CoreWeave Forge 将投机解码优化流程整合到单一环境
精选理由
CoreWeave 讲了个真事：优化 Kimi K2.6 推理速度靠工程师熬夜手动切换训练和评估，现在 Forge 把这套投机解码流程自动化了。
CoreWeave 介绍了 Kimi K2.6 推理优化的案例：投机解码由小模型起草 token、大模型负责验证，再由工程师 @aaronbatilo 在训练、评估、推理服务之间手动流转每一轮迭代。该方案曾让 Kimi K2.6 登上 Artificial Analysis 的输出速度榜首。CoreWeave Forge 把这一循环固化成可重复的流程，全部步骤在同一环境中完成，不再依赖人工交接。
原文 · CoreWeave
Speculative decoding has a small model draft tokens and a big one verifies them. Ours also had @aaronbatilo, who manually moved the work between training, evaluation, and serving every round.
It topped @ArtificialAnlys for Kimi K2.6 output speed at the time, and all that hand work cost Aaron a lot of sleep.
CoreWeave Forge makes that loop repeatable, every step in one connected environment, so the handoffs don't depend on Aaron staying up.
Try it here: https://t.co/yQOKK1IqyX