skills.rest
devMoez
devMoezCOMMUNITY
Explore all skills by this author →
GitHub

slime-rl-training

Coordinate Megatron-LM training with SGLang rollouts for GRPO-style RL post-training.

1|Updated May 16, 2026
Search Tags:#data pipeline#reinforcement learning#distributed training#sglang#grpo#megatron-lm#llm post-training
One-click install
npx skills add https://github.com/devMoez/titan --skill slime-rl-training-devmoez
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: slime-rl-training
Source: https://github.com/devMoez/titan/tree/main/optional-skills/mlops/slime
Command: npx skills add https://github.com/devMoez/titan --skill slime-rl-training-devmoez

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill requires sglang-router, ray, torch, transformers, and includes references (resource) components.

What problem does it solve?

Helps you run LLM post-training with reinforcement learning by connecting Megatron-LM training with SGLang rollout generation for efficient GRPO-style workflows.

Core Features & Use Cases

  • Megatron-LM RL Training: Runs actor/critic (optional) training with TP/PP/DP/SP parallelism for large models.
  • SGLang High-Throughput Rollouts: Generates rollout responses via SGLang and a router for scalable sample collection.
  • Data Buffer & Custom Workflows: Manages prompts, stores samples, supports off-policy buffering, and allows multi-turn/tool-using generation and reward logic.
  • Use Case: Post-train a GLM-family model using GRPO on JSONL prompts and labels, producing improved reasoning performance by repeatedly generating rollouts and updating the policy.

Quick Start

Ask Titan to train GLM with slime in GRPO mode using your prompt data: run the GRPO training command and point --prompt-data to your JSONL dataset while sourcing the model config script.

Frequently Asked Questions about slime-rl-training

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

FAQPage Schema
How do I scale LLM post-training with reinforcement learning using Megatron-LM and SGLang?▼

To scale LLM post-training with reinforcement learning, slime connects Megatron-LM for actor/critic training with SGLang router-based rollouts for scalable sample collection. It applies GRPO-style workflows to generate rollout responses and update the policy efficiently.

Can I use GRPO training for multi-turn or tool-calling agent scenarios?▼

Yes, GRPO training supports multi-turn and tool-calling agent scenarios. Slime manages custom data-generation buffers, off-policy buffering, and reward logic to facilitate complex generation workflows during LLM post-training.

What is the best way to configure distributed training parallelism for LLM post-training?▼

The best way to configure distributed training parallelism is using Megatron-LM, which runs actor and optional critic training with TP, PP, DP, and SP parallelism for large models during LLM post-training.

Does slime support custom data pipelines for reinforcement learning rollouts?▼

Yes, slime supports custom data pipelines for RL rollouts by managing prompts and storing samples in data buffers. It allows off-policy buffering and custom reward logic integration for your specific generation requirements.

How do I start GRPO training on a GLM-family model with JSONL prompt data?▼

To start GRPO training on a GLM-family model, run the GRPO training command and point the --prompt-data parameter to your JSONL dataset while sourcing the model config script to drive rollout generation and policy updates.

What RL loop parameters can I control during LLM post-training?▼

During LLM post-training, you can control RL loop parameters including advantage estimators and KL-penalty controls. Satisfying requirements for Megatron-LM parallel training and SGLang router-based rollouts ensures proper policy optimization.

Related Skills

Explore More Software Engineering Skills→
InverterNetworkInverterNetwork

slime-rl-training

Orchestrate Megatron-LM training and SGLang rollout generation for large-scale RL post-training.

Official
Advanced
zeronx798zeronx798

evaluating-llms-harness

Benchmark large language models across academic and reasoning datasets.

Community
Advanced
heysuhasheysuhas
2

slime-rl-training

Integrate Megatron-LM and SGLang for large-scale LLM reinforcement learning training.

Community
Advanced
jacardljacardl

slime-rl-training

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

Community
Advanced