slime-rl-training

Integrates Megatron-LM and SGLang for scalable GRPO-based LLM reinforcement post-training.

1|Updated Feb 21, 2026
One-click install
npx skills add https://github.com/tianhao909/AI-Research-SKILLs-cn --skill slime-rl-training-tianhao909
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: slime-rl-training
Source: https://github.com/tianhao909/AI-Research-SKILLs-cn/tree/main/06-post-training/slime
Command: npx skills add https://github.com/tianhao909/AI-Research-SKILLs-cn --skill slime-rl-training-tianhao909

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill requires sglang-router>=0.2.3, ray, torch>=2.0.0, transformers>=4.40.0, and includes scripts (resource) and references (resource) components.

What problem does it solve?

This Skill streamlines the complex process of post-training Large Language Models (LLMs) using Reinforcement Learning (RL), specifically by integrating Megatron-LM for efficient training and SGLang for high-throughput generation.

Core Features & Use Cases

  • RL Post-Training: Enables advanced RL techniques like GRPO for LLM fine-tuning.
  • Megatron-LM Integration: Leverages Megatron-LM's distributed training capabilities for large models.
  • SGLang Rollout: Utilizes SGLang for fast and scalable response generation during RL rollouts.
  • Use Case: Fine-tune a Llama 3 model using Reinforcement Learning from Human Feedback (RLHF) to improve its conversational abilities and adherence to safety guidelines, leveraging distributed training for faster iteration.

Quick Start

Launch GRPO training for a Qwen3 model using the provided script and your training data.

Frequently Asked Questions about slime-rl-training

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I scale reinforcement learning post-training for large language models?

To scale reinforcement learning post-training for LLMs, you can utilize frameworks that integrate distributed training engines like Megatron-LM with high-throughput generation engines like SGLang to handle GRPO and RLHF efficiently.

Can I use Megatron-LM and SGLang for GRPO training on Llama 3 models?

Yes, you can use Megatron-LM and SGLang for GRPO training on Llama 3 models. This combination leverages Megatron-LM for distributed training and SGLang for scalable rollout generation during reinforcement learning.

What is the best way to perform high-throughput generation rollouts during LLM reinforcement learning?

The best way to perform high-throughput generation rollouts during LLM reinforcement learning is using SGLang. It provides fast and scalable response generation required for efficient RL workflows like GRPO.

Does this reinforcement learning framework support training for Qwen3 and DeepSeek V3 models?

Yes, this reinforcement learning framework supports training for Qwen3 and DeepSeek V3 models. It is designed to offer a research-grade environment with production backing for these specific architectures.

Do I need Ray and PyTorch to run distributed LLM post-training with GRPO?

Yes, you need Ray and PyTorch to run distributed LLM post-training with GRPO. These dependencies are required alongside Megatron-LM and SGLang to facilitate scalable reinforcement learning workflows.

How do I improve a model's conversational abilities and safety using RLHF?

To improve conversational abilities and safety using RLHF, you fine-tune models like Llama 3 with reinforcement learning. This process adjusts responses based on human feedback using distributed training for faster iteration.