int num_dims);

// Get a Matrix TensorHandle with given float values and dimensions
TFE_TensorHandle* TestTensorHandleWithDimsFloat(TFE_Context* ctx, float data[],
                                                int64_t dims[], int num_dims);

// Get a Matrix TensorHandle with given int values and dimensions

    const TF_Graph* fn_body, const char* fn_name,
    unsigned char append_hash_to_fn_name, int num_opers,
    const TF_Operation* const* opers, int ninputs, const TF_Output* inputs,
    int noutputs, const TF_Output* outputs, const char* const* output_names,
    int ncontrol_outputs, const TF_Operation* const* control_outputs,
    const char* const* control_output_names, const TF_FunctionOptions* opts,

//
// This method uses `sendfile` if available (i.e., linux 2.6.33 or later) or an
// intermediate buffer if not.
//
// Returns number of bytes transferred or -1 on failure.
int CopyFileContents(int dst_fd, int src_fd, off_t size);

}  // namespace tf_posix_filesystem

 public:
  // Eager async execution is only supported when remote eager is not in use
  // (b/157523095).
  explicit ParallelDevice(const std::vector<std::string>& devices,
                          bool is_async = false, int in_flight_nodes_limit = 0);

  ~ParallelDevice();

  // Helper to copy a tensor handle from another device once for each component
  // of the ParallelDevice.
  //

  TFE_OpSetAttrInt(op.get(), "N", num_replicas);
  for (int i = 0; i < num_replicas; ++i) {
    TFE_OpAddInput(op.get(), components[i], status);
    if (TF_GetCode(status) != TF_OK) return nullptr;
  }
  TFE_OpSetDevice(op.get(), device, status);
  if (TF_GetCode(status) != TF_OK) return nullptr;

  TFE_TensorHandle* result_handle;
  int num_retvals = 1;
  TFE_Execute(op.get(), &result_handle, &num_retvals, status);

// passed to TFE_RegisterCustomDevice. The `device_name` arguments must match.
void AllocateParallelDevice(const char* device_name,
                            const char* const* underlying_devices,
                            int num_underlying_devices,
                            TFE_CustomDevice* device, void** device_info);

}  // namespace parallel_device
}  // namespace tensorflow

void PathExists(const TF_Filesystem* filesystem, const char* path,
                TF_Status* status);
void CreateDir(const TF_Filesystem* filesystem, const char* path,
               TF_Status* status);
int GetChildren(const TF_Filesystem* filesystem, const char* path,
                char*** entries, TF_Status* status);
void DeleteFile(const TF_Filesystem* filesystem, const char* path,
                TF_Status* status);

  // 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.
  int numa_node;
} TF_ThreadOptions;

// Creates the specified directory. Typical status code are:
//  * TF_OK - successfully created the directory
//  * TF_ALREADY_EXISTS - directory already exists

                                                 const TF_DataType* values,
                                                 int num_values);
TF_CAPI_EXPORT extern void TFE_OpSetAttrShapeList(
    TFE_Op* op, const char* attr_name, const int64_t** dims,
    const int* num_dims, int num_values, TF_Status* out_status);
TF_CAPI_EXPORT extern void TFE_OpSetAttrFunctionList(TFE_Op* op,

  // Releases the cleanup function instead of running it.
  // Hint: use c.release()() to run early.
  F release() {
    released_ = true;
    return std::move(f_);
  }

  bool is_released() const { return released_; }

 private:
  static_assert(!std::is_reference<F>::value, "F must not be a reference");

  bool released_ = false;
  F f_;
};

template <int&... ExplicitParameterBarrier, typename F,