LMSYS：Blog（Chatbot Arena 团队）·· 13 小时前AI 评分42
Miles 实现 Blackwell 原生 MXFP8 与 NVFP4 强化学习方案
Blog Towards Blackwell-Native 8-bit and 4-bit RL: End-to-End MXFP8 and NVFP4 RL in Miles TL;DR: We implemented two Blackwell-native RL recipes in Miles: end-to-end MXFP8 and per-token NVFP4 for MoE experts. Both are supported by fine-grained precision control across checkpoint conversion,... Ziang Li, humans& and Miles Team July 29, 2026
AI 导读
Miles 团队在 Miles 中实现了两种 Blackwell 原生 RL 方案：端到端 MXFP8，以及面向 MoE 专家的 per-token NVFP4，覆盖 checkpoint 转换、Megatron 训练、SGLang rollout 与实时权重更新。
来源：LMSYS：Blog（Chatbot Arena 团队） · lmsys.org