miles-rl-training
Category: ml-training · License: MIT · Version: 1.0.0
Tags: Reinforcement Learning, MoE, FP8, INT4, Enterprise, SGLang, Megatron-LM
When to use it
Provides guidance for enterprise-grade RL training using miles, a production-ready fork of slime. Use when training large MoE models with FP8/INT4, needing train-inference alignment, or requiring speculative RL for maximum throughput.Example requests
Ask your agent in plain language, for example:- “How do I turn on FP8 rollout and training for a Qwen3-MoE run so RL doesn’t collapse?”
- “I’m running GRPO on DeepSeek-V3 with miles on 4 nodes of H200. After about 200 steps the reward drops sharply and the logprobs from SGLang don’t match Megatron’s. Is this a train-inference mismatch, and how do I fix it? My args currently include —use-tis —tis-threshold 0.9 but I don’t use any routing replay.”
- “My RL rollouts are the bottleneck when post-training a big MoE model. Generation takes far longer than the gradient update. I’d like to speed up sampling by having a small helper model propose tokens that the main model checks, and the helper should stay in sync with the policy as it changes. What setup do I need?”
Install
/plugin marketplace add KalarisLabs/research-agent-skills and install the whole ml-training category with /plugin install ml-training@research-agent-skills.