LinkedIn 用多教师蒸馏将 AI 职位搜索训练提速 8 倍
原标题：How LinkedIn Trains AI Job Search 8x Faster with Multi-Teacher Distillation
AI 摘要
LinkedIn 公布了其 AI 职位搜索背后的训练基础设施，采用多教师蒸馏流水线，将 8B 相关性教师模型和 1.7B 互动教师模型的知识压缩进 0.6B 参数的排序模型。系统基于开源 SGLang 构建自定义框架，在训练循环中直接服务教师模型，并通过在线/离线蒸馏拆分、LiGer、多节点训练、FSDP2 和 H200 集群等优化实现约 8 倍训练加速。该 0.6B 学生模型使职位搜索 NDCG@10 从 0.7583 提升至 0.9432（+24.48%），推理吞吐从约 290 提升至每 GPU 每秒 2000 条以上，目前已在美国用户的自然语言职位搜索中投产。
正文节选
LinkedIn has published details of the training infrastructure behind its AI-powered job search, describing a multi-teacher distillation pipeline that compresses knowledge from large teacher models into a compact 0.6B-parameter ranking model. The main contribution isn’t just the distillation technique. It’s the systems work that makes it fast enough for iteration. This includes a custom framework built on SGLang, which serves teacher models directly in the training loop. Training a small language