// gradient registerer to instantiate a GradientFunction.
struct ForwardOperation {
 public:
  string op_name;
  std::vector<AbstractTensorHandle*> inputs;
  std::vector<AbstractTensorHandle*> outputs;
  std::vector<int64_t> skip_input_indices;
  AttrBuilder attrs;
};

using GradientFunctionFactory =
    std::function<GradientFunction*(const ForwardOperation& op)>;

// Map from op name to a `GradientFunctionFactory`.

  // Returns the AbstractFunction as a FunctionDef.
  virtual void GetNameAttrList(
      tensorflow::NameAttrList* name_and_attrs) const = 0;

  virtual bool GetInt(absl::string_view, int64_t* result) const = 0;
  virtual bool GetFloat(absl::string_view attr_name, float* result) const = 0;
  virtual bool GetBool(absl::string_view attr_name, bool* result) const = 0;

  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  const std::vector<std::unique_ptr<ParallelTensor>>& handles = *outputs;
  const std::vector<int64_t>* shape;
  absl::Status s = handles[0]->Shape(&shape);
  ASSERT_TRUE(s.ok());
  EXPECT_EQ(0, shape->size());
}

TEST(PARALLEL_DEVICE_LIB, TestCancelOnError) {

      CancellationManager* cancellation_manager) = 0;

  // Returns the stack trace set by `SetStackTrace` if exists.
  virtual absl::optional<ManagedStackTrace> GetStackTrace() = 0;

  virtual void SetStepId(int64_t step_id) = 0;

  // For LLVM style RTTI.
  static bool classof(const AbstractOperation* ptr) {
    return ptr->getKind() == kEager || ptr->getKind() == kTfrt;
  }

 protected:

  // `keep_alive_secs` of inactivity.
  virtual absl::Status SetOrUpdateServerDef(
      const ServerDef& server_def, bool reset_context, int keep_alive_secs,
      int64_t init_timeout_in_ms, int retries,
      bool clear_existing_contexts = false) = 0;

  // Initializes context for the local worker and no contexts will be created

  const std::vector<int64_t>* shape;
  absl::Status s = reinterpret_cast<ParallelTensor*>(data)->Shape(&shape);
  if (!s.ok()) {
    tsl::Set_TF_Status_from_Status(status, s);
    return -1;
  }
  return shape->size();
}

// Used as an argument to TFE_NewCustomDeviceTensorHandle, for computing a
// dimension of a parallel tensor.

  /* Want to test simple MatMul example:
    [[0,0],    *   [[0,0],    =   [[0,0],
     [0,0]]         [0,0]]         [0,0]]
  */

  // Build an abstract input tensor.
  int64_t dims[] = {2, 2};  // Matrices will be 2 x 2
  int num_dims = sizeof(dims) / sizeof(dims[0]);

  float vals[] = {0.0f, 0.0f, 0.0f, 0.0f};
  TFE_Context* eager_ctx = TF_ExecutionContextGetTFEContext(ctx, status.get());

void TF_DeleteExecutionContext(TF_ExecutionContext*);

// Represents a (partially-defined) shape.
typedef struct TF_Shape {
  int num_dims;  // Must be >= -1; -1 represents unknown rank.
  int64_t* dim_sizes;
} TF_Shape;

// Add a new parameter to a TensorFlow Function.
TF_AbstractTensor* TF_AddFunctionParameter(TF_ExecutionContext* func,
                                           TF_DataType dtype, TF_Shape shape,

void TF_GraphSetOutputHandleShapesAndTypes(TF_Graph* graph, TF_Output output,
                                           int num_shapes_and_types,
                                           const int64_t** shapes,
                                           const int* ranks,
                                           const TF_DataType* types,
                                           TF_Status* status);

};

struct LoggedTensor {
  TFE_TensorHandle* tensor;
  LoggedTensor() = delete;
  explicit LoggedTensor(TFE_TensorHandle* tensor) : tensor(tensor) {}
  ~LoggedTensor() { TFE_DeleteTensorHandle(tensor); }
};

int64_t LoggedTensorDim(void* data, int dim_index, TF_Status* status) {
  return TFE_TensorHandleDim(reinterpret_cast<LoggedTensor*>(data)->tensor,
                             dim_index, status);
}