TapeVSpace vspace(ctx);
  std::vector<int64_t> target_tensor_ids = MakeTensorIDList(targets);
  std::vector<int64_t> source_tensor_ids = MakeTensorIDList(sources);
  tensorflow::gtl::FlatSet<int64_t> sources_set(source_tensor_ids.begin(),
                                                source_tensor_ids.end());
  std::unordered_map<int64_t, TapeTensor> sources_that_are_targets;

    return absl::OkStatus();
  }
  absl::Status SetAttrShape(const char* attr_name, const int64_t* dims,
                            const int num_dims) override {
    PartialTensorShape shape;
    if (num_dims >= 0) {
      shape = PartialTensorShape(ArraySlice<int64_t>(
          reinterpret_cast<const int64_t*>(dims), num_dims));
    }
    op_->node_builder.Attr(attr_name, shape);
    return absl::OkStatus();

  return th;
}

TFE_TensorHandle* DoubleTestMatrixTensorHandle(TFE_Context* ctx) {
  int64_t dims[] = {2, 2};
  double data[] = {1.0, 2.0, 3.0, 4.0};
  TF_Status* status = TF_NewStatus();
  TF_Tensor* t = TFE_AllocateHostTensor(ctx, TF_DOUBLE, &dims[0],
                                        sizeof(dims) / sizeof(int64_t), status);
  memcpy(TF_TensorData(t), &data[0], TF_TensorByteSize(t));

  // Get a list of the names of functions that have been registered.
  virtual std::vector<string> ListFunctionNames() = 0;

  struct CacheStats {
    int64_t kernel_cache_size;
    int64_t device_cache_size;
    std::map<std::string, int64_t> func_kernel_cache_entries;
    int64_t local_rendezvous_cache_active_size;
  };
  virtual CacheStats GetCacheStats() = 0;

}

std::unique_ptr<ParallelTensor> ParallelDevice::DeviceIDs(
    TFE_Context* context, TF_Status* status) const {
  std::vector<int32_t> ids;
  ids.reserve(num_underlying_devices());
  for (int i = 0; i < num_underlying_devices(); ++i) {
    ids.push_back(i);
  }
  return ScalarsFromSequence<int32_t>(ids, context, status);
}

absl::optional<std::vector<std::unique_ptr<ParallelTensor>>>

// Atomically increments the value of the cell. The value must be non-negative.
TF_CAPI_EXPORT extern void TFE_MonitoringCounterCellIncrementBy(
    TFE_MonitoringCounterCell* cell, int64_t value);

// Retrieves the current value of the cell.
TF_CAPI_EXPORT extern int64_t TFE_MonitoringCounterCellValue(
    TFE_MonitoringCounterCell* cell);

// APIs for Counter without label.
typedef struct TFE_MonitoringCounter0 TFE_MonitoringCounter0;

    TF_DeleteShapeAndTypeList(output_shapes);
  }

  absl::optional<std::vector<int64_t>> make_shape(
      std::vector<int64_t>&& dims) const {
    return absl::make_optional(dims);
  }

  absl::optional<std::vector<int64_t>> unknown_shape() const {
    return absl::nullopt;
  }

  static constexpr int64_t kUnknownDim =
      shape_inference::InferenceContext::kUnknownDim;
  TF_Status* status_;

}

// Checks whether the stride array matches the layout of compact, row-majored
// data.
bool IsValidStrideCompactRowMajorData(int64_t* shape_arr, int64_t* stride_arr,
                                      int ndim) {
  bool valid = true;
  int64_t expected_stride = 1;
  for (int i = ndim - 1; i >= 0; --i) {
    // Empty tensors are always compact regardless of strides.
    if (shape_arr[i] == 0) return true;

  return num_dims;
}

int64_t TFE_TensorHandleNumElements(TFE_TensorHandle* h, TF_Status* status) {
  if (h == nullptr) {
    status->status = tensorflow::errors::InvalidArgument("Invalid handle");
    return -1;
  }

  int64_t num_elements = -1;
  status->status = tensorflow::unwrap(h)->NumElements(&num_elements);
  return num_elements;
}

int64_t TFE_TensorHandleDim(TFE_TensorHandle* h, int dim_index,

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