TF_DeleteTensor(model_out_tensor);

  // If the output is a scalar, then return the scalar output
  if (num_dims_out == 0) {
    outputs[0] = model_out.release();
    return absl::OkStatus();
  }

  // Else, reduce sum the output to get a scalar

  // Will sum all dimensions, so get a Tensor containing [0,...,num_dims_out-1].
  AbstractTensorHandlePtr sum_dims;
  {
    vector<int32_t> vals(num_dims_out);

        // Non-parallel tensors from _EagerConst/tf.constant are implicitly
        // broadcast, i.e. set as the input to each parallel operation. This
        // allows code like "tf.constant(1.)" or "tf.reduce_sum(..., axis=1)"
        // (where the value starts on the host), without allowing other implicit
        // copies/broadcasts. Other implicit copies may be supported eventually,