训练速度慢

您好，使用到辅助信息时，模型的训练速度很慢，不知道是不是因为辅助信息是放在cpu上存储和操作的，导致速度慢的？比如使用ShallowEncoder（我用的是稀疏特征，可能是get_sparse_feature等导致的速度慢；不用稀疏特征时速度很快）。
是否有改进的方法或者思路呢？谢谢！

Start by profiling training with auxiliary information through ShallowEncoder, comparing sparse and non-sparse features. Inspect get_sparse_feature and the CPU-side auxiliary-feature operations mentioned in the report to determine where time is spent. Done means the bottleneck is confirmed and a concrete optimization scope is agreed.

Written by the indexing model from the issue text.