轻松训练推测性解码模型，并顺利将其迁移到 SGLang 服务中。
Train speculative decoding models effortlessly and port them smoothly to SGLang serving.
