// 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*);

  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,

  }
  if (!ShouldRecord(input_tensor_id, input_dtypes)) {
    return absl::OkStatus();
  }

  // We may need to allocate zero inputs for trainable dtypes we don't have JVPs
  // for. Make sure they get cleaned up.
  std::vector<Gradient*> new_zeros;
  auto delete_new_zeros = gtl::MakeCleanup([&new_zeros, this] {
    for (Gradient* tensor : new_zeros) {
      this->vspace_.DeleteGradient(tensor);
    }

//
// 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;

  void MaybeFetch(const Key& key, const std::shared_ptr<Block>& block,
                  TF_Status* status) ABSL_LOCKS_EXCLUDED(mu_);

  /// Trim the block cache to make room for another entry.
  void Trim() ABSL_EXCLUSIVE_LOCKS_REQUIRED(mu_);

  /// Update the LRU iterator for the block at `key`.
  void UpdateLRU(const Key& key, const std::shared_ptr<Block>& block,