const TFE_OpAttrs* attributes, int expected_max_outputs,
                    CancellationManager& cancellation_manager,
                    absl::optional<int64_t> step_id = absl::nullopt);
  // Block until the previous `StartExecute` operation has executed. Forwards
  // the status from `TFE_Execute` and returns outputs if the status is OK.

  // well with `MatMul` and remove `TestMatMul*` in
  // `mnist_gradients_test` when done.
  GTEST_SKIP();

  float A_vals[] = {1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f, 7.0f, 8.0f, 9.0f};
  int64_t A_dims[] = {3, 3};
  AbstractTensorHandlePtr A;
  {
    AbstractTensorHandle* A_raw;
    status_ = TestTensorHandleWithDims<float, TF_FLOAT>(
        immediate_execution_ctx_.get(), A_vals, A_dims, 2, &A_raw);

TF_CAPI_EXPORT extern void TFE_CollectiveOpsCheckPeerHealth(
    TFE_Context* ctx, const char* task, int64_t timeout_in_ms,
    TF_Status* status);

// Information about the shape of a Tensor and its type.
struct TF_ShapeAndType {
  // Number of dimensions. -1 indicates unknown rank.
  int num_dims;
  // Array of dimensions. -1 indicates unknown dim.
  int64_t* dims;
  // The data type. May be 0 to denote unknown type.
  TF_DataType dtype;
};

                                 const int64_t* dims, int num_dims) {
  DCHECK(index >= 0 && index < shape_list->num_items);
  TF_ShapeAndType& shape = shape_list->items[index];
  DCHECK(shape.dims == nullptr) << "Shape at " << index << " is already set!";
  DCHECK(num_dims >= 0) << "Number of dimensions cannot be negative!";
  shape.num_dims = num_dims;
  shape.dims = new int64_t[num_dims];

  if (metadata)
    return ::testing::AssertionSuccess();
  else
    return ::testing::AssertionFailure() << metadata.status().message();
}

::testing::AssertionResult CompareSubString(int64_t offset, size_t length,
                                            absl::string_view result,
                                            size_t read) {
  // Result isn't a null-terminated string so we have to wrap it inside a

    CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
    size_t size = TFE_TensorHandleDeviceMemorySize(copy, status);
    CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
    int64_t dims[] = {2, 2};
    TFE_TensorHandle* copy_aliased = TFE_NewTensorHandleFromDeviceMemory(
        ctx, name, TF_FLOAT, dims, 2, data, size, &Deleter, copy, status);
    CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  const std::vector<std::unique_ptr<ParallelTensor>>& handles = *outputs;
  const std::vector<int64_t>* shape;
  Status s = handles[0]->Shape(&shape);
  ASSERT_TRUE(s.ok());
  EXPECT_EQ(0, shape->size());
}

TEST(PARALLEL_DEVICE_LIB, TestCancelOnError) {

  const std::vector<int64_t>* shape;
  Status s = reinterpret_cast<ParallelTensor*>(data)->Shape(&shape);
  if (!s.ok()) {
    tsl::Set_TF_Status_from_Status(status, s);
    return -1;
  }
  return shape->size();
}

// Used as an argument to TFE_NewCustomDeviceTensorHandle, for computing a
// dimension of a parallel tensor.

// another option.

namespace tensorflow {
namespace parallel_device {

Variable* Variable::Create(TFE_Context* context, TF_DataType type,
                           const int64_t* dims, const int num_dims,
                           const char* device, TF_Status* status) {
  std::unique_ptr<TFE_Op, decltype(&TFE_DeleteOp)> op(
      TFE_NewOp(context, "VarHandleOp", status), TFE_DeleteOp);