Why does training eagle with my own data perform worse than medusa

[skylee-01]

Hello, our data is for an agent scenario, with 100 tokens as input and 20 tokens as output. The input is relatively fixed. The acceleration effect of using Eagle's code for training is weaker than that of Medusa. How should we troubleshoot this? I understand that Eagle's performance should definitely be better than Medusa's. What data should I collect for analysis?

[skylee-01]

Using the original eagle version

[xinlong-yang]

May be the output is too short (...