: forward_outputs_(f_outputs) {
    for (auto output : forward_outputs_) {
      if (output) {
        output->Ref();
      }
    }
  }

  Status Compute(AbstractContext* ctx,
                 absl::Span<AbstractTensorHandle* const> grad_outputs,
                 absl::Span<AbstractTensorHandle*> grad_inputs) override {
    AbstractTensorHandle* upstream_grad = grad_outputs[0];

    grad_inputs[0] = grad_outputs[0];
    if (grad_inputs[0]) {
      grad_inputs[0]->Ref();
    }
    return absl::OkStatus();
  }
};

// Computes:
//
// @tf.custom_gradient
// def f(input):
//   def grad(grads):
//     return grads[0]
//   return tf.exp(input), grad
// outputs = [f(inputs[0])]
Status ExpWithPassThroughGrad(AbstractContext* ctx,

#include "tensorflow/c/eager/abstract_context.h"

namespace tensorflow {
namespace gradients {
namespace {
class IdentityNGradientFunction : public GradientFunction {
 public:
  Status Compute(AbstractContext* ctx,
                 absl::Span<AbstractTensorHandle* const> grad_outputs,
                 absl::Span<AbstractTensorHandle*> grad_inputs) override {
    for (int i = 0; i < grad_outputs.size(); i++) {

                 TF_Message(first_bad_status.get()));
    return result;
  }
  // For each output of the original operation, pack the per-device
  // TensorHandles we've computed into a single parallel TensorHandle.
  std::vector<std::unique_ptr<ParallelTensor>> per_device_outputs;
  per_device_outputs.reserve(first_op_output_count);
  for (int i = 0; i < first_op_output_count; ++i) {

        "If the intent is to treat this op as non-differentiable consider "
        "using RegisterNotDifferentiable or "
        "NotDifferentiableGradientFunction.");
  }
  return gradient_function->Compute(ctx_, output_gradients, result);
}

Status TapeVSpace::BuildOnesLike(const TapeTensor& t,
                                 AbstractTensorHandle** result) const {
  AbstractOperationPtr op(ctx_->CreateOperation());

      control_output_names_vec.push_back(string(control_output_names[i]));
    }
  }

  // Compute body nodes.
  std::vector<const Node*> body_nodes;
  status->status = tensorflow::ComputeBodyNodes(
      fn_body, fn_name, num_opers, opers, input_nodes, &body_nodes);
  if (TF_GetCode(status) != TF_OK) return nullptr;

  // Compute body nodes.
  std::vector<const Node*> control_output_nodes;

        "Expected numeric or variant tensor, got dtype ", dtype);
  }
  return tensorflow::ops::Conj(ctx, input, output, name);
}

class AddGradientFunction : public GradientFunction {
 public:
  Status Compute(AbstractContext* ctx,
                 absl::Span<AbstractTensorHandle* const> grad_outputs,
                 absl::Span<AbstractTensorHandle*> grad_inputs) override {
    // TODO(b/161805092): Support broadcasting.

    }
    ++count;
  }

  return result;
}

static bool FolderExists(GCSFile* gcs_file, std::string dir,
                         TF_Status* status) {
  ExpiringLRUCache<GcsFileStat>::ComputeFunc compute_func =
      [gcs_file](const std::string& dir, GcsFileStat* stat, TF_Status* status) {
        auto children =
            GetChildrenBounded(gcs_file, dir, 1, true, true, status);

  buf->Unref();
  return tensorflow::wrap(tensorflow::TensorHandle::CreateLocalHandle(
      std::move(t), device, device, context));
}

// This function will block till the operation that produces `h` has
// completed. This is only valid on local TFE_TensorHandles. Returns the size in
// bytes of the memory pointed to by the device pointer returned above.
size_t TFE_TensorHandleDeviceMemorySize(TFE_TensorHandle* h,

class TestUnavailableErrorOp : public tensorflow::OpKernel {
 public:
  explicit TestUnavailableErrorOp(tensorflow::OpKernelConstruction* ctx)
      : tensorflow::OpKernel(ctx) {}
  void Compute(tensorflow::OpKernelContext* ctx) override {
    ctx->SetStatus(tensorflow::errors::Unavailable("Test error."));
  }
};
REGISTER_KERNEL_BUILDER(
    Name("TestNonCommUnavailable").Device(tensorflow::DEVICE_DEFAULT),