miles-rl-training

Optimize large-scale MoE reinforcement learning training with Megatron parallelism and FP8 or INT4 quantization.

Updated May 4, 2026
One-click install
npx skills add https://github.com/Supporter09/Face_Anti_Spoofing_Biometric --skill miles-rl-training-supporter09
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: miles-rl-training
Source: https://github.com/Supporter09/Face_Anti_Spoofing_Biometric/tree/main/.claude/skills/miles
Command: npx skills add https://github.com/Supporter09/Face_Anti_Spoofing_Biometric --skill miles-rl-training-supporter09

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

This Skill helps teams run large-scale reinforcement learning training for modern MoE language models without getting lost in brittle configuration, unstable low-precision setups, or train-inference mismatch.

Core Features & Use Cases

  • Enterprise MoE Training: Plan and configure high-throughput training for large expert models with Megatron-style parallelism.
  • Low-Precision Workflows: Apply FP8 and INT4 strategies when memory pressure and throughput are critical.
  • Train-Inference Alignment: Use routing replay, deterministic settings, and validation checks to keep rollout behavior consistent with training.
  • Speculative RL: Speed up rollout generation with draft-model speculation and online updates.
  • Troubleshooting Support: Diagnose collapse, routing mismatch, quantization regressions, and synchronization issues with practical recovery steps.

Quick Start

Ask the Skill to create a miles training plan for your MoE model, including alignment, precision settings, rollout acceleration, and the safest configuration choices for your hardware.

Frequently Asked Questions about miles-rl-training

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

FAQPage Schema
How do I configure large-scale reinforcement learning training for MoE language models?

Configure large-scale RL training for MoE models by applying Megatron-style parallelism, routing replay, and train-inference alignment to stabilize high-throughput enterprise post-training workflows.

What is the best way to maintain train-inference alignment during SGLang rollout integration?

Maintain train-inference alignment during SGLang rollout integration by using deterministic expert-routing support, routing replay, and validation checks to keep rollout behavior consistent with training.

Can I use FP8 or INT4 quantization for MoE reinforcement learning training?

You can apply FP8 and INT4 quantization strategies to MoE reinforcement learning training when memory pressure and throughput are critical for your enterprise post-training setup.

How does speculative decoding speed up rollout generation in RL training?

Speculative decoding speeds up rollout generation in RL training by using draft-model speculation and online updates to accelerate generation without compromising training stability.

Why does expert routing collapse or mismatch during low-precision MoE training?

Expert routing collapse or mismatch during low-precision MoE training occurs from quantization regressions and synchronization issues, requiring deterministic settings and practical recovery steps to diagnose.

Do I need validated checkpoint conversion for Megatron parallelism in RL training?

Validated checkpoint conversion is required for Megatron parallelism in RL training to ensure miles-compatible configuration and deterministic expert-routing support across distributed setups.