// Set temporarily while in the Accumulate method; if backward_tape is not
    // nullptr then we forward op executions to it so Accumulate can compute a
    // backward pass on its backward function.
    //
    // Not owned by the ForwardAccumulator. The method which sets
    // `backward_tape` keeps ownership.
    GradientTape<Gradient, BackwardFunction, TapeTensor>* backward_tape;