// will eventually contain the full while loop.
  TF_Graph* parent;
  TF_Output* parent_inputs;
};

struct TF_OperationDescription {
  TF_OperationDescription(TF_Graph* g, const char* op_type,
                          const char* node_name)
      : node_builder(node_name, op_type, g->graph.op_registry()), graph(g) {}

  tensorflow::NodeBuilder node_builder;
  TF_Graph* graph;

#include <vector>

#include "tensorflow/c/c_api.h"
#include "tensorflow/c/experimental/grappler/grappler.h"
#include "tensorflow/core/framework/graph.pb.h"
#include "tensorflow/core/grappler/optimizers/custom_graph_optimizer.h"
#include "tensorflow/core/grappler/optimizers/custom_graph_optimizer_registry.h"
#include "tensorflow/core/platform/status.h"
#include "tensorflow/core/protobuf/rewriter_config.pb.h"

namespace tensorflow {

// TF_InitGraph is used to do graph optimizer registration.
// Plugin should implement TF_InitGraph to register graph optimizers.
void TF_InitGraph(TP_OptimizerRegistrationParams* params, TF_Status* status);

// Get a set of node names that must be preserved. They can not be transformed
// or removed during the graph transformation. This includes feed and fetch

// A TF_ExecutionContext stores knowledge about how to execute an operation.
// E.g. it could know whether we're in eager mode or graph mode, keeps track
// of gradient tapes, etc.
typedef struct TF_ExecutionContext TF_ExecutionContext;

// A TF_AbstractTensor is an input to an operation. E.g. it could be a union
// type of eager and graph tensors. It is also the result of executing an
// operation.
typedef struct TF_AbstractTensor TF_AbstractTensor;

 * 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.
 */

                                       TF_Status* status);

// Enables only graph collection in RunMetadata on the functions executed from
// this context.
TF_CAPI_EXPORT extern void TFE_ContextEnableGraphCollection(TFE_Context* ctx);

// Disables only graph collection in RunMetadata on the functions executed from
// this context.

// The new graph construction API, still under development.

// Represents a computation graph.  Graphs may be shared between sessions.
// Graphs are thread-safe when used as directed below.
typedef struct TF_Graph TF_Graph;

// Return a new graph object.
TF_CAPI_EXPORT extern TF_Graph* TF_NewGraph(void);

// Destroy an options object. Graph will be deleted once no more

// (for now at least).
// Implementations of the context may contain some state e.g. an execution
// environment, a traced representation etc.
class AbstractContext {
 protected:
  enum AbstractContextKind { kGraph, kMlir, kEager, kTfrt, kTape, kOpHandler };
  explicit AbstractContext(AbstractContextKind kind) : kind_(kind) {}
  virtual ~AbstractContext() {}

 public:
  AbstractContextKind getKind() const { return kind_; }

TF_CAPI_EXPORT extern TF_Buffer* TF_CreateRunOptions(
    unsigned char enable_full_trace);

// Returns the graph content in a human-readable format, with length set in
// `len`. The format is subject to change in the future.
// The returned string is heap-allocated, and caller should call free() on it.
TF_CAPI_EXPORT extern const char* TF_GraphDebugString(TF_Graph* graph,
                                                      size_t* len);

// representation between various supported formats e.g. FunctionDef and Mlir
// function.
class AbstractFunction : public core::RefCounted {
 protected:
  enum AbstractFunctionKind { kGraph, kMlir };
  explicit AbstractFunction(AbstractFunctionKind kind) : kind_(kind) {}

 public:
  // Returns which subclass is this instance of.
  AbstractFunctionKind getKind() const { return kind_; }