TFE_DeleteTensorHandle(input2);
  TFE_DeleteTensorHandle(identity_ret[0]);
  TFE_DeleteTensorHandle(identity_ret[1]);

  TFE_DeleteOp(cloned_identity);
  TFE_DeleteOp(original_identity);
  TFE_DeleteOp(original_var_op);
  TFE_DeleteOp(cloned);
  TF_DeleteStatus(status);
  TFE_DeleteContext(ctx);
}

TEST(CAPI, ShareVariableAcrossContextsWorks) {

  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);

                    /*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.

  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]);

    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_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);

    // 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);

    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

  status->status = new_op->Reset(op_or_function_name, nullptr);
  if (!status->status.ok()) {
    new_op->Release();
    new_op = nullptr;
  }
  return tensorflow::wrap(new_op);
}

void TFE_DeleteOp(TFE_Op* op) {
  if (op == nullptr) {
    return;
  }

  tensorflow::unwrap(op)->Release();
}

const char* TFE_OpGetName(const TFE_Op* op, TF_Status* status) {

#include "tensorflow/c/tf_status.h"
#include "tensorflow/c/tf_status_helper.h"

namespace tensorflow {
namespace parallel_device {
namespace {

class OpDeleter {
 public:
  void operator()(TFE_Op* to_delete) const { TFE_DeleteOp(to_delete); }
};

using OpPtr = std::unique_ptr<TFE_Op, OpDeleter>;

using MaybeParallelTensorOwned =
    absl::variant<std::unique_ptr<ParallelTensor>, TensorHandlePtr>;

using MaybeParallelTensorUnowned =