skills.rest
uv-xiao
uv-xiaoCOMMUNITY
Explore all skills by this author →
Website𝕏@youwei_xiaoGitHub

uv-miles-rl-training

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

1|Updated Feb 6, 2026
Search Tags:#reinforcement learning#speculative decoding#fp8#moe#int4#enterprise training
One-click install
npx skills add https://github.com/uv-xiao/pkbllm --skill uv-miles-rl-training
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: uv-miles-rl-training
Source: https://github.com/uv-xiao/pkbllm/tree/main/knowledge/ML/training/miles
Command: npx skills add https://github.com/uv-xiao/pkbllm --skill uv-miles-rl-training

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?

This Skill provides a robust framework for training large-scale Mixture-of-Experts (MoE) models, addressing challenges like training stability, low-precision quantization, and train-inference alignment in enterprise environments.

Core Features & Use Cases

  • Large MoE Training: Optimized for training models over 1TB, supporting DeepSeek V3 and Qwen3-MoE.
  • Low-Precision Training: Enables FP8 and INT4 quantization-aware training for reduced memory footprint and increased throughput.
  • Train-Inference Alignment: Ensures bit-wise identical alignment between training and inference using techniques like Rollout Routing Replay (R3).
  • Speculative RL: Achieves up to 25%+ rollout speedup through speculative decoding.
  • Use Case: Train a 1TB MoE model using FP8 quantization on H100 GPUs, ensuring bit-wise alignment with inference and maximizing throughput via speculative RL.

Quick Start

Use the uv-miles-rl-training skill to train a Qwen3-30B model with FP8 quantization and speculative RL enabled.

Frequently Asked Questions about uv-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 with FP8 quantization using a framework that supports FP8 quantization-aware training. This reduces memory footprint and increases throughput for models exceeding 1TB, such as DeepSeek V3 and Qwen3-MoE.

What is train-inference alignment for Mixture-of-Experts models?▼

Train-inference alignment for Mixture-of-Experts models ensures bit-wise identical states between training and inference. Techniques like Rollout Routing Replay (R3) achieve this alignment to maintain stability in enterprise production environments.

Does speculative decoding speed up reinforcement learning rollouts?▼

Yes, speculative decoding speeds up reinforcement learning rollouts. By applying speculative RL techniques during the rollout phase, you can achieve a 25% or greater increase in rollout speedup for large-scale training.

Can I use INT4 quantization-aware training for 1TB MoE models?▼

Yes, you can use INT4 quantization-aware training for 1TB MoE models. This low-precision training capability reduces memory footprint and increases throughput while maintaining stability for large-scale enterprise models.

What's the best way to achieve enterprise-grade RL training stability for MoE?▼

The best way to achieve enterprise-grade RL training stability for MoE is by using a framework that combines low-precision quantization-aware training with Rollout Routing Replay. This ensures train-inference alignment for models over 1TB.

Related Skills

Explore More Software Engineering Skills→
zhuangbiaoweizhuangbiaowei
2

slime-rl-training

Trains large language models with GRPO reinforcement learning using Megatron-LM and SGLang.

Community
Advanced
ccfccf

vector-index-tuning

Tune HNSW parameters and quantization for large-scale vector search indexes.

Community
Advanced
ovachieverovachiever
52

grpo-rl-training

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

Community
Advanced
Orchestra-ResearchOrchestra-Research
11.5k

miles-rl-training

Train large-scale MoE models with FP8 and INT4 low-precision RL.

Official
Advanced