slime-rl-training

Orchestrate Megatron-LM and SGLang for end-to-end RL post-training.

Updated May 20, 2026
One-click install
npx skills add https://github.com/SriRamkunamsetty/SITA2.0-HermesAgent --skill slime-rl-training-sriramkunamsetty
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: slime-rl-training
Source: https://github.com/SriRamkunamsetty/SITA2.0-HermesAgent/tree/main/hermes-agent/optional-skills/mlops/slime
Command: npx skills add https://github.com/SriRamkunamsetty/SITA2.0-HermesAgent --skill slime-rl-training-sriramkunamsetty

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

slime-rl-training guides ML teams to accelerate RL post-training of large language models by integrating Megatron-LM with SGLang, enabling scalable, reproducible RL workflows.

Core Features & Use Cases

  • End-to-end RL post-training orchestration combining Megatron-LM training with SGLang rollout, reward modeling, and weight synchronization.
  • Flexible data buffering, custom generate functions, and multi-turn, tool-enabled interactions for research and production experiments.
  • Pre-configured model scripts and diverse workflows (GRPO, async, multi-turn) with evaluation utilities for rapid iteration.

Quick Start

Install slime, load a pre-configured model script, and start an RL training job over your dataset with Megatron-LM and SGLang.

Frequently Asked Questions about slime-rl-training

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

FAQPage Schema
How do I orchestrate RL post-training for large language models using Megatron-LM and SGLang?

You can orchestrate RL post-training for large language models by integrating Megatron-LM for training and SGLang for rollout within the slime framework, automating weight synchronization and reward modeling.

Can I run GRPO-style training and multi-turn tool-enabled conversations with slime?

Yes, slime supports running GRPO-style training, asynchronous workflows, and multi-turn tool-enabled conversations across diverse datasets, utilizing pre-configured model scripts and custom generate functions.

What is the best way to scale reinforcement learning workflows for LLMs with configurable arguments?

Scaling reinforcement learning workflows for LLMs is handled by configuring Megatron and SGLang arguments within slime, enabling flexible data buffering and reward modeling for rapid iteration.

Does slime support custom generate functions and data buffering for RL experiments?

Yes, slime provides flexible data buffering and allows custom generate functions for multi-turn interactions, enabling ML engineers to run production and research RL experiments.

Do I need pre-configured model scripts to start an RL training job over my dataset?

You can load pre-configured model scripts to quickly start an RL training job over your dataset, utilizing the integrated Megatron-LM and SGLang environment for scalable execution.