slime-rl-training

Integrate Megatron-LM training with SGLang rollout for RL fine-tuning.

Updated Apr 19, 2026
One-click install
npx skills add https://github.com/gqf2008/hermez-ai --skill slime-rl-training-gqf2008
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: slime-rl-training
Source: https://github.com/gqf2008/hermez-ai/tree/main/skills/mlops/slime
Command: npx skills add https://github.com/gqf2008/hermez-ai --skill slime-rl-training-gqf2008

SYSTEM DOCUMENTATION & REQUIREMENTS

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

What problem does it solve?

slime provides a structured approach for post-training reinforcement learning of large language models by bridging Megatron-LM training with SGLang-driven rollout, enabling scalable RL fine-tuning and managed data workflows.

Core Features & Use Cases

  • Megatron-LM based training with full parallelism and SGLang rollout
  • Flexible data generation workflows and prompt/data buffers for scalable RL experiments
  • Supports GLM-family models and related architectures (Qwen3, DeepSeek V3, Llama 3) with production-grade tooling
  • Use cases include research-scale RL, rapid experimentation, and production-ready RL training pipelines

Quick Start

Install slime, choose a pre-configured model script under scripts/models, and start the GRPO-based training workflow using the provided arguments.

Frequently Asked Questions about slime-rl-training

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

FAQPage Schema
How do I streamline reinforcement learning fine-tuning for large language models using Megatron-LM?

You can streamline RL fine-tuning by integrating Megatron-LM training with SGLang-driven rollout to enable scalable post-training and managed data workflows. This approach bridges high-throughput data generation with parallel training.

Does RL post-training with SGLang support GLM-family models and Qwen3?

RL post-training supports GLM-family models and related architectures including Qwen3, DeepSeek V3, and Llama 3. It applies production-grade tooling to these models for research-scale and production-ready training pipelines.

Can I use asynchronous training and multi-turn workflows for LLM post-training?

You can use asynchronous training, GRPO training, and custom multi-turn workflows for LLM post-training. These standard workflows are supported through specific data buffers, scripts, and configuration options.

What do I need to set up before starting GRPO-based training with slime?

You need a Megatron-LM setup, SGLang routing, and slime-specific data buffers, scripts, and configuration. Required dependencies include sglang-router, ray, torch, and transformers to ensure proper environment readiness.

What's the best way to start scalable RL experiments after installing slime?

After installing slime, choose a pre-configured model script under scripts/models and start the GRPO-based training workflow using provided arguments. This facilitates rapid experimentation and scalable reinforcement learning.