// in tests only.)
TF_CAPI_EXPORT int TF_PickUnusedPortOrDie(void);

// Fast path method that makes constructing a single scalar tensor require less
// overhead and copies.
TF_CAPI_EXPORT extern TFE_TensorHandle* TFE_NewTensorHandleFromScalar(
    TF_DataType data_type, void* data, size_t len, TF_Status* status);

// Specify the server_def that enables collective ops.

      TF_NewStatus(), TF_DeleteStatus);
  TFE_Context* eager_ctx =
      TF_ExecutionContextGetTFEContext(wrap(ctx), status.get());
  TF_RETURN_IF_ERROR(StatusFromTF_Status(status.get()));
  TFE_TensorHandle* input_eager =
      TestTensorHandleWithDims<T, datatype>(eager_ctx, data, dims, num_dims);
  *tensor =
      unwrap(TF_CreateAbstractTensorFromEagerTensor(input_eager, status.get()));
  return absl::OkStatus();
}

  ~AbstractTensorHandle() override {}

 public:
  // Returns tensor dtype.
  virtual tensorflow::DataType DataType() const = 0;

  // Returns the status of the tensor handle. If it is a tfrt::TensorHandle,
  // the tensor handle can be an error and return non-OK status.
  virtual tensorflow::Status TensorHandleStatus() const;

  // Returns tensor shape. If tensor has unknown rank, shape remains untouched.

  virtual ImmediateExecutionTensorHandle* CreateLocalHandleFromTFTensor(
      tensorflow::Tensor& t, const char* d_name) = 0;

  // Convert a TFRT TensorHandle to tensorflow::TensorHandle.
  virtual ImmediateExecutionTensorHandle* TFTensorHandleFromInterface(
      ImmediateExecutionTensorHandle* handle) = 0;

  virtual std::vector<std::string> GetLoggedOpsTestonly() { return {}; }

// (also known as adjoint function) to compute, given downstream gradients,
// upstream gradients.
//
// TODO(apassos) provide concrete template instantiations for TFE_TensorHandle
// specialization, which is blocked by quite a few things needing to loop back
// into python now.
template <typename Gradient, typename BackwardFunction, typename TapeTensor>
class VSpace {
 public: