The type mlpackage(mlprogram)'s inferencing time and load time are much more than mlmodel(Neural network).

As to your official documents, the mlprogram's load time should be less than mlmodel. But now the test result is contrary.

No source file, test, or reproduction details are named. Start by reproducing the load and inference timing comparison between mlpackage(mlprogram) and mlmodel (Neural network), then compare the results with the official documentation; done means the discrepancy and whether it reflects an implementation or documentation problem are established.

Written by the indexing model from the issue text.