"/job:localhost/replica:0/task:0/device:CPU:1"};
  ParallelDevice parallel_device(std::move(devices));
  std::unique_ptr<TFE_Op, decltype(&TFE_DeleteOp)> handle_op(
      TFE_NewOp(context.get(), "VarHandleOp", status.get()), TFE_DeleteOp);
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  TFE_OpSetAttrType(handle_op.get(), "dtype", TF_FLOAT);

  ASSERT_TRUE(arrived);
  ASSERT_FALSE(executed);
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  std::unique_ptr<TFE_Op, decltype(&TFE_DeleteOp)> matmul(
      MatMulOp(context, hcpu, hdevice), TFE_DeleteOp);
  TFE_OpSetDevice(matmul.get(), name, status.get());
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  TFE_TensorHandle* retval;
  int num_retvals = 1;

        << TF_Message(status.get());
  }

  {
    // Try to pass a non-parallel tensor to TPUReplicatedOutput
    std::unique_ptr<TFE_Op, decltype(&TFE_DeleteOp)> op(
        TFE_NewOp(context.get(), "TPUReplicatedOutput", status.get()),
        TFE_DeleteOp);
    if (TF_GetCode(status.get()) != TF_OK) return;
    TFE_OpSetAttrInt(op.get(), "num_replicas", 1);
    TFE_OpAddInput(op.get(), value_one.get(), status.get());

                    /*input_shapes*/ {unknown_shape(), unknown_shape()},
                    /*input_tensors*/ {},
                    /*expected_shape*/ make_shape({kUnknownDim, kUnknownDim}));

  TFE_DeleteOp(matmul_op);
  // TODO(bgogul): Add some death tests where status is not OK.
}

TEST_F(ShapeInferenceTest, InfersShapesFromInputTensors) {
  // Prepare some tensors for shape.

  CheckTFE_TensorHandleHasFloats(retval_task0, {7, 10, 15, 22});

  TFE_DeleteTensorHandle(retval_task0);
  TFE_DeleteTensorHandle(h0_task0);
  TFE_DeleteTensorHandle(retvals[0]);

  TFE_DeleteOp(matmul);

  TFE_Executor* executor = TFE_ContextGetExecutorForThread(ctx);
  TFE_ExecutorWaitForAllPendingNodes(executor, status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  TFE_TensorHandle* retvals[1] = {nullptr};
  int num_retvals = 1;
  TFE_Execute(func, &retvals[0], &num_retvals, status);
  EXPECT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);
  ASSERT_EQ(1, num_retvals);
  TFE_DeleteOp(func);
  TFE_DeleteTensorHandle(packed_handle);
  TF_Tensor* t = TFE_TensorHandleResolve(retvals[0], status);
  ASSERT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);
  TFE_DeleteTensorHandle(retvals[0]);

  TFE_TensorHandle* var_handle = nullptr;
  int num_retvals = 1;
  TFE_Execute(op, &var_handle, &num_retvals, status);
  if (TF_GetCode(status) != TF_OK) return nullptr;
  TFE_DeleteOp(op);
  if (TF_GetCode(status) != TF_OK) return nullptr;
  CHECK_EQ(1, num_retvals);

  // Assign 'value' to it.
  op = TFE_NewOp(ctx, "AssignVariableOp", status);

    std::vector<TFE_TensorHandle*> result;
    result.push_back(nullptr);
    int num_retvals = 1;
    TFE_Execute(op, result.data(), &num_retvals, status);
    TFE_DeleteOp(op);
    ASSERT_TRUE(TF_GetCode(status) == TF_OK) << TF_Message(status);
    ASSERT_EQ(num_retvals, 1);

    TF_Tensor* r = TFE_TensorHandleResolve(result[0], status);

    TFE_TensorDebugInfo* debug_info, int dim_index);

// Description of the TensorFlow op to execute.
//
// Assumes that the provided 'ctx' outlives the returned TFE_Op, i.e.,
// TFE_DeleteOp() is called before TFE_DeleteContext().
//
// Very similar to TF_OperationDescription with some differences:
// (1) TF_Output or TFE_TensorHandle* as arguments to TF_AddInput,
//     TF_AddInputList

    // on to one TFE_Op per device and just call TFE_ResetOp to avoid parsing
    // device names repeatedly.
    std::unique_ptr<TFE_Op, decltype(&TFE_DeleteOp)> const_op(
        TFE_NewOp(context, "Const", status), TFE_DeleteOp);
    if (TF_GetCode(status) != TF_OK) return nullptr;
    TFE_OpSetDevice(const_op.get(), underlying_devices_[device_index].c_str(),
                    status);