skills.rest
ricable
ricableCOMMUNITY
Explore all skills by this author →
GitHub

miles-rl-training

Configure and run RL training for large language models with MoE, FP8, and INT4 workflows.

Updated Feb 15, 2026
Search Tags:#reinforcement learning#speculative decoding#fp8#sglang#int4#train-inference alignment#mmoe training
One-click install
npx skills add https://github.com/ricable/mcai --skill miles-rl-training-ricable
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: miles-rl-training
Source: https://github.com/ricable/mcai/tree/main/.agents/skills/miles-rl-training
Command: npx skills add https://github.com/ricable/mcai --skill miles-rl-training-ricable

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

This Skill helps teams run stable, high-throughput reinforcement learning training for large language models, especially when standard setups fail on scale, precision, or train-inference alignment.

Core Features & Use Cases

  • Large-scale MoE training: Coordinate production RL workflows for DeepSeek, Qwen MoE, and other large expert models.
  • Low-precision optimization: Apply FP8 and INT4 training strategies to reduce memory pressure while preserving throughput.
  • Speculative and aligned rollouts: Improve rollout speed and keep training behavior aligned with inference using replay, correction, and draft-model workflows.
  • Use case: A research or infrastructure team can use this Skill to configure an end-to-end RL run with parallelism settings, routing replay, and safety checks for stable production training.

Quick Start

Ask the assistant to configure a miles RL training run for your target model, cluster size, precision mode, and rollout strategy.

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 reinforcement learning training for large MoE models like DeepSeek?▼

To configure reinforcement learning training for large MoE models, specify your target model, cluster size, precision mode, and rollout strategy. The setup coordinates production RL workflows using Megatron parallelism settings and SGLang for stable high-throughput execution across distributed nodes.

What is train-inference alignment in production RL systems?▼

Train-inference alignment in production RL systems ensures training behavior matches inference using replay, correction, and draft-model workflows. This alignment is maintained through speculative rollouts and routing replay, preventing policy drift during large-scale language model training.

Can I use FP8 and INT4 quantization-aware workflows for large language model training?▼

Yes, you can apply FP8 and INT4 training strategies to reduce memory pressure while preserving throughput in large language model training. These low-precision optimization workflows are supported natively, allowing stable execution without sacrificing performance during production RL runs.

Do I need miles-compatible cluster orchestration to run SGLang and Megatron parallelism?▼

Yes, miles-compatible cluster orchestration is required to manage SGLang and Megatron parallelism settings effectively. This orchestration coordinates optional R3, TIS, and MTP controls necessary for maintaining stability and high throughput during enterprise-grade RL training execution.

How do speculative rollouts improve reinforcement learning speed for large models?▼

Speculative rollouts improve reinforcement learning speed by using draft-model workflows to predict and verify multiple tokens simultaneously. This technique reduces generation latency while maintaining train-inference alignment through correction mechanisms during the rollout process.

Why does my large-scale MoE training fail on precision or train-inference alignment?▼

Large-scale MoE training fails on precision or alignment when standard setups lack quantization-aware workflows and replay mechanisms. Applying FP8 or INT4 strategies with routing replay and speculative rollouts resolves memory pressure and corrects inference drift.

Related Skills

Explore More Software Engineering Skills→
KappTech88KappTech88

miles-rl-training

Guide RL training of large MoE models with Miles using FP8/INT4 optimization.

Community
Advanced
OpenLAIROpenLAIR
1.0k

miles-rl-training

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

Official
Advanced
ovachieverovachiever
52

grpo-rl-training

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

Community
Advanced
Supporter09Supporter09

miles-rl-training

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

Community
Advanced