How does speculative decoding contribute to fast inference from transformers?
A) By reducing the number of layers in the transformer
B) By parallelizing the decoding process
C) By increasing the number of attention heads
D) By using beam search to generate multiple candidate outputs