inference based DeepSpeed implementation for ensemble feature.

Input

3871-dim embedding

Encoder

88 x CNN with 30 heads

Output

auc-roc projection