NVIDIA-NeMo/RL

Health v2 · maintenance only

Adoption not part of the score

1961 stars · 534 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

NeMo RL is NVIDIA's open-source post-training library for scaling reinforcement learning methods (GRPO, PPO, DPO, SFT, distillation) on LLMs and vision-language models. It supports PyTorch DTensor and Megatron Core training backends with vLLM/SGLang generation, Ray-based orchestration, and multi-node Slurm or Kubernetes deployment.

Use cases

• train an LLM with GRPO on math reasoning tasks
• run RLHF post-training on a Qwen or Llama model across multiple GPUs
• fine-tune a vision-language model with SFT or GRPO
• perform on-policy distillation from a teacher model
• scale RL training on a multi-node Slurm or Kubernetes cluster
• reproduce NVIDIA Nemotron model post-training recipes
• run multi-turn RL with tool use environments

When to choose

• you need scalable, multi-node RL post-training for large LLMs or VLMs
• you want Megatron Core parallelism (TP/PP/CP/EP) for very large models
• you want Hugging Face model support with fast vLLM rollouts
• you need reproducible recipes for GRPO, PPO, DPO, or distillation

When to avoid

• you only need simple single-GPU supervised fine-tuning without RL
• you have no NVIDIA GPUs or CUDA environment
• you need a lightweight library without Docker or cluster infrastructure
• you are training small non-transformer models with standard RL algorithms

Facets

library · maturity active

machine-learning llm-training reinforcement-learning gpu-computing benchmarking large-language-models machine-learning deep-learning gpu-computing developer-tools python cloud post-training grpo ppo dpo sft rlhf vllm megatron ray vlm distillation multi-node-training linux gpu docker kubernetes

Member repositories

For agents

markdown · JSON · MCP: product_card(name="NVIDIA-NeMo/RL")