skills.rest
eddielueng
eddieluengCOMMUNITY
Explore all skills by this author →
GitHub

slime-rl-training

Run GRPO RL post-training for LLMs with Megatron-LM and SGLang rollouts.

1|Updated Apr 14, 2026
Search Tags:#reinforcement learning#distributed training#sglang#grpo#megatron-lm#llm post-training#agentic data generation
One-click install
npx skills add https://github.com/eddielueng/hermes-agent-zh --skill slime-rl-training-eddielueng
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: slime-rl-training
Source: https://github.com/eddielueng/hermes-agent-zh/tree/main/optional-skills/mlops/slime
Command: npx skills add https://github.com/eddielueng/hermes-agent-zh --skill slime-rl-training-eddielueng

SYSTEM DOCUMENTATION & REQUIREMENTS

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

What problem does it solve?

LLM teams need a practical way to do post-training with reinforcement learning while generating rollouts at high throughput and keeping tight integration between training and inference.

Core Features & Use Cases

  • Megatron-LM + SGLang integration for RL post-training: orchestrates parallel training (TP/PP/DP/SP) and high-throughput rollout generation using SGLang with a router.
  • Group-based RL algorithms and KL regularization: supports GRPO and related advantage estimators, including optional KL loss for stability.
  • Data buffering and flexible prompt workflows: manages prompt initialization, custom data generation/filtering, and off-policy style buffered reuse of generated samples.
  • Use Case: Train a GLM or other large reasoning model using GRPO/variants by repeatedly generating multiple responses per prompt, computing rewards, and updating the policy.

Quick Start

Run GRPO RL post-training by asking your AI to launch slime training with a sourced model config, a JSONL dataset path, and batch/rule parameters for rollout generation and policy updates.

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 reinforcement learning post-training with high-throughput rollout generation?▼

Scale LLM reinforcement learning post-training by integrating Megatron-LM training with SGLang-based rollout generation using slime. This orchestrates parallel training and high-throughput inference to efficiently update policies.

Can I use slime for GRPO training on Qwen and Llama models?▼

Yes, you can use slime for GRPO training on Qwen and Llama models. The framework applies to GLM and DeepSeek families as well, supporting group-based RL algorithms and optional KL regularization.

How does distributed training handle rollout batch constraints and off-policy sampling?▼

Distributed training handles rollout batch constraints and off-policy sampling through data buffering strategies. This manages prompt initialization, custom data filtering, and flexible reuse of generated samples for policy updates.

Do I need Ray and Megatron-LM to run multi-turn agentic training?▼

Yes, you need Ray and Megatron-LM setup to run multi-turn agentic training. The workflow requires a correct dataset schema mapping alongside SGLang router configurations to ensure scalable RL execution.

What is the best way to generate custom data for reinforcement learning workflows?▼

The best way to generate custom data for reinforcement learning workflows is using slime's built-in prompt workflows. It manages custom data generation and filtering while applying buffered off-policy sampling for generated samples.

Related Skills

Explore More Software Engineering Skills→
ovachieverovachiever
52

grpo-rl-training

Apply GRPO/RL training with TRL to optimize language models.

Community
Advanced
carterwayneskhizeinecarterwayneskhizeine
19

slime-rl-training

Orchestrate Megatron-LM training with SGLang rollout generation for RL post-training.

Community
Advanced
heysuhasheysuhas
2

slime-rl-training

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

Community
Advanced
evalstateevalstate

trl

Submit TRL training jobs on Hugging Face Jobs and convert models to GGUF.

Community
Advanced