How to use a linear layer so the weights are softmax normalized

I want my linear layer to have a positive interaction with previous layer inputs
So whether to create a new layer and use weights that require grad and init them with new layer so they can be included in computation graph OR how?