Enterprise RL-обучение крупных MoE-моделей

★ 7.7 · ai-tooling

miles is a Claude Code skill that guides enterprise-grade reinforcement learning training using miles, a production-ready fork of slime built by Orchestra Research for large-scale model post-training. It covers training 1TB+ MoE models such as DeepSeek V3 and Qwen3-MoE with FP8/INT4 quantization, bit-wise identical train-inference alignment via Rollout Routing Replay (R3), and speculative RL using EAGLE decoding through SGLang for 25–40% rollout speedup. The skill includes two structured workflows: distributed MoE training with Megatron-LM tensor and expert parallelism, and speculative RL setup with optional online MTP draft-model training. Dependencies are sglang-router ≥ 0.2.3, ray, torch ≥ 2.0.0, and transformers ≥ 4.40.0. It is aimed at teams that need production stability when fine-tuning large language models with reinforcement learning on H100/H200 clusters.

• #reinforcement-learning
• #moe-training
• #quantization
• #llm-training
• #sglang
• #megatron-lm