std::vector<Tensor> all_input_tensors;
  // Update the vector with information from `input_tensors` if provided.
  if (input_tensors != nullptr) {
    // Note that we take the address of the elements in `all_input_tensors`
    // below. Allocate enough space so that no reallocation happens, which will
    // make the pointers invalid.
    all_input_tensors.reserve(num_inputs);

// Tensor passed to Accumulate goes out of scope.
//
// Not thread-safe.
template <typename Gradient, typename BackwardFunction, typename TapeTensor>
class ForwardAccumulator {
 public:
  // Does not take ownership of `vspace`, which must outlive the
  // ForwardAccumulator.
  explicit ForwardAccumulator(
      const VSpace<Gradient, BackwardFunction, TapeTensor>& vspace,
      bool use_batch)

build:rbe_cross_compile_macos_x86 --bes_upload_mode=nowait_for_upload_complete
test:rbe_cross_compile_macos_x86 --config=rbe_cross_compile_base
# Increase the test timeout as tests often take longer on mac.
test:rbe_cross_compile_macos_x86 --test_timeout=300,450,1200,3600
# Limit jobs to 100 to avoid running into "out of memory" issues (b/316266643)
build:rbe_cross_compile_macos_x86 --jobs=100

    theta_inputs[input_index] = thetaMinus.get();
    TF_RETURN_IF_ERROR(
        RunAndMaybeSum(ctx, forward, theta_inputs, f_outputs, use_function));
    AbstractTensorHandlePtr fMinus(f_outputs[0]);

    // Take Difference of both estimates: (f(theta + eps) - f(theta - eps)).
    TF_RETURN_IF_ERROR(
        ops::Sub(ctx, fPlus.get(), fMinus.get(), f_outputs, "sub_top"));
    AbstractTensorHandlePtr fDiff(f_outputs[0]);

//   TF_Output values_inputs[5] = {{...}, ..., {...}};
//   TF_AddInputList(desc, values_inputs, 5);

// For inputs that take a single tensor.
TF_CAPI_EXPORT extern void TF_AddInput(TF_OperationDescription* desc,
                                       TF_Output input);

// For inputs that take a list of tensors.
// inputs must point to TF_Output[num_inputs].
TF_CAPI_EXPORT extern void TF_AddInputList(TF_OperationDescription* desc,