estimator group_norm \\\n  --ref_num_nodes 1 --ref_num_gpus_per_node 8 \\\n  --reward_num_nodes 1 --reward_num_gpus_per_node 8 \\\n  --actor_num_nodes 1 --actor_num_gpus_per_node 8 \\\n  --vllm_num_engines 4 --vllm_tensor_parallel_size 2 \\\n  --colocate_all_models \\\n  --pretrain OpenRLHF\u002FLlama-3-8b-sft-mixture \\\n  --reward_pretrain OpenRLHF\u002FLlama-3-8b-rm-700k \\\n  --save_path .\u002Foutput\u002Fllama3-8b-grpo \\\n  --micro_train_batch_size 8 --train_batch_size 128 \\\n  --micro_rollout_batch_size 16 --rollout_batch_size 1024 \\\n  --max_epochs 1 --bf16 \\\n  --actor_learning_rate 5e-7 \\\n  --init_kl_coef 0.01 --use_kl_loss --kl_estimator k3 \\\n  --normalize_reward --no_advantage_std_norm\n```\n\n**Key GRPO parameters**:\n- `--advantage_estimator group_norm` - Enables GRPO\n- `--use_kl_loss` - KL loss from GRPO paper\n- `--kl_estimator k3` - Loss function (k2 ≈ k1)\n- `--no_advantage_std_norm` - Disables std normalization\n\n### Workflow 3: DPO training (preference optimization)\n\nSimpler alternative without reward model:\n\n```bash\ndeepspeed --module openrlhf.cli.train_dpo \\\n  --save_path .\u002Foutput\u002Fllama3-8b-dpo \\\n  --save_steps -1 --logging_steps 1 \\\n  --eval_steps -1 --train_batch_size 256 \\\n  --micro_train_batch_size 2 --pretrain meta-llama\u002FMeta-Llama-3-8B \\\n  --bf16 --max_epochs 1 --max_len 8192 \\\n  --zero_stage 3 --learning_rate 5e-7 --beta 0.1 \\\n  --dataset OpenRLHF\u002Fpreference_dataset_mixture2_and_safe_pku \\\n  --apply_chat_template --chosen_key chosen \\\n  --rejected_key rejected --flash_attn --gradient_checkpointing\n```\n\n## When to use vs alternatives\n\n**Use OpenRLHF when**:\n- Training large models (7B-70B+) with RL\n- Need vLLM inference acceleration\n- Want distributed architecture with Ray\n- Have multi-node GPU cluster\n- Need PPO\u002FGRPO\u002FRLOO\u002FDPO in one framework\n\n**Algorithm selection**:\n- **PPO**: Maximum control, best for complex rewards\n- **GRPO**: Memory-efficient, no critic needed\n- **RLOO**: Modified PPO with per-token KL\n- **REINFORCE++**: More stable than GRPO, faster than PPO\n- **DPO**: Simplest, no reward model needed\n\n**Use alternatives instead**:\n- **TRL**: Single-node training, simpler API\n- **veRL**: ByteDance's framework for 671B models\n- **DeepSpeedChat**: Integrated with DeepSpeed ecosystem\n\n## Common issues\n\n**Issue: GPU OOM with large models**\n\nDisable model colocation:\n```bash\n# Remove --colocate_all_models flag\n# Allocate separate GPUs for each model\n--actor_num_gpus_per_node 8 \\\n--critic_num_gpus_per_node 8 \\\n--reward_num_gpus_per_node 8 \\\n--ref_num_gpus_per_node 8\n```\n\n**Issue: DeepSpeed GPU index out of range**\n\nSet environment variable:\n```bash\nexport RAY_EXPERIMENTAL_NOSET_CUDA_VISIBLE_DEVICES=1\n```\n\n**Issue: Training instability**\n\nUse Hybrid Engine instead of async:\n```bash\n--colocate_all_models \\\n--vllm_enable_sleep \\\n--deepspeed_enable_sleep\n```\n\nAdjust KL coefficient:\n```bash\n--init_kl_coef 0.05  # Increase from 0.01\n```\n\n**Issue: Slow generation during PPO**\n\nEnable vLLM acceleration:\n```bash\n--vllm_num_engines 4 \\\n--vllm_tensor_parallel_size 2 \\\n--vllm_gpu_memory_utilization 0.5\n```\n\n## Advanced topics\n\n**Hybrid Engine GPU sharing**: See [references\u002Fhybrid-engine.md](references\u002Fhybrid-engine.md) for vLLM sleep mode, DeepSpeed sleep mode, and optimal node allocation.\n\n**Algorithm comparison**: See [references\u002Falgorithm-comparison.md](references\u002Falgorithm-comparison.md) for PPO vs GRPO vs RLOO vs REINFORCE++ benchmarks and hyperparameters.\n\n**Multi-node setup**: See [references\u002Fmulti-node-training.md](references\u002Fmulti-node-training.md) for Ray cluster configuration and fault tolerance.\n\n**Custom reward functions**: See [references\u002Fcustom-rewards.md](references\u002Fcustom-rewards.md) for reinforced fine-tuning and agent RLHF.\n\n## Hardware requirements\n\n- **GPU**: NVIDIA A100\u002FH100 recommended\n- **VRAM**:\n  - 7B model: 8× A100 40GB (Hybrid Engine)\n  - 70B model: 48× A100 80GB (vLLM:Actor:Critic = 1:1:1)\n- **Multi-node**: Ray cluster with InfiniBand recommended\n- **Docker**: NVIDIA PyTorch container 25.02+\n\n**Performance**:\n- 2× faster than DeepSpeedChat\n- vLLM inference acceleration\n- Hybrid Engine minimizes GPU idle time\n\n## Resources\n\n- Docs: https:\u002F\u002Fgithub.com\u002FOpenRLHF\u002FOpenRLHF\n- Paper: https:\u002F\u002Farxiv.org\u002Fabs\u002F2405.11143\n- Examples: https:\u002F\u002Fgithub.com\u002FOpenRLHF\u002FOpenRLHF\u002Ftree\u002Fmain\u002Fexamples\n- Discord: Community support\n\n\n\n",{"name":5,"tags":23,"author":36,"license":14,"version":37,"description":7,"dependencies":38},[24,25,26,27,28,29,30,31,32,33,34,35],"Post-Training","OpenRLHF","RLHF","PPO","GRPO","RLOO","DPO","Ray","vLLM","Distributed Training","Large Models","ZeRO-3","Orchestra Research","1.0.0",[39,40,41,42,43,44],"openrlhf","ray","vllm","torch","transformers","deepspeed","main","46e03791269608c5ac886a64c08f5d94147e8786","labs\u002F.cursor\u002Fskills\u002Fopenrlhf","https:\u002F\u002Fgithub.com\u002Fbeyondai\u002Fsandbox","https:\u002F\u002Fgithub.com\u002Fbeyondai","2026-11-08 04:18:42",{"source":8,"skillName":5,"commands":52,"prompt":62,"zipAvailable":63},[53,56,59],{"tool":54,"command":55},"npx","npx skills add beyondai\u002Fsandbox --skill openrlhf-training",{"tool":57,"command":58},"bunx","bunx skills add beyondai\u002Fsandbox --skill openrlhf-training",{"tool":60,"command":61},"pnpm","pnpm dlx skills add beyondai\u002Fsandbox --skill openrlhf-training","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fbeyondai\u002Fsandbox\u002Fblob\u002Fmain\u002Flabs\u002F.cursor\u002Fskills\u002Fopenrlhf\u002FSKILL.md",true,[65,71,75,79],{"path":66,"resourceType":67,"resourceTypeDesc":68,"sizeBytes":69,"sourceHtmlUrl":70},"references\u002Falgorithm-comparison.md","REFERENCE","参考文档",9783,"https:\u002F\u002Fgithub.com\u002Fbeyondai\u002Fsandbox\u002Fblob\u002Fmain\u002Flabs\u002F.cursor\u002Fskills\u002Fopenrlhf\u002Freferences\u002Falgorithm-comparison.md",{"path":72,"resourceType":67,"resourceTypeDesc":68,"sizeBytes":73,"sourceHtmlUrl":74},"references\u002Fcustom-rewards.md",15865,"https:\u002F\u002Fgithub.com\u002Fbeyondai\u002Fsandbox\u002Fblob\u002Fmain\u002Flabs\u002F.cursor\u002Fskills\u002Fopenrlhf\u002Freferences\u002Fcustom-rewards.md",{"path":76,"resourceType":67,"resourceTypeDesc":68,"sizeBytes":77,"sourceHtmlUrl":78},"references\u002Fhybrid-engine.md",7266,"https:\u002F\u002Fgithub.com\u002Fbeyondai\u002Fsandbox\u002Fblob\u002Fmain\u002Flabs\u002F.cursor\u002Fskills\u002Fopenrlhf\u002Freferences\u002Fhybrid-engine.md",{"path":80,"resourceType":67,"resourceTypeDesc":68,"sizeBytes":81,"sourceHtmlUrl":82},"references\u002Fmulti-node-training.md",11099,"https:\u002F\u002Fgithub.com\u002Fbeyondai\u002Fsandbox\u002Fblob\u002Fmain\u002Flabs\u002F.cursor\u002Fskills\u002Fopenrlhf\u002Freferences\u002Fmulti-node-training.md",[84,91,98,105,112,119],{"skillNo":85,"name":86,"nameZh":6,"description":87,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":88,"categories":89,"keywords":90,"publishedAt":18},"01M7D3GMPGJZRG8EPY404YF8TR","weights-and-biases","Track ML experiments with automatic logging, visualize training in real-time, optimize hyperparameters with sweeps, and manage model registry with W&B - collaborative MLOps platform","https:\u002F\u002Fgithub.com\u002Fbeyondai\u002Fsandbox\u002Fblob\u002Fmain\u002Flabs\u002F.opencode\u002Fskills\u002Fweights-and-biases\u002FSKILL.md",[],[],{"skillNo":92,"name":93,"nameZh":6,"description":94,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":95,"categories":96,"keywords":97,"publishedAt":18},"01M7D3GFZH440V2J532BSZMPVP","serving-llms-vllm","Serves LLMs with high throughput using vLLM's PagedAttention and continuous batching. Use when deploying production LLM APIs, optimizing inference latency\u002Fthroughput, or serving models with limited GPU memory. Supports OpenAI-compatible endpoints, quantization (GPTQ\u002FAWQ\u002FFP8), and tensor parallelism.","https:\u002F\u002Fgithub.com\u002Fbeyondai\u002Fsandbox\u002Fblob\u002Fmain\u002Flabs\u002F.opencode\u002Fskills\u002Fvllm\u002FSKILL.md",[],[],{"skillNo":99,"name":100,"nameZh":6,"description":101,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":102,"categories":103,"keywords":104,"publishedAt":18},"01M7D3GAE8MQCPGF72EP9G9GKY","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.","https:\u002F\u002Fgithub.com\u002Fbeyondai\u002Fsandbox\u002Fblob\u002Fmain\u002Flabs\u002F.opencode\u002Fskills\u002Fverl\u002FSKILL.md",[],[],{"skillNo":106,"name":107,"nameZh":6,"description":108,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":109,"categories":110,"keywords":111,"publishedAt":18},"01M7D3G7KKNE5RGNYJNXV4F0MM","unsloth","Expert guidance for fast fine-tuning with Unsloth - 2-5x faster training, 50-80% less memory, LoRA\u002FQLoRA optimization","https:\u002F\u002Fgithub.com\u002Fbeyondai\u002Fsandbox\u002Fblob\u002Fmain\u002Flabs\u002F.opencode\u002Fskills\u002Funsloth\u002FSKILL.md",[],[],{"skillNo":113,"name":114,"nameZh":6,"description":115,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":116,"categories":117,"keywords":118,"publishedAt":18},"01M7D3G2RCZQ08GS0KAW5CBWSC","fine-tuning-with-trl","Fine-tune LLMs using reinforcement learning with TRL - SFT for instruction tuning, DPO for preference alignment, PPO\u002FGRPO for reward optimization, and reward model training. Use when need RLHF, align model with preferences, or train from human feedback. Works with HuggingFace Transformers.","https:\u002F\u002Fgithub.com\u002Fbeyondai\u002Fsandbox\u002Fblob\u002Fmain\u002Flabs\u002F.opencode\u002Fskills\u002Ftrl-fine-tuning\u002FSKILL.md",[],[],{"skillNo":120,"name":121,"nameZh":6,"description":122,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":123,"categories":124,"keywords":125,"publishedAt":18},"01M7D3FX08BPB172CYGM597NZF","torchforge-rl-training","Provides guidance for PyTorch-native agentic RL using torchforge, Meta's library separating infra from algorithms. Use when you want clean RL abstractions, easy algorithm experimentation, or scalable training with Monarch and TorchTitan.","https:\u002F\u002Fgithub.com\u002Fbeyondai\u002Fsandbox\u002Fblob\u002Fmain\u002Flabs\u002F.opencode\u002Fskills\u002Ftorchforge\u002FSKILL.md",[],[]]