merging tokenizer of deepseekv3 to distilled model for speculative decoding in llama.cpp?

[BarfingLemurs]

Hi @cg123, would there be a way to swap the original distilled deepseek model's tokenizers to make them somewhat compatible with llama.cpp speculative decoding?

Another case, could deepseekvl2, which has the same vocabulary as deepseekv3 be converted to a text-only model, for speculative use?

Until llama.cpp supports MTP inference, this would be very useful for that model!