How to reproduce average acceptance length?

[garipovroma]

Hello,

Thank you for your work and contribution to the community.

I've been reproducing the numbers from the EAGLE-2 paper, particularly Table 1 for the L2-7B model. However, I haven't been able to find the code that calculates $\tau$ (average acceptance length). While I found that it's possible to compute acceptance length by modifying evaluation/gen_ea_answer_llama2chat.py and model/ea_model.py (as mentioned in this issue #146), I am concerned that this approach may introduce inaccuracies due to discrepancies with your approach to measure $\tau$. Could you please confirm whether this is how you measured $\tau$ and provide your code for measuring average acceptance length, along with a smaller example of how to run it?

Thank you!

[Liyuhui-12]

The average acceptance length can be calculated by dividing the total number of generated tokens by the number of steps, which can be obtained from the output files.