Saving Computational Time with DNN Pruning

I am following this tutorial on pruning.
It seems particularly attractive to me as it is said that if we reduce the number of parameters of the model we will reduce the execution time, which is relevant in my application.

However, from this example that I applied to my case, it just seems to me that a certain percentage of parameters are zeroed, but actually, they remain there in the model with a value of 0.

They are not removed, so it seems to me that the execution time remains the same.

What am I doing wrong? How to do it?