Why GLM on llama.cpp has no MTP?

I spent a while last summer reverse engineering GLM 4.5's MTP layer with Claude. I was able to poke around in these layers and see the model thinking ahead about concepts, and get some level of next token prediction - but the extra computation needed to validate the tokens meant that I needed to consistently predict 2 tokens ahead to beat baseline speeds - because the next pass would be predicting the next token anyway, so if you're going to do a forward pass for the validation step, you need to have an extra token to check for it to end up being faster. So, normal speculative decoding with a smaller base model more effective really. I considered setting up a system to be able to do speculative decoding with different architectures, but then I got distracted building a KV caching system, since that saves you \*minutes\* of processing the system prompt every time you boot up a local coding agent or switch mode, and GLM Air was already giving me me \~60tps on inference. Of course I'm not an ML expert so maybe I was also doing something wrong, but on the other hand, maybe that's why they've never bothered to enable MTP on release - for inference, it's much more architectural complexity, without consistent benefit. My feeling after playing around with it for a week is that the real strength of MTP is that it likely speeds up training being able to process a few tokens at a time, and the model is probably smarter because it has a richer internal representation per token. This might be why GLM 4 series models basically seem to perform as well as models almost twice their size? And again this is just me spitballing, I haven't thought about this for a while.