TF_DataType datatype_enum(
      static_cast<TF_DataType>(DataTypeToEnum<DataType>().value));
  for (int device_index = 0; device_index < num_underlying_devices();
       ++device_index) {
    auto device_value = absl::make_unique<DataType>();
    *device_value = values[device_index];
    std::unique_ptr<TF_Tensor, decltype(&TF_DeleteTensor)> tensor(
        TF_NewTensor(
            datatype_enum, /*dims=*/nullptr, /*num_dims=*/0,

//   TF_Output values_inputs[5] = {{...}, ..., {...}};
//   TF_AddInputList(desc, values_inputs, 5);

// For inputs that take a single tensor.
TF_CAPI_EXPORT extern void TF_AddInput(TF_OperationDescription* desc,
                                       TF_Output input);

// For inputs that take a list of tensors.
// inputs must point to TF_Output[num_inputs].
TF_CAPI_EXPORT extern void TF_AddInputList(TF_OperationDescription* desc,

//
// Very similar to TF_OperationDescription with some differences:
// (1) TF_Output or TFE_TensorHandle* as arguments to TF_AddInput,
//     TF_AddInputList
// (2) TF_ColocateWith, TF_AddControlInput etc. do not make sense.
// (3) Implementation detail: Avoid use of NodeBuilder/NodeDefBuilder since
//     the additional sanity checks there seem unnecessary;
typedef struct TFE_Op TFE_Op;

#include <memory>

#include "tensorflow/core/lib/monitoring/cell_reader.h"

struct TFE_MonitoringCounterReader {
  explicit TFE_MonitoringCounterReader(const char* name) {
    counter = std::make_unique<
        ::tensorflow::monitoring::testing::CellReader<int64_t>>(name);
  }
  template <typename... LabelType>
  int64_t Read(const LabelType&... labels);

      ABSL_EXCLUSIVE_LOCKS_REQUIRED(mu_) {
    lru_list_.push_front(key);
    Entry entry{timer_seconds_(), value, lru_list_.begin()};
    auto insert = cache_.insert(std::make_pair(key, entry));
    if (!insert.second) {
      lru_list_.erase(insert.first->second.lru_iterator);
      insert.first->second = entry;
    } else if (max_entries_ > 0 && cache_.size() > max_entries_) {

// Tensor passed to Accumulate goes out of scope.
//
// Not thread-safe.
template <typename Gradient, typename BackwardFunction, typename TapeTensor>
class ForwardAccumulator {
 public:
  // Does not take ownership of `vspace`, which must outlive the
  // ForwardAccumulator.
  explicit ForwardAccumulator(
      const VSpace<Gradient, BackwardFunction, TapeTensor>& vspace,
      bool use_batch)

// Deletes the eager Executor without waiting for enqueued nodes. Please call
// TFE_ExecutorWaitForAllPendingNodes before calling this API if you want to
// make sure all nodes are finished.
TF_CAPI_EXPORT extern void TFE_DeleteExecutor(TFE_Executor*);

// Returns true if the executor is in async mode.
TF_CAPI_EXPORT extern bool TFE_ExecutorIsAsync(TFE_Executor*);

namespace internal {
struct AbstractTensorHandleDeleter {
  void operator()(AbstractTensorHandle* p) const {
    if (p != nullptr) {
      p->Unref();
    }
  }
};
}  // namespace internal

// TODO(b/185908092): Make AbstractTensorHandlePtr an IntrusivePtr.
using AbstractTensorHandlePtr =
    std::unique_ptr<AbstractTensorHandle,
                    internal::AbstractTensorHandleDeleter>;

}  // namespace tensorflow

Status SetAttrFunctionList(AbstractOperation*, const char* attr_name,
                           absl::Span<const AbstractOperation*> values,
                           ForwardOperation*);

// Make the call to `Tape::RecordOperation`.
Status Execute(AbstractOperation*, AbstractContext*,
               absl::Span<AbstractTensorHandle*> retvals, int* num_retvals,
               ForwardOperation*, Tape*, const GradientRegistry&);

  virtual Status AddParameter(DataType dtype, const PartialTensorShape& shape,
                              TracingTensorHandle**) = 0;

  // Finalize this context and make a function out of it. The context is in a
  // invalid state after this call and must be destroyed.
  virtual Status Finalize(OutputList* outputs, AbstractFunction**) = 0;

  // For LLVM style RTTI.