Train speculative decoding models effortlessly and port them smoothly to SGLang serving. - View it on GitHub