PyTorch - Sharing gradients between different iterations doesn't improve performance

Question

I need to find the largest eigenvalue of the NN, so I\'m implementing the power method. The main function is a hessian-vector product (HVP), based on PyHessian.

The m