skills.rest
ihatesea69
ihatesea69COMMUNITY
Explore all skills by this author →
WebsiteGitHub

miles-rl-training

Train large-scale Mixture-of-Experts models with FP8/INT4 quantization-aware training.

3|Updated Feb 21, 2026
Search Tags:#model training#reinforcement learning#speculative decoding#fp8#enterprise ai#moe#int4
One-click install
npx skills add https://github.com/ihatesea69/HieuNghi-AI-Skills --skill miles-rl-training-ihatesea69
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: miles-rl-training
Source: https://github.com/ihatesea69/HieuNghi-AI-Skills/tree/main/airesearch_skills/06-post-training/miles
Command: npx skills add https://github.com/ihatesea69/HieuNghi-AI-Skills --skill miles-rl-training-ihatesea69

SYSTEM DOCUMENTATION & REQUIREMENTS

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

What problem does it solve?

This Skill streamlines the complex process of training large-scale Mixture-of-Experts (MoE) models, addressing challenges like stability, low-precision training, and train-inference alignment in enterprise environments.

Core Features & Use Cases

  • Large MoE Training: Optimized for 1TB+ MoE models (e.g., DeepSeek V3, Qwen3-MoE).
  • Low-Precision Training: Supports FP8 and INT4 quantization-aware training for reduced memory footprint.
  • Train-Inference Alignment: Ensures bit-wise identical alignment between training and inference.
  • Speculative RL: Achieves maximum throughput via speculative decoding.
  • Use Case: Train a 1TB Qwen3-MoE model on H200 GPUs using FP8 quantization, ensuring that the trained model behaves identically during inference as it did during training.

Quick Start

Use the miles skill to train a Qwen3-MoE model using FP8 quantization and speculative RL.

Frequently Asked Questions about miles-rl-training

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

FAQPage Schema
How do I train large-scale MoE models with FP8 quantization?▼

You can train large-scale MoE models using FP8 quantization-aware training to reduce memory footprint. This approach supports 1TB+ models like DeepSeek V3 and Qwen3-MoE on H200 GPUs while maintaining stability.

What is bit-wise train-inference alignment in reinforcement learning?▼

Bit-wise train-inference alignment ensures a trained model behaves identically during inference as it did during training. This process eliminates discrepancies caused by low-precision quantization when deploying large-scale MoE models.

Can I use speculative decoding to improve RL training throughput?▼

Yes, you can use speculative RL to achieve maximum throughput during training. Speculative decoding accelerates the reinforcement learning process for large-scale Mixture-of-Experts models.

Does this Skill support INT4 quantization-aware training for Mixture-of-Experts models?▼

Yes, it supports INT4 quantization-aware training alongside FP8. This capability further reduces the memory footprint required for training large-scale Mixture-of-Experts models in enterprise environments.

How do I integrate distributed training with SGLang and Megatron-LM?▼

Distributed training integrates with SGLang and Megatron-LM to facilitate enterprise-grade reinforcement learning. This combination supports advanced features like speculative RL and bit-wise alignment for massive models.

What is the best way to ensure stability when training 1TB MoE models?▼

To ensure stability when training 1TB MoE models, use quantization-aware training with bit-wise train-inference alignment. This method addresses low-precision challenges and maintains consistent behavior during inference.

Related Skills

Explore More Software Engineering Skills→
SalesforceCommerceCloudSalesforceCommerceCloud
51

b2c-ods

Provision, manage, and monitor on-demand B2C sandboxes using the b2c CLI.

Official
Intermediate
hhhi21ghhhi21g

miles-rl-training

Optimize reinforcement learning training for large MoE models with low-precision and speculative RL.

Community
Advanced
webdevtodayjasonwebdevtodayjason

fine-tuning-with-trl

Fine-tune language models with TRL for RLHF and preference alignment.

Community
Advanced
BagelHoleBagelHole
46

llm-fine-tuning

Set up QLoRA, LoRA, and full fine-tuning infrastructure with Hugging Face TRL and Axolotl.

Community
Advanced