slime-rl-training

Integrate Megatron-LM with SGLang for RL-based post-training of large language models.

Updated Aug 23, 2026
One-click install
npx skills add https://github.com/travelinman1013/leroys-agent --skill slime-rl-training-travelinman1013
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: slime-rl-training
Source: https://github.com/travelinman1013/leroys-agent/tree/main/optional-skills/mlops/slime
Command: npx skills add https://github.com/travelinman1013/leroys-agent --skill slime-rl-training-travelinman1013

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

RL-based post-training of large language models (LLMs) requires an integrated workflow to coordinate large-scale data generation, policy optimization, and distributed training. slime provides a unified RL post-training framework that combines Megatron-LM with SGLang to enable scalable, end-to-end training loops.

Core Features & Use Cases

  • Training: Megatron-LM with full parallelism support (TP, PP, DP, SP)
  • Rollout: SGLang-based high-throughput generation with router
  • Data Management: Flexible prompt buffers and sample storage for efficient RL pipelines
  • Use Case: Develop and refine RL policies for GLM-like models with automated data generation and reward tuning.

Quick Start

Run the slime training workflow with your dataset and model configuration to start RL-based post-training.

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 RL post-training for large language models with Megatron-LM and SGLang?

RL post-training for large language models is executed by integrating Megatron-LM for distributed policy optimization with SGLang for scalable rollout generation. You configure your dataset, model parameters, and reward models to run the end-to-end training loop.

What is the best way to scale rollout generation for LLM reinforcement learning?

Scaling rollout generation for LLM reinforcement learning is handled by SGLang, which provides high-throughput generation with a router. This works alongside Megatron-LM to coordinate data buffers and reward-driven optimization.

Can I use slime for RL post-training on GLM-family transformer architectures?

Yes, slime supports RL-based post-training for GLM-family models and other transformer architectures. It applies data-generation workflows, policy optimization, and distributed training to refine these models.

What parallelism strategies are supported during RL training with Megatron-LM?

RL training with Megatron-LM supports full parallelism including tensor, pipeline, data, and sequence parallelism. This enables distributed training at scale for large language models during the post-training phase.

How do I manage prompt data and samples for reinforcement learning pipelines?

Reinforcement learning pipelines manage prompt data and samples using flexible prompt buffers and sample storage. This supports efficient data generation and reward tuning within the training workflow.

Do I need a distributed compute environment for LLM reinforcement learning with slime?

Yes, a distributed compute environment is required for LLM reinforcement learning with slime. The framework relies on Megatron-LM and SGLang to support scalable rollout generation and distributed training across multiple nodes.