Created by: jczaja

MKL-DNN is often keeping params (weights, bias) in its own blocked format. Essentaily it means that each operators that is having params is keeping two sets of weights (one paddle Tensors and then converted(reordered) weights into MKL-DNN format). This may lead to increased memory cosnumption. Goal of this is task is to analyze if it is possible to keep only one set of weights&biases not two. For example if it is possible to release paddle weights in inference and keep only mkl-dnn weights.