slime-rl-training

Orchestrate Megatron-LM training and SGLang rollout generation for large-scale RL post-training.

Updated May 4, 2026
One-click install
npx skills add https://github.com/InverterNetwork/hermes-agent --skill slime-rl-training-inverternetwork
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: slime-rl-training
Source: https://github.com/InverterNetwork/hermes-agent/tree/main/optional-skills/mlops/slime
Command: npx skills add https://github.com/InverterNetwork/hermes-agent --skill slime-rl-training-inverternetwork

SYSTEM DOCUMENTATION & REQUIREMENTS

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

What problem does it solve?

This Skill addresses the complexity of scaling Reinforcement Learning (RL) for Large Language Models by providing a unified framework that bridges Megatron-LM training with high-throughput SGLang rollout generation.

Core Features & Use Cases

  • High-Throughput RL: Integrates Megatron-LM for actor training and SGLang for efficient rollout generation, supporting GRPO, PPO, and other algorithms.
  • Flexible Data Buffering: Manages prompt initialization, sample storage, and custom filtering logic for off-policy training.
  • Use Case: Use this framework to train reasoning models like GLM-4 or Qwen3 by automating the generation of rollouts, reward computation, and iterative policy updates across multi-node GPU clusters.

Quick Start

Execute the slime training script by providing the path to your prompt data and selecting a pre-configured model script from the models directory.

Frequently Asked Questions about slime-rl-training

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

FAQPage Schema
How do I run reinforcement learning post-training for LLMs with Megatron-LM and SGLang?

Reinforcement learning post-training is facilitated by orchestrating Megatron-LM for distributed actor training and SGLang for high-throughput rollout generation across multi-node GPU clusters. You execute a training script by providing prompt data and selecting a pre-configured model script.

Can I use GRPO and PPO for large-scale LLM training with this framework?

Yes, GRPO and PPO are supported for large-scale LLM training. The framework automates iterative policy updates by integrating Megatron-LM training with SGLang rollout generation and flexible data buffering for off-policy training.

Do I need Ray and high-performance GPU clusters to scale LLM post-training?

Yes, Ray and high-performance GPU clusters are required. Ray handles distributed task scheduling across multi-node environments, while the GPU clusters provide the necessary hardware acceleration for Megatron-LM model parallelism.

What is the best way to manage prompt initialization and sample storage during off-policy LLM training?

The best way to manage off-policy LLM training data is through the framework's flexible data buffering. It handles prompt initialization, sample storage, and custom filtering logic to support complex reinforcement learning algorithms like GRPO and PPO.

Does this framework support training reasoning models like GLM-4 or Qwen3?

Yes, training reasoning models like GLM-4 or Qwen3 is a primary use case. The framework automates the generation of rollouts, reward computation, and iterative policy updates across distributed multi-node GPU clusters.

Why does SGLang need to be integrated with Megatron-LM for RL post-training?

SGLang needs to be integrated with Megatron-LM to bridge high-throughput rollout generation with distributed actor training. This unified approach solves the complexity of scaling reinforcement learning for Large Language Models by automating the iterative update cycle.