NovaSky-AI/SkyRL

Health v2 · maintenance only

Adoption not part of the score

2201 stars · 410 forks observed · 2026-08-28

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

SkyRL is a modular full-stack reinforcement learning library for post-training large language models, combining a training framework (skyrl-train), a Tinker API-compatible backend (skyrl-tx), an agent layer for long-horizon tasks (skyrl-agent), and a Gymnasium-style environment library (skyrl-gym). It supports FSDP and Megatron training backends with vLLM inference on your own NVIDIA GPUs, including LoRA and full-parameter fine-tuning.

Use cases

• train llm with reinforcement learning on my own gpus
• run tinker api training scripts locally without code changes
• fine-tune llm with grpo on gsm8k
• build custom gym environments for rlhf training
• train multi-turn tool-use agents with rl
• rl post-training for sql and search agents
• supervised fine-tuning of large language models with lora
• train vision-language models with reinforcement learning

When to choose

• you want to run RL post-training (GRPO, PPO, DAPO) for LLMs on your own NVIDIA hardware
• you need Tinker API compatibility with zero code changes on local GPUs
• you are training long-horizon, multi-turn tool-use agents
• you want a modular RL stack with pluggable environments, trainers, and inference engines
• you need to scale from 0.5B to 200B+ parameter models with FSDP or Megatron

When to avoid

• you only need simple SFT without RL and prefer lightweight fine-tuning tools
• you have no NVIDIA GPUs or CUDA 13.0-capable drivers
• you want a fully managed hosted training service rather than self-managed infrastructure
• you need RL for non-LLM domains like robotics control

Facets

library · maturity active

machine-learning llm-training agent-framework gpu-computing sdk large-language-models reinforcement-learning machine-learning gpu-computing python cloud reinforcement-learning rlhf grpo tinker-api fsdp megatron vllm ray lora post-training gymnasium-environments multi-turn-agents ai-agents gpu docker linux

Member repositories

For agents

markdown · JSON · MCP: product_card(name="NovaSky-AI/SkyRL")