When modifying the model, there was a severe drop in GFLOPS

[w13ww]

I want to modify the Conv in the code to kervolution, but a strange phenomenon appeared after I completed the modification。
Here are the initial model parameters and GFLOPS.

This is after modification。

And before and after the modification, the use of gpu during training has also undergone major changes. Before modification, each round of training gpu uses 4.6G, after modification, the gpu is 36G！