e8a945606e0b3dafe71ce2f1bf1b52ab3b05428c,gpytorch/lazy/kronecker_product_added_diag_lazy_tensor.py,KroneckerProductAddedDiagLazyTensor,_root_inv_decomposition,#KroneckerProductAddedDiagLazyTensor#Any#,217

Before Change

                symm_prod = KroneckerProductLazyTensor(
                    *[d.matmul(k).matmul(d) for k, d in zip(lt.lazy_tensors, dlt_inv_root.lazy_tensors)]
                )
                evals, evecs = symm_prod.symeig(eigenvectors = True)
                evals_p_i_root = DiagLazyTensor((evals + 1.).reciprocal().sqrt())
                return MatmulLazyTensor(dlt_inv_root, MatmulLazyTensor(evecs, evals_p_i_root))
        
        return super()._root_inv_decomposition(initial_vectors=initial_vectors)

After Change

                return MatmulLazyTensor(S, Lambda_I.inverse().sqrt())
            else:
                // again, we compute the root decomposition by pulling across the diagonals
                dlt_sqrt, evals_p_i, evecs = _symmetrize_kpadlt_constructor(lt, dlt)
                dlt_inv_root = dlt_sqrt.inverse()
                evals_p_i_root = DiagLazyTensor(evals_p_i.diag().reciprocal().sqrt())
                return MatmulLazyTensor(dlt_inv_root, MatmulLazyTensor(evecs, evals_p_i_root))

Instances