Yeah important conceptually to remember MTP is kind of just more weights, but speculative decoding is the runtime algorithm that’s a significant add to whatever code is serving the model.