miles-rl-training - Enterprise RL Training Framework
Enterprise-grade RL training framework for large MoE models with FP8/INT4 support
Tags
Updated: 2026-05-11Capabilities
Typical Inputs
Typical Outputs
What this skill does
- train MoE models
- configure FP8 training
- configure INT4 training
- enable speculative RL
- record expert routing
- replay routing decisions
Inputs
- model checkpoint
- prompt dataset
- draft model checkpoint
- training configuration
Outputs
- trained model checkpoint
- training logs
- loss metrics
Requirements
- H100/H200 GPUs with FP8 support
- Docker environment
- CUDA
- sglang-router
- Ray
- PyTorch
- transformers