Skip to content
Research Agent Skillsby Kalaris Labs
Install
ML Trainingv1.0.0AdaptedMIT

openrlhf-training

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.

Install this skill

Works with Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, OpenCode, and 70+ other harnesses.

skills.sh CLI (Recommended)
npx skills add KalarisLabs/research-agent-skills --skill openrlhf-training
uvx (Python tool runner)
uvx research-agent-skills install openrlhf-training

Documentation & Instructions

OpenRLHF - High-Performance RLHF Training

Quick start

OpenRLHF is a Ray-based RLHF framework optimized for distributed training with vLLM inference acceleration.

Installation:

# Launch Docker container
docker run --runtime=nvidia -it --rm --shm-size="10g" --cap-add=SYS_ADMIN \
  -v $PWD:/openrlhf nvcr.io/nvidia/pytorch:25.02-py3 bash

# Uninstall conflicts
sudo pip uninstall xgboost transformer_engine flash_attn pynvml -y

# Install OpenRLHF with vLLM
pip install openrlhf[vllm]

PPO Training (Hybrid Engine):

ray start --head --node-ip-address 0.0.0.0 --num-gpus 8

ray job submit --address="http://127.0.0.1:8265" \
  --runtime-env-json='{"working_dir": "/openrlhf"}' \
  -- python3 -m openrlhf.cli.train_ppo_ray \
  --ref_num_nodes 1 --ref_num_gpus_per_node 8 \
  --reward_num_nodes 1 --reward_num_gpus_per_node 8 \
  --critic_num_nodes 1 --critic_num_gpus_per_node 8 \
  --actor_num_nodes 1 --actor_num_gpus_per_node 8 \
  --vllm_num_engines 4 --vllm_tensor_parallel_size 2 \
  --colocate_all_models \
  --vllm_gpu_memory_utilization 0.5 \
  --pretrain OpenRLHF/Llama-3-8b-sft-mixture \
  --reward_pretrain OpenRLHF/Llama-3-8b-rm-700k \
  --save_path ./output/llama3-8b-rlhf \
  --micro_train_batch_size 8 --train_batch_size 128 \
  --micro_rollout_batch_size 16 --rollout_batch_size 1024 \
  --max_epochs 1 --prompt_max_len 1024 --generate_max_len 1024 \
  --zero_stage 3 --bf16 \
  --actor_learning_rate 5e-7 --critic_learning_rate 9e-6 \
  --init_kl_coef 0.01 --normalize_reward \
  --gradient_checkpointing --packing_samples \
  --vllm_enable_sleep --deepspeed_enable_sleep

GRPO Training (Group Normalized Policy Optimization):

# Same command as PPO, but add:
--advantage_estimator group_norm

Common workflows

Workflow 1: Full RLHF pipeline (SFT → Reward Model → PPO)

Step 1: Train reward model (DPO):

deepspeed --module openrlhf.cli.train_rm \
  --save_path ./output/llama3-8b-rm \
  --save_steps -1 --logging_steps 1 \
  --eval_steps -1 --train_batch_size 256 \
  --micro_train_batch_size 1 --pretrain meta-llama/Meta-Llama-3-8B \
  --bf16 --max_epochs 1 --max_len 8192 \
  --zero_stage 3 --learning_rate 9e-6 \
  --dataset OpenRLHF/preference_dataset_mixture2_and_safe_pku \
  --apply_chat_template --chosen_key chosen \
  --rejected_key rejected --flash_attn --gradient_checkpointing

Step 2: PPO training:

ray start --head --node-ip-address 0.0.0.0 --num-gpus 8

ray job submit --address="http://127.0.0.1:8265" \
  -- python3 -m openrlhf.cli.train_ppo_ray \
  --ref_num_nodes 1 --ref_num_gpus_per_node 8 \
  --reward_num_nodes 1 --reward_num_gpus_per_node 8 \
  --critic_num_nodes 1 --critic_num_gpus_per_node 8 \
  --actor_num_nodes 1 --actor_num_gpus_per_node 8 \
  --vllm_num_engines 4 --vllm_tensor_parallel_size 2 \
  --colocate_all_models \
  --pretrain OpenRLHF/Llama-3-8b-sft-mixture \
  --reward_pretrain ./output/llama3-8b-rm \
  --save_path ./output/llama3-8b-ppo \
  --micro_train_batch_size 8 --train_batch_size 128 \
  --micro_rollout_batch_size 16 --rollout_batch_size 1024 \
  --max_epochs 1 --prompt_max_len 1024 --generate_max_len 1024 \
  --zero_stage 3 --bf16 \
  --actor_learning_rate 5e-7 --critic_learning_rate 9e-6 \
  --init_kl_coef 0.01 --normalize_reward \
  --vllm_enable_sleep --deepspeed_enable_sleep

Workflow 2: GRPO training (no critic model needed)

Memory-efficient alternative to PPO:

ray job submit --address="http://127.0.0.1:8265" \
  -- python3 -m openrlhf.cli.train_ppo_ray \
  --advantage_estimator group_norm \
  --ref_num_nodes 1 --ref_num_gpus_per_node 8 \
  --reward_num_nodes 1 --reward_num_gpus_per_node 8 \
  --actor_num_nodes 1 --actor_num_gpus_per_node 8 \
  --vllm_num_engines 4 --vllm_tensor_parallel_size 2 \
  --colocate_all_models \
  --pretrain OpenRLHF/Llama-3-8b-sft-mixture \
  --reward_pretrain OpenRLHF/Llama-3-8b-rm-700k \
  --save_path ./output/llama3-8b-grpo \
  --micro_train_batch_size 8 --train_batch_size 128 \
  --micro_rollout_batch_size 16 --rollout_batch_size 1024 \
  --max_epochs 1 --bf16 \
  --actor_learning_rate 5e-7 \
  --init_kl_coef 0.01 --use_kl_loss --kl_estimator k3 \
  --normalize_reward --no_advantage_std_norm

Key GRPO parameters:

  • --advantage_estimator group_norm - Enables GRPO
  • --use_kl_loss - KL loss from GRPO paper
  • --kl_estimator k3 - Loss function (k2 ≈ k1)
  • --no_advantage_std_norm - Disables std normalization

Workflow 3: DPO training (preference optimization)

Simpler alternative without reward model:

deepspeed --module openrlhf.cli.train_dpo \
  --save_path ./output/llama3-8b-dpo \
  --save_steps -1 --logging_steps 1 \
  --eval_steps -1 --train_batch_size 256 \
  --micro_train_batch_size 2 --pretrain meta-llama/Meta-Llama-3-8B \
  --bf16 --max_epochs 1 --max_len 8192 \
  --zero_stage 3 --learning_rate 5e-7 --beta 0.1 \
  --dataset OpenRLHF/preference_dataset_mixture2_and_safe_pku \
  --apply_chat_template --chosen_key chosen \
  --rejected_key rejected --flash_attn --gradient_checkpointing

When to use vs alternatives

Use OpenRLHF when:

  • Training large models (7B-70B+) with RL
  • Need vLLM inference acceleration
  • Want distributed architecture with Ray
  • Have multi-node GPU cluster
  • Need PPO/GRPO/RLOO/DPO in one framework

Algorithm selection:

  • PPO: Maximum control, best for complex rewards
  • GRPO: Memory-efficient, no critic needed
  • RLOO: Modified PPO with per-token KL
  • REINFORCE++: More stable than GRPO, faster than PPO
  • DPO: Simplest, no reward model needed

Use alternatives instead:

  • TRL: Single-node training, simpler API
  • veRL: ByteDance's framework for 671B models
  • DeepSpeedChat: Integrated with DeepSpeed ecosystem

Common issues

Issue: GPU OOM with large models

Disable model colocation:

# Remove --colocate_all_models flag
# Allocate separate GPUs for each model
--actor_num_gpus_per_node 8 \
--critic_num_gpus_per_node 8 \
--reward_num_gpus_per_node 8 \
--ref_num_gpus_per_node 8

Issue: DeepSpeed GPU index out of range

Set environment variable:

export RAY_EXPERIMENTAL_NOSET_CUDA_VISIBLE_DEVICES=1

Issue: Training instability

Use Hybrid Engine instead of async:

--colocate_all_models \
--vllm_enable_sleep \
--deepspeed_enable_sleep

Adjust KL coefficient:

--init_kl_coef 0.05  # Increase from 0.01

Issue: Slow generation during PPO

Enable vLLM acceleration:

--vllm_num_engines 4 \
--vllm_tensor_parallel_size 2 \
--vllm_gpu_memory_utilization 0.5

Advanced topics

Hybrid Engine GPU sharing: See references/hybrid-engine.md for vLLM sleep mode, DeepSpeed sleep mode, and optimal node allocation.

Algorithm comparison: See references/algorithm-comparison.md for PPO vs GRPO vs RLOO vs REINFORCE++ benchmarks and hyperparameters.

Multi-node setup: See references/multi-node-training.md for Ray cluster configuration and fault tolerance.

Custom reward functions: See references/custom-rewards.md for reinforced fine-tuning and agent RLHF.

Hardware requirements

  • GPU: NVIDIA A100/H100 recommended
  • VRAM:
    • 7B model: 8× A100 40GB (Hybrid Engine)
    • 70B model: 48× A100 80GB (vLLM:Actor:Critic = 1:1:1)
  • Multi-node: Ray cluster with InfiniBand recommended
  • Docker: NVIDIA PyTorch container 25.02+

Performance:

  • 2× faster than DeepSpeedChat
  • vLLM inference acceleration
  • Hybrid Engine minimizes GPU idle time

Resources

Agent operating procedure

  1. Check the environment. Check GPU type, memory and driver/CUDA versions (nvidia-smi), framework versions, and dataset location and size.
  2. Pin down the inputs. Confirm formats, identifiers and parameters from the data or the user. Ask rather than guess any value that changes the result.
  3. Run a small version first. Do a smoke run: tiny model or subset, few steps, and confirm loss decreases and checkpoints save.
  4. Execute the full task using the instructions and references above.
  5. Validate the result. Track metrics on held-out data, compare against a baseline, and record seeds, configs and hardware.
  6. Report. State what was run (versions, commands, parameters), what was checked, and what is still uncertain.
If this happens Do this
CUDA out-of-memory Reduce batch size, enable gradient accumulation/checkpointing or mixed precision, or shard the model.
Loss is NaN or diverges Lower the learning rate, check data for invalid values, and enable gradient clipping.
A function, flag or endpoint in these instructions is missing in the installed version Check the installed version's own documentation (help(), --help, official docs), adapt, and tell the user. Never invent an API.
A required input, identifier or parameter is ambiguous Ask the user, or state the assumption explicitly before running.

Integrity rules

  • Never fabricate results, parameters, identifiers, citations or statistics. If something cannot be run or verified, say so plainly.
  • Never claim training results without logs; estimate compute cost before launching large jobs.
  • Treat version-specific details here as possibly outdated: confirm them against the official documentation for the installed version.
  • Ask before actions that cost money, consume shared GPUs or cloud quota, touch personal or patient data, or cannot be undone.

Related skills

  • fine-tuning-with-trl: Fine-tune LLMs using reinforcement learning with TRL - SFT for instruction tuning, DPO for preference alignment, PPO/GRPO for reward optimi…
  • verl-rl-training: Provides guidance for training LLMs with reinforcement learning using verl (Volcano Engine RL).
  • simpo-training: Simple Preference Optimization for LLM alignment.

Trigger Prompts & Evaluation Corpus

These evaluation prompts test that the agent correctly selects openrlhf-training for relevant tasks and leaves it out for near misses.

Should trigger (3)

  • should trigger

    “How do I run GRPO with OpenRLHF on 8 GPUs?”

    expects Gives an OpenRLHF GRPO launch using --advantage_estimator group_norm with --use_kl_loss and a kl_estimator setting. Notes that no critic is needed, and suggests a small smoke run and checking the installed version's --help before the full job.

  • should trigger

    “I'm running PPO on a Llama-3 8B SFT checkpoint (./ckpt/llama3-8b-sft) with a reward model at ./ckpt/rm-8b, on 2 nodes of 8×A100 80GB with a Ray cluster. Using the hybrid engine I get CUDA OOM during the actor update, and DeepSpeed also throws 'GPU index out of range'. How do I fix this?”

    expects Suggests disabling model colocation or adjusting the vLLM/actor/critic GPU allocation, and setting the environment variable for the DeepSpeed GPU index error. Also suggests reducing micro-batch size or enabling gradient checkpointing, asks for or states assumptions about the actual launch command, and checks versions with nvidia-smi and --help instead of inventing flags.

  • should trigger

    “I want to align a 70B model with a learned reward signal, with rollouts generated by a fast inference engine and the training spread over a multi-node GPU cluster managed by a distributed scheduler. Which approach and setup would you recommend, and what hardware would I need?”

    expects Recommends a Ray plus vLLM based RLHF setup such as OpenRLHF, with algorithm guidance (PPO vs GRPO vs REINFORCE++). Gives a rough hardware estimate (for example, about 48× A100 80GB with a 1:1:1 vLLM:Actor:Critic split), mentions InfiniBand, and warns that these figures are approximate and should be confirmed. Asks about budget and shared cluster use before launching anything.

Near misses (2)

  • near miss

    “I have a single RTX 4090 and want to do DPO on Qwen2.5-1.5B with Hugging Face TRL's DPOTrainer and LoRA. Can you write the script?”

    expects Writes a TRL DPOTrainer script with a LoRA config for a single GPU. It does not use OpenRLHF, Ray or a multi-node setup.

  • near miss

    “Help me set up verl with the FSDP backend to run GRPO on a Qwen 72B model, and explain how to configure the rollout workers.”

    expects Uses verl's own configuration and rollout worker setup (Volcano Engine RL), not OpenRLHF commands or flags.