: device_(device),
        tensors_(std::move(tensors)),
        shape_(std::vector<int64_t>(shape.begin(), shape.end())),
        dtype_(dtype) {}
  ParallelTensor(const ParallelDevice& device,
                 std::vector<TensorHandlePtr> tensors, const TF_DataType dtype)
      : device_(device),
        tensors_(std::move(tensors)),
        shape_(absl::nullopt),
        dtype_(dtype) {}

  Cleanup(Cleanup&& src)  // NOLINT
      : released_(src.is_released()), f_(src.release()) {}

  // Implicitly move-constructible from any compatible Cleanup<G>.
  // The source will be released as if src.release() were called.
  // A moved-from Cleanup can be safely destroyed or reassigned.
  template <typename G>
  Cleanup(Cleanup<G>&& src)  // NOLINT

      const char* description, LabelDesc&&... label) {
    sampler = absl::WrapUnique(tensorflow::monitoring::Sampler<NumLabels>::New(
        {name, description, label...}, std::move(buckets)));
  }

  std::unique_ptr<tensorflow::monitoring::Sampler<NumLabels>> sampler;
};

struct TFE_MonitoringSampler0 : TFE_MonitoringSampler<0> {
  using TFE_MonitoringSampler::TFE_MonitoringSampler;

// the same shape, although this requirement may be relaxed in the future.
//
// `device_name` must not name an existing physical or custom device (see
// the documentation for TFE_RegisterCustomDevice for more information).
//
// Tensors may be copied on or off the device explicitly using
// TPUReplicatedInput and TPUReplicatedOutput respectively. For example, with

#include "tensorflow/c/eager/parallel_device/parallel_device_lib.h"
#include "tensorflow/core/platform/test.h"

namespace tensorflow {
namespace parallel_device {

// A helper for performing common operations on variables. A much more
// restricted stand-in for tf.Variable in Python.
class Variable {
 public:
  // Construct a Variable from a resource-dtype TFE_TensorHandle and an
  // indication of the dtype of the variable's value.
  //

// will fail. If `prefix` is nullptr, the default prefixing behaviour takes
// place, see TF_AddGradients for more details.
//
// WARNING: This function does not yet support all the gradients that python
// supports. See
// https://www.tensorflow.org/code/tensorflow/cc/gradients/README.md
// for instructions on how to add C++ more gradients.
TF_CAPI_EXPORT void TF_AddGradientsWithPrefix(TF_Graph* g, const char* prefix,

TF_CAPI_EXPORT extern void TF_DeleteFile(const char* filename,
                                         TF_Status* status);

// Retrieves the next item from the given TF_StringStream and places a pointer
// to it in *result. If no more items are in the list, *result is set to NULL
// and false is returned.
//
// Ownership of the items retrieved with this function remains with the library.
// Item points are invalidated after a call to TF_StringStreamDone.

#include "tensorflow/c/eager/tfe_op_attrs_internal.h"  // IWYU pragma: export
#include "tensorflow/c/eager/tfe_tensor_debug_info_internal.h"  // IWYU pragma: export

// TODO(b/154564140): Move this to its own header. This requires splitting
// c_api_experimental.h
struct TFE_ContextOptions {
  TF_SessionOptions session_options;
  // true if async execution is enabled.
  bool async = false;

#include "tensorflow/core/platform/status.h"

namespace tensorflow {

// Abstract interface to an operation.
// This interface allows building and executing an operation in either
// tracing or immediate execution mode.
class AbstractOperation {
 protected:
  enum AbstractOperationKind {
    kGraph,
    kMlir,
    kEager,
    kTfrt,
    kTape,
    kOpHandler
  };

  TFE_DEVICE_PLACEMENT_SILENT_FOR_INT32 = 3,
} TFE_ContextDevicePlacementPolicy;
// LINT.ThenChange(//tensorflow/c/eager/immediate_execution_context.h)

// Sets the default execution mode (sync/async). Note that this can be
// overridden per thread using TFE_ContextSetExecutorForThread.
TF_CAPI_EXPORT extern void TFE_ContextOptionsSetAsync(TFE_ContextOptions*,