const char* worker_name,
                                                 TF_Status* status);

// Sync pending nodes in local executors (including the context default executor
// and thread executors) and streaming requests to remote executors, and get the
// combined status.
TF_CAPI_EXPORT extern void TFE_ContextAsyncWait(TFE_Context* ctx,

typedef struct TF_Thread TF_Thread;

typedef struct TF_ThreadOptions {
  // Thread stack size to use (in bytes), zero implies that the system default
  // will be used.
  size_t stack_size;

  // Guard area size to use near thread stacks to use (in bytes), zero implies
  // that the system default will be used.
  size_t guard_size;

  // The NUMA node to use, -1 implies that there should be no NUMA affinity for
  // this thread.

#else
#define TF_CAPI_EXPORT __declspec(dllimport)
#endif  // TF_COMPILE_LIBRARY
#else
#ifdef TF_CAPI_WEAK
#define TF_CAPI_EXPORT \
  __attribute__((visibility("default"))) __attribute((weak))
#else
#define TF_CAPI_EXPORT __attribute__((visibility("default")))
#endif  // TF_CAPI_WEAK
#endif  // _WIN32
#endif  // SWIG

// TF_Bool is the C API typedef for unsigned char, while TF_BOOL is
// the datatype for boolean tensors.

DEFINE_CONVERSION_FUNCTIONS(tensorflow::AbstractOpAttrs, TFE_OpAttrs);

// Set an AttrValue on the op. Doesn't handle the list types.
void SetOpAttrValueScalar(TFE_Context* ctx, TFE_Op* op,
                          const tensorflow::AttrValue& default_value,
                          const char* attr_name, TF_Status* status);
}  // namespace tensorflow

  tensorflow::NodeBuilder node_builder;
  TF_Graph* graph;
  std::set<tensorflow::string> colocation_constraints;
};

struct TF_Operation {
  tensorflow::Node node;

 private:
  ~TF_Operation() = default;
};

struct TF_Session {
  TF_Session(tensorflow::Session* s, TF_Graph* g);

  tensorflow::Session* session;
  TF_Graph* const graph;

  tensorflow::mutex mu TF_ACQUIRED_AFTER(TF_Graph::mu);

  // are executed asynchronously this must outlive the queued op, so it can't be
  // function-local to Execute.
  mutable std::unique_ptr<CancellationManager> default_cancellation_manager_;
};

// Contains a tuple of tensors, one on each of the `underlying_devices_` of the
// ParallelDevice.
class ParallelTensor {
 public:

  // least one input of the op is watched and has trainable dtype.
  //
  // op_type is used to decide which of the incoming gradients can be left as
  // nullptr instead of building zeros when build_default_zeros_grads == true.
  void RecordOperation(
      const string& op_type, const std::vector<TapeTensor>& output_tensors,
      absl::Span<const int64_t> input_tensor_id,

  explicit AbstractOpAttrs(AbstractOpAttrsKind kind) : kind_(kind) {}

 public:
  // Returns which subclass is this instance of.
  AbstractOpAttrsKind getKind() const { return kind_; }
  virtual ~AbstractOpAttrs() = default;

  // 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 PreferCustomSummarizer() const { return false; }

  // Returns a string which summarizes the value of this TensorHandle, for
  // debugging. Does not include a shape or dtype.
  //
  // Included in the default implementation of DebugString.
  virtual Status SummarizeValue(std::string& summary) const;

  // For LLVM style RTTI.
  static bool classof(const AbstractTensorHandle* ptr) {

#include "tensorflow/c/tf_status_helper.h"
#include "tensorflow/c/tf_tensor.h"
#include "tensorflow/core/platform/status.h"

namespace tensorflow {

// Builds and returns a `TracingContext` using the default tracing impl.
AbstractContext* BuildFunction(const char* fn_name);

// Creates parameters (placeholders) in the tracing `ctx` using the shape and
// dtype of `inputs`.
Status CreateParamsForInputs(AbstractContext* ctx,