0.1.2

generalize to any number of leading start tokens for prophet

0.1.1

the extra norm in projecting to prophet model dimensions hurt for som…

…e reason

0.1.0

complete speculative sampling with prophet net on cached embeddings i…

…dea!

0.0.15

fix start token and determining input sequence

0.0.14

move towards completing what is started

0.0.12

more fixes

0.0.11

just copy batched spec decoding and make batch early exit strategy work

0.0.10

0.0.10

0.0.9

make batched spec decoding work

0.0.8

necessary setup for batched speculative decoding