skills.rest
blueskies1818
blueskies1818COMMUNITY
Explore all skills by this author →
WebsiteGitHub

slime-rl-training

Guide LLM post-training with RL using Megatron-LM and SGLang.

1|Updated May 21, 2026
Search Tags:#LLM#GLM#RL#SGLang#Megatron-LM#Reinforcement Learning
One-click install
npx skills add https://github.com/blueskies1818/hermesALIone --skill slime-rl-training-blueskies1818
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: slime-rl-training
Source: https://github.com/blueskies1818/hermesALIone/tree/main/Agent/optional-skills/mlops/slime
Command: npx skills add https://github.com/blueskies1818/hermesALIone --skill slime-rl-training-blueskies1818

SYSTEM DOCUMENTATION & REQUIREMENTS

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

What problem does it solve?

This Skill provides guidance for LLM post-training with Reinforcement Learning (RL) using slime, a Megatron+SGLang framework, enabling efficient training of GLM models and custom data generation workflows.

Core Features & Use Cases

  • Megatron-LM Training: Offers full parallelism support for actor model training with Megatron-LM.
  • SGLang Rollout: Utilizes SGLang for high-throughput rollout generation.
  • Data Buffer: Provides flexible prompt management and sample storage.
  • Model Support: Supports GLM, Qwen3, DeepSeek V3/R1, Llama 3 models.
  • Use Case: Ideal for scaling RL in Megatron-LM with tight integration and custom data generation capabilities.

Quick Start

Run the following command to initiate a GRPO training session:

python train.py \
    --actor-num-nodes 1 \
    --actor-num-gpus-per-node 4 \
    --rollout-num-gpus 4 \
    --advantage-estimator grpo \
    --use-kl-loss --kl-loss-coef 0.001 \
    --rollout-batch-size 32 \
    --n-samples-per-prompt 8 \
    --global-batch-size 256 \
    --num-rollout 3000 \
    --prompt-data /path/to/data.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 scale reinforcement learning post-training for LLMs using Megatron-LM?▼

You can scale reinforcement learning post-training for LLMs using Megatron-LM by utilizing the slime framework, which connects Megatron-LM for full parallelism actor training and SGLang for high-throughput rollout generation.

Can I train GLM models with reinforcement learning and SGLang?▼

Yes, you can train GLM models with RL and SGLang. The framework supports GLM, Qwen3, DeepSeek V3/R1, and Llama 3 models, using SGLang specifically for high-throughput rollout generation.

What do I need to run GRPO training for large language models?▼

To run GRPO training for large language models, you need Linux or macOS, specific dependencies including sglang-router, ray, torch, and transformers, and a JSONL prompt dataset to initiate the training session.

Does the slime framework support custom data generation workflows for RL?▼

Yes, the slime framework supports custom data generation workflows for RL. It provides a flexible data buffer for prompt management and sample storage alongside its SGLang rollout generation capabilities.

How does SGLang integrate with Megatron-LM for LLM post-training?▼

SGLang integrates with Megatron-LM for LLM post-training through the slime framework, where SGLang handles high-throughput rollout generation and Megatron-LM manages the actor model training with full parallelism support.

Related Skills

Explore More Software Engineering Skills→
richardnguyen0715richardnguyen0715

slime-rl-training

Train LLMs with reinforcement learning using Megatron-LM and SGLang.

Community
Advanced
tianhao909tianhao909
1

slime-rl-training

Integrates Megatron-LM and SGLang for scalable GRPO-based LLM reinforcement post-training.

Community
Advanced
Z43LZ43L

slime-rl-training

Integrates Megatron-LM training loops with accelerated SLang rollouts for LLM RL fine-tuning.

Community
Advanced
MesferAliMesferAli

slime-rl-training

Train GLM models with RL using Megatron-LM and SGLang.

Community
Advanced