End-to-End Speedup in Figure 2 if Draft Token Acceptance Rate = 100%

[longbowzhang]

Hi Authors,

Your remarkable work is both inspiring and thought-provoking. As I delve deeper into the details, I have a question about the relationship between "Draft Token Acceptance Rate" and "End-to-End Speedup." Specifically, in Figure 2, if the "Draft Token Acceptance Rate" were 100%, how would the End-to-End Speedup values change for K={1,2,4}?

I assume these values can be analytically calculated. Could you confirm this, and if so, provide insights into how these calculations might be derived? Any guidance or clarification would be greatly appreciated.

[junzhang-zj]

Hi @longbowzhang, this is a toy example to visualize the theoretical impact of draft token number and acceptance rate on speedup. In this scenario, the draft model is assumed to be 2× faster than the verify model. Please note that this setup is purely illustrative and does not reflect actual experimental conditions.