How can the pruned model with sparse matrix save model size and computation cost?

[JiachuanDENG]

Thank you so much for sharing this work and code!
But from my understanding, Wanda is actually not decreasing the number of parameter of the model, instead, it simply set a lot of parameter to 0.

1. In this way, the total number of parameter remains the same as the original model. So the model size will be the same. I checked this by print out the #param of the pruned model and original model, the #param looks identical
   

2. Although a lot of parameters are set to 0, but if no special optimization is designed for the forward pass, 0 is still involved in the multiplicaiton and addition operation. So the computation cost will be the same.
   Am I understanding right?

[Eric-mingjie]

Yes, you are right. This is how most researchers implement unstructured pruning methods (you can think of it as similar to how people implement magnitude pruning in practice). Structured N:M sparsity however can deliver practical speedup.