miles-rl-training

Configure miles for large MoE RL training with FP8/INT4 and train-inference alignment.

1.0k|117|Updated Feb 26, 2026
One-click install
npx skills add https://github.com/OpenLAIR/dr-claw --skill miles-rl-training-openlair
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: miles-rl-training
Source: https://github.com/OpenLAIR/dr-claw/tree/main/skills/post-training/miles
Command: npx skills add https://github.com/OpenLAIR/dr-claw --skill miles-rl-training-openlair

SYSTEM DOCUMENTATION & REQUIREMENTS

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

What problem does it solve?

Enterprise-grade RL training for large MoE models requires specialized tooling, stability, and production-oriented configurations. miles provides a production-ready fork of slime to address MoE training stability, FP8/INT4 support, and train-inference alignment.

Core Features & Use Cases

  • Enterprise RL workflows: scalable training for large MoEs, FP8/INT4 support, speculative RL, online SFT.
  • Production readiness: tooling, deployment, and stability for enterprise environments.
  • Use Case: Accelerate rollout of MoE models in a production research lab with end-to-end training and evaluation.

Quick Start

Begin enterprise RL training with miles by configuring a large MoE model for FP8 training and enabling train-inference alignment.

Frequently Asked Questions about miles-rl-training

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

FAQPage Schema
How do I stabilize RL training for large MoE models in a production environment?

Stabilize RL training for large MoE models by deploying production-ready configurations that support FP8/INT4 and train-inference alignment. This addresses training instability through specialized tooling and optimized deployment options for enterprise environments.

What is train-inference alignment and when do I need it for MoE training?

Train-inference alignment ensures consistency between training and inference phases in large MoE models. You need it during speculative RL and online SFT to prevent deployment discrepancies and maintain production readiness in research labs.

Can I use sgLang and Megatron-style parallelism for enterprise RL training?

Yes, enterprise RL training integrates with sgLang and supports Megatron-style parallelism. This enables scalable training for large MoEs with optimized production configurations and train-inference alignment.

How do I configure FP8 or INT4 support for large MoE model training?

Configure FP8 or INT4 support for large MoE models by applying specialized training configurations during deployment. This optimizes memory usage and accelerates rollout in production research lab environments.

Does Ray work with torch and transformers for speculative RL workflows?

Yes, Ray works with torch and transformers to execute speculative RL workflows. These dependencies support distributed execution and model handling for enterprise-grade large MoE training and online SFT.

What are the limitations of using slime-like configurations for enterprise MoE deployment?

Slime-like configurations may require specialized forks to address MoE training stability and production readiness. Without train-inference alignment and FP8/INT4 support, enterprise deployment may face scalability and consistency limitations.