Why the bias of config in train scripts doesn't aligning at infererence?

[Haodong-Lei-Ray]

When training, I found the bias doesn't align with the config. This file is in the EAGLE3/eagle/train/main.py. I think if do that, the result will produce the bias in "fc" layer due to following the default. Maybe you should edit it because I find the model_state_dict of your official model doesn't have the fc.bias.

[Haodong-Lei-Ray]

I raise this question because I find Model follow this json.