std::vector<int64_t> shape;
  int rank = -1;
  *status = handle.NumDims(&rank);
  if (!status->ok()) {
    return shape;
  }
  shape.reserve(rank);
  for (int i = 0; i < rank; ++i) {
    int64_t dim;
    *status = handle.Dim(i, &dim);
    if (!status->ok()) {
      return shape;
    }
    shape.push_back(dim);
  }
  return shape;
}

}  // namespace

extern "C" {

  EXPECT_TRUE(
      std::equal(std::begin(partial_shape), std::end(partial_shape), values));
}

TEST_F(CApiAttributesTest, ShapeList) {
  const int64_t shape_1[] = {1, 3};
  const int64_t shape_2[] = {2, 4, 6};
  const int64_t* list[] = {&shape_1[0], &shape_2[0]};
  const size_t list_size = TF_ARRAYSIZE(list);
  const int ndims[] = {TF_ARRAYSIZE(shape_1), TF_ARRAYSIZE(shape_2)};

      TFE_NewOp(context, "VarHandleOp", status), TFE_DeleteOp);
  if (TF_GetCode(status) != TF_OK) return nullptr;
  TFE_OpSetAttrType(op.get(), "dtype", type);
  TFE_OpSetAttrShape(op.get(), "shape", dims, num_dims, status);
  TFE_OpSetAttrString(op.get(), "container", "", 0);
  // Use the special GUID for no buffer sharing
  //
  // TODO(allenl): Should we provide a better API for this? AFAIK this is the

      break;
    case tensorflow::AttrValue::kShape: {
      const auto& tensor_shape = default_value.shape();
      if (tensor_shape.unknown_rank()) {
        TFE_OpSetAttrShape(op, attr_name, nullptr, -1, status);
      } else {
        const auto num_dims = tensor_shape.dim_size();
        std::unique_ptr<int64_t[]> dims(new int64_t[num_dims]);

                                                           TF_DeleteStatus);

  TF_Tensor* tensor_shape = Int32Tensor({37, 1});
  TF_Operation* shape = Const(tensor_shape, func_graph.get(), s.get(), "shape");
  TF_Operation* random =
      RandomUniform(shape, TF_FLOAT, func_graph.get(), s.get());

  TF_Output outputs[] = {{random, 0}};
  *func = TF_GraphToFunction(func_graph.get(), name,

  // Create the variable handle.
  TFE_Op* op = TFE_NewOp(ctx, "VarHandleOp", status);
  if (TF_GetCode(status) != TF_OK) return nullptr;
  TFE_OpSetAttrType(op, "dtype", TF_FLOAT);
  TFE_OpSetAttrShape(op, "shape", {}, 0, status);
  TFE_OpSetAttrString(op, "container", "localhost", 0);
  TFE_OpSetAttrString(op, "shared_name", "", 0);
  if (!device_name.empty()) {
    TFE_OpSetDevice(op, device_name.c_str(), status);
  }

    ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());

    TFE_Op* shape_op = ShapeOp(ctx, hgpu);
    TFE_OpSetDevice(shape_op, gpu_device_name.c_str(), status.get());
    ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
    TFE_TensorHandle* retvals[1];
    int num_retvals = 1;
    TFE_Execute(shape_op, &retvals[0], &num_retvals, status.get());

    ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());

    TFE_Op* shape_op = ShapeOp(ctx, hgpu);
    TFE_OpSetDevice(shape_op, gpu_device_name.c_str(), status.get());
    ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
    TFE_TensorHandle* retvals[1];
    int num_retvals = 1;
    TFE_Execute(shape_op, &retvals[0], &num_retvals, status.get());

  }

  Status Compute(AbstractContext* ctx,
                 absl::Span<AbstractTensorHandle* const> grad_outputs,
                 absl::Span<AbstractTensorHandle*> grad_inputs) override {
    // TODO(vnvo2409): Add shape broadcasting
    /* Given upstream grad U and a Div op: Z = X/Y, the gradients are:
     *
     *    dX = U / Y
     *    dY = -U*X / Y^2 = (X/Y) * -U / Y = -U*Z / Y
     *
     */