speculative decoding in llama.cpp

[mellowwire98]

I don't think that's the whole story. I wouldn't optimize around that first. On my side the bigger swing came from the runtime and not the model.