skills.rest
ruiyangruiyi
ruiyangruiyiCOMMUNITY
Explore all skills by this author →
GitHub

slime-rl-training

Orchestrates Megablock-LM RL training with SGLab-based rollout generation for GRPO-style workflows.

Updated May 26, 2026
Search Tags:#reinforcement learning#distributed training#sglang#grpo#megatron-lm#llm post-training#rl fine-tuning
One-click install
npx skills add https://github.com/ruiyangruiyi/hermes-agent --skill slime-rl-training-ruiyangruiyi
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: slime-rl-training
Source: https://github.com/ruiyangruiyi/hermes-agent/tree/main/optional-skills/mlops/slime
Command: npx skills add https://github.com/ruiyangruiyi/hermes-agent --skill slime-rl-training-ruiyangruiyi

SYSTEM DOCUMENTATION & REQUIREMENTS

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

What problem does it solve?

Provides practical guidance for post-training large language models with reinforcement learning (RL), coordinating Megatron-LM training and SGLang-based high-throughput rollouts so you can scale RL fine-tuning for GLM-family and related models.

Core Features & Use Cases

  • Megatron-LM RL Training: Runs actor/critic training with tensor/pipeline/data parallelism for scalable optimization.
  • SGLang Rollout Generation with Routing: Generates rollouts efficiently through SGLang routers to support high-throughput reward collection.
  • Flexible Data Buffer & Workflows: Manages prompts and generated samples with off-policy buffering and custom generation functions for multi-turn or tool-use settings.
  • Algorithms & Config Options: Supports GRPO-style group-relative advantage training, KL loss options, asynchronous training, and evaluation across multiple tasks.

Quick Start

Ask the AI to walk you through running GRPO training by sourcing a model script, selecting rollout and batch parameters, and launching train.py with your JSONL prompt dataset and label keys.

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 reinforcement learning fine-tuning on LLMs using Megatron-LM and SGLang?▼

Reinforcement learning fine-tuning is streamlined by orchestrating Megatron-LM actor and critic training with SGLang rollout generation. You can launch the train.py script with your JSONL prompt dataset, selecting rollout and batch parameters to scale RL optimization efficiently.

Can I use GRPO workflows for qwen3 and deepseek-style models?▼

GRPO workflows are supported for qwen3, deepseek-style, and GLM-4.x models. The system applies group-relative advantage training, allowing you to configure GRPO-related arguments, KL loss options, and asynchronous training loops for these specific LLM architectures.

How does SGLang router handle high-throughput rollout generation during distributed training?▼

SGLang router generates rollouts efficiently by managing high-throughput reward collection alongside Megatron-LM parallelism. It wires the rollout engine to support compatible tensor, pipeline, and data parallelism during scalable RL post-training optimization.

Do I need Ray and PyTorch to manage multi-turn agentic training loops with a data buffer?▼

Ray and PyTorch are required dependencies to manage multi-turn agentic training loops. The flexible data buffer system handles off-policy buffering and custom generation functions, enabling complex tool-calling scenarios and prompt-sample workflows within your distributed environment.

What is the best way to configure custom data generation pipelines for RL post-training?▼

Configuring custom data generation pipelines involves using slime's data buffer system with your JSONL prompt datasets and label keys. You can define custom generation functions for multi-turn or tool-use settings, managing off-policy buffering for GRPO-style training.

Does asynchronous training support evaluation across multiple tasks during RL fine-tuning?▼

Asynchronous training supports evaluation across multiple tasks during RL fine-tuning. You can configure GRPO-related arguments and KL loss options to run concurrent evaluations while maintaining scalable optimization through Megatron-LM and SGLang integration.

Related Skills

Explore More Software Engineering Skills→
zhuangbiaoweizhuangbiaowei
2

slime-rl-training

Trains large language models with GRPO reinforcement learning using Megatron-LM and SGLang.

Community
Advanced
hhhi21ghhhi21g

slime-rl-training

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

Community
Advanced
qcmuuqcmuu

torchforge-rl-training

Train PyTorch-native reinforcement learning with GRPO objectives and vLLM rollouts.

Community
Advanced
OpenLAIROpenLAIR
1.0k

verl-rl-training

Configure Verl backends and launch GRPO-based RL fine-tuning jobs.

Official
Advanced