LMSYS：Blog（Chatbot Arena 团队）·· 13 小时前AI 评分34
SGLang RL 团队实现 INT4 QAT 端到端流程：约 1TB 模型单张 H200 完成 rollout
Blog Squeezing 1TB Model Rollout into a Single H200: INT4 QAT RL End-to-End Practice 💡 TL;DR: Inspired by the Kimi K2 team, the SGLang RL team successfully landed an INT4 Quantization-Aware Training (QAT) pipeline. By combining fake quantization during training with real quantizati... SGLang RL Team, InfiXAI Team, Ant Group Asystem & AQ Infra Team, slime Team, RadixArk Team January 26, 2026
AI 导读
SGLang RL 团队联合 InfiXAI、蚂蚁 Asystem & AQ Infra、slime、RadixArk 团队，在 slime 框架上落地了 INT4 QAT 端到端流程，训练用 fake quantization、推理用 W4A16，实现与 BF16 全精度相当的训练-推理一致性与稳定性。
来源：LMSYS：Blog（Chatbot Arena 团队） · lmsys.org