Each of the tokens generated by the draft model has to be verified by the parent/original model, but if this acceptance rate falls, then the speedup from speculative decoding would be eliminated. This acceptance rate, and more directly the speedup from draft models, is what "performance" refer s to in the article.
By the way, can it be slower than without speculative decoding in worst case then?
> can it be slower than without speculative decoding in worst case then?
In practice, for single-user (non-batched) inference with a working configuration, you pretty much always get some speedup. For non-coding tasks I've seen it be nearly a wash for some people, in which case you might want to avoid it due to the extra memory usage (you'd rather use that memory to run a bigger quant/model, even at a slightly lower speed).