Speculative decoding accelerates autoregressive generation by letting a faster draft model propose multiple tokens that a target model verifies in parallel without changing the target distribution.
Some results have been hidden because they may be inaccessible to you
Show inaccessible results