Keras gradient, must be calculated wrt weights or wrt inputs?

Question

I\'m building with Keras (2.4.3 with Tensorflow 2.3.0) a custom layer implementing the function f_theta(x) whose x is a single real-valued inpu