slime-rl-training

Scale RL post-training for LLMs with Megatron-LM and SGLang.

Updated Apr 11, 2026
One-click install
npx skills add https://github.com/hhhi21g/HealthCenter --skill slime-rl-training-hhhi21g
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: slime-rl-training
Source: https://github.com/hhhi21g/HealthCenter/tree/main/.codex/skills/slime
Command: npx skills add https://github.com/hhhi21g/HealthCenter --skill slime-rl-training-hhhi21g

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 addresses the challenge of scaling Reinforcement Learning (RL) for large Language Models (LLMs), particularly with the integration of Megatron-LM and SGLang, to streamline the post-training process of LLMs.

Core Features & Use Cases

  • LLM Post-Training: Provides a framework for post-training LLMs using reinforcement learning.
  • Megatron-LM Integration: Enables seamless training of models with Megatron-LM.
  • SGLang Rollout: Integrates SGLang for high-throughput rollout generation.
  • Use Case: Ideal for training GLM models or implementing custom data generation workflows that require tight integration with Megatron-LM for RL scaling.

Quick Start

Run the following command to initiate the GRPO training process: python train.py ...

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 for LLM post-training with Megatron-LM?

Scaling RL for LLM post-training involves integrating Megatron-LM for model training and SGLang for high-throughput rollout generation. This framework provides the necessary infrastructure to manage and execute RL at scale.

Can I use SGLang for high-throughput rollout generation in LLM post-training?

Yes, SGLang is integrated into the framework specifically to enable high-throughput rollout generation during the reinforcement learning phase. It works in tandem with Megatron-LM to streamline the post-training process of large language models.

What dependencies do I need to run reinforcement learning training for GLM models?

You need specific software dependencies including sglang-router (>=0.2.3), ray, torch (>=2.0.0), and transformers (>=4.40.0). These packages provide the necessary environment for scaling reinforcement learning and executing the training script.

Does this framework support custom data generation workflows for LLM post-training?

Yes, the framework is ideal for implementing custom data generation workflows that require tight integration with Megatron-LM. It enables seamless training of models like GLM while handling reinforcement learning scaling.

What is the best way to initiate GRPO training for large language models?

The best way to initiate GRPO training is by running the `python train.py` command with the appropriate arguments. This starts the reinforcement learning process using the integrated Megatron-LM and SGLang backend.

Why integrate Megatron-LM with SGLang for reinforcement learning scaling?

Integrating Megatron-LM with SGLang addresses the challenge of scaling RL by combining efficient distributed training with high-throughput rollout generation. This synergy streamlines the post-training phase for large language models.