SafeAILab/EAGLE

Question regarding sampling technique in EAGLE papers

Closed this issue · 2 comments

llsj14 commented

Hello,

Thank you for your excellent work on EAGLE and EAGLE-2. I have been reading both papers, as well as Medusa.

The accept rate and speed-up metrics from the various experiments in the paper are valuable.
However, I have a question regarding the sampling method you used. I’m curious whether you applied rejection sampling or the typical acceptance scheme proposed by Medusa.

The papers typically show a 10-15% difference in accept rate, aside from the sampling method. It would be helpful to understand which method you used in your experiments.

Thank you

Liyuhui-12 commented

Thank you for your interest. We employ strict speculative sampling to ensure lossless results, rather than using the typical acceptance scheme found in Medusa.

  • 👍1
llsj14 commented

Ah, so it is rejection sampling. Thank you for your answer.