* parameter specified by `input_index`.
 *
 * I.e. if y = <output of the forward model> and w = inputs[input_index],
 * this will calculate dy/dw numerically.
 *
 * `use_function` indicates whether to use graph mode(true) or eager(false).
 *
 * `numerical_grad` is the pointer to the AbstractTensorHandle* which will
 * hold the numerical gradient data at the end of the function.
 */

// status.
TF_CAPI_EXPORT extern void TF_AppendWritableFile(TF_WritableFileHandle* handle,
                                                 const char* data,
                                                 size_t length,
                                                 TF_Status* status);

// Deletes the named file and indicates whether successful in *status.

// C API for Graph. The API is under active development and eventually
// should allow registering a plugin graph optimizer with TensorFlow.
//
// Conventions:
//   * Struct prefix indicates whether struct fields should be filled by the
//     plugin or core implementation:
//     * Struct that should be filled by the plugin: `TP_OptimizerConfigs`,
//       `TP_Optimizer`, `TP_OptimizerRegistrationParams`

    TF_Status* status);

// Information about the shape of a Tensor and its type.
struct TF_ShapeAndType {
  // Number of dimensions. -1 indicates unknown rank.
  int num_dims;
  // Array of dimensions. -1 indicates unknown dim.
  int64_t* dims;
  // The data type. May be 0 to denote unknown type.
  TF_DataType dtype;
};

typedef struct TF_ShapeAndType TF_ShapeAndType;

// following maps the operation name to the indices at which the corresponding
// gradient function can accept None values. e.g. FusedBatchNorm outputs 5
// values and hence receives 5 gradient values during backprop. However the
// gradient function uses only the first of those values and ignores the rest.
// The entry, "FusedBatchNorm": [1, 2, 3, 4], indicates that only the gradient

// tensors/ops/etc. and usable in APIs like OpSetDevice/ResetOp/etc.

#define TFE_CUSTOM_DEVICE_VERSION 4

// Struct to be filled in. Functions are required except where indicated.
typedef struct TFE_CustomDevice {
  int version = TFE_CUSTOM_DEVICE_VERSION;
  // Method to copy a tensor to the custom device.
  TFE_TensorHandle* (*copy_tensor_to_device)(TFE_Context* context,

// `oper` and `num_dims` with the corresponding number of dimensions. On return,
// for every i where `num_dims[i]` > 0, `dims[i]` will be an array of
// `num_dims[i]` elements. A value of -1 for `num_dims[i]` indicates that the
// i-th shape in the list is unknown.
//
// The elements of `dims` will point to addresses in `storage` which must be
// large enough to hold at least `storage_size` int64_ts.  Ideally, `num_shapes`

  virtual Status NumDims(int* num_dims) const = 0;
  // Returns number of elements across all dimensions.
  virtual Status NumElements(int64_t* num_elements) const = 0;
  // Returns size of specified dimension
  //
  // -1 indicates an unknown axis length; this is unreachable for most standard
  // ImmediateExecutionTensorHandles, but comes up for example when computing
  // the shape of a parallel tensor with component shapes differing across
  // devices.

typedef struct TFE_TensorHandle TFE_TensorHandle;

TF_CAPI_EXPORT extern TFE_TensorHandle* TFE_NewTensorHandle(const TF_Tensor* t,
                                                            TF_Status* status);
// Indicates that the caller will not be using `h` any more.
TF_CAPI_EXPORT extern void TFE_DeleteTensorHandle(TFE_TensorHandle* h);
TF_CAPI_EXPORT extern TF_DataType TFE_TensorHandleDataType(TFE_TensorHandle* h);

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`.
class GradientRegistry {