// i.e., d(y_1 + y_2 + ...)/dx_1, d(y_1 + y_2 + ...)/dx_2...
//
// `dx` are used as initial gradients (which represent the symbolic partial
// derivatives of some loss function `L` w.r.t. `y`).
// `dx` must be nullptr or have size `ny`.
// If `dx` is nullptr, the implementation will use dx of `OnesLike` for all
// shapes in `y`.

  op_tape_.erase(op_it);
}

// Terminology:
//
//  - op: a possibly composite operation, which has an entry in the tape
//  - target: dy in dx/dy
//  - source: dx in dx/dy
//  - tensor: one of the many inputs or outputs of an operation
//
// Below here we do the gradient algorithm. It works as follows:
//