slime-rl-training

Run GRPO training jobs for Qwen3-4B with Megatron-LM and SGLang.

Updated Aug 27, 2026
One-click install
npx skills add https://github.com/DoanNgocCuong/continuous-training-pipeline_T3_2026 --skill slime-rl-training-doanngoccuong
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: slime-rl-training
Source: https://github.com/DoanNgocCuong/continuous-training-pipeline_T3_2026/tree/main/.claude/skills/slime
Command: npx skills add https://github.com/DoanNgocCuong/continuous-training-pipeline_T3_2026 --skill slime-rl-training-doanngoccuong

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes scripts (resource) and references (resource) components.

What problem does it solve?

This Skill streamlines the complex process of post-training Large Language Models (LLMs) using Reinforcement Learning (RL), specifically designed for scaling and integrating with high-performance training frameworks.

Core Features & Use Cases

  • Megatron-LM Integration: Native support for Megatron-LM's advanced parallelism (TP, PP, DP, SP) for efficient training of massive models.
  • High-Throughput Rollout: Leverages SGLang for fast response generation, crucial for RL data collection.
  • Flexible Data Handling: Supports custom data generation workflows and flexible prompt management.
  • Use Case: Fine-tuning a large language model like Qwen3 or Llama 3 for a specific task using RLHF, ensuring efficient GPU utilization and scalable data processing.

Quick Start

Use the slime-rl-training skill to launch a GRPO training job for the qwen3-4B model using data from /path/to/train.jsonl.

Frequently Asked Questions about slime-rl-training

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

FAQPage Schema
How do I fine-tune a large language model with Reinforcement Learning using Megatron-LM?

To fine-tune a large language model with Reinforcement Learning, you can use this framework to integrate Megatron-LM for distributed training and SGLang for high-throughput inference. It supports GRPO algorithms to streamline post-training workflows.

Can I use SGLang for high-throughput rollout during RLHF training?

Yes, you can use SGLang for high-throughput rollout during RLHF training. This framework leverages SGLang specifically to facilitate fast response generation, which is crucial for efficient RL data collection and scaling pipelines.

Does this RL training framework support Qwen and Llama models?

Yes, this RL training framework supports Qwen and Llama models, alongside GLM and DeepSeek. You can launch GRPO training jobs for these models, such as a qwen3-4B model, using your custom prompt data.

What is the best way to scale distributed LLM post-training with GRPO?

The best way to scale distributed LLM post-training with GRPO is utilizing native Megatron-LM parallelism, including TP, PP, DP, and SP. This ensures efficient GPU utilization and scalable data processing for massive models.

How do I handle custom data generation for LLM Reinforcement Learning?

You can handle custom data generation for LLM Reinforcement Learning by using the framework's flexible data handling workflows. It supports custom prompt management and data generation from JSONL files to drive your GRPO training jobs.

Why use Megatron-LM and SGLang together for LLM RL training?

You use Megatron-LM and SGLang together for LLM RL training to combine advanced distributed training parallelism with high-throughput inference. This integration solves the bottleneck of fast response generation during RL data collection.