TF_Status* status = TF_NewStatus();
  TF_Tensor* t = TFE_AllocateHostTensor(ctx, datatype, dims, num_dims, status);
  memcpy(TF_TensorData(t), data, TF_TensorByteSize(t));
  TFE_TensorHandle* th = TFE_NewTensorHandleFromTensor(ctx, t, status);
  CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TF_DeleteTensor(t);
  TF_DeleteStatus(status);
  return th;
}

#ifndef TENSORFLOW_C_EAGER_ABSTRACT_FUNCTION_H_
#define TENSORFLOW_C_EAGER_ABSTRACT_FUNCTION_H_

#include "absl/status/statusor.h"
#include "tensorflow/core/framework/function.pb.h"
#include "tensorflow/core/platform/intrusive_ptr.h"
#include "tensorflow/core/platform/refcount.h"
#include "tensorflow/core/platform/status.h"

namespace tensorflow {

class FunctionRecord;

  // When `reset_context` is true, initialize new cluster context state based
  // on cluster configurations provided in `server_def`; otherwise, update
  // existing context state with the provided `server_def`. Contexts created
  // on remote tasks will be considered stale and garbage collected after
  // `keep_alive_secs` of inactivity.
  virtual Status SetOrUpdateServerDef(const ServerDef& server_def,

  std::unique_ptr<TF_Status, decltype(&TF_DeleteStatus)> status(
      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 =
      TestScalarTensorHandle<T, datatype>(eager_ctx, value);
  *tensor =
      unwrap(TF_CreateAbstractTensorFromEagerTensor(input_eager, status.get()));

  // Returns the device ID which created the handle.
  virtual int DeviceId(Status* status) const = 0;
  // Returns a tensor for the handle. If tensor is remote, it will be copied.
  virtual AbstractTensorInterface* Resolve(Status* status) = 0;

  std::string DebugString() const override;

  // Returns a Boolean hint indicating whether callers should prefer
  // `SummarizeValue` to resolving this handle and formatting the tensor.

      ::tensorflow::Set_TF_Status_from_Status(C_STATUS, _s);      \
      TF_OpKernelContext_Failure(CTX, C_STATUS);                  \
      TF_DeleteStatus(C_STATUS);                                  \
      return;                                                     \
    }                                                             \
  } while (0)

#define TF_CLEANUP_AND_RETURN_IF_ERROR(C_STATUS, BUFFER, __VA_ARGS__) \

  void GetTensor(const string& name,
                 std::unique_ptr<tensorflow::Tensor>* out_tensor,
                 TF_Status* out_status) const;

 private:
  // Uses "v2_reader_" to build "var name -> shape" and "var name -> data type"
  // maps; both owned by caller.
  // REQUIRES: "v2_reader_ != nullptr && v2_reader_.status().ok()".
  std::pair<std::unique_ptr<TensorSliceReader::VarToShapeMap>,

                                           TF_Status* status);

void RecordMutation(TF_Graph* graph, const TF_Operation& op,
                    const char* mutation_type)
    TF_EXCLUSIVE_LOCKS_REQUIRED(graph->mu);

bool ExtendSessionGraphHelper(TF_Session* session, TF_Status* status)
    TF_LOCKS_EXCLUDED(session->graph->mu, session->mu);

//   // forward `op`.
//   return new AddGradientFunction;
// }
//
// Status RegisterGradients(GradientRegistry* registry) {
//   return registry->Register("Add", AddRegisterer);
// }
class GradientFunction {
 public:
  virtual Status Compute(AbstractContext* ctx,
                         absl::Span<AbstractTensorHandle* const> grad_outputs,

  ~ModularWritableFile() override { ops_->cleanup(file_.get()); }

  Status Append(StringPiece data) override;
  Status Close() override;
  Status Flush() override;
  Status Sync() override;
  Status Name(StringPiece* result) const override;
  Status Tell(int64_t* position) override;

 private:
  std::string filename_;
  std::unique_ptr<TF_WritableFile> file_;