TFE_DeleteTensorHandle(h1_task1);
  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_DeleteExecutor(executor);
  TFE_DeleteContext(ctx);

  TF_DeleteStatus(status);

  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_DeleteExecutor(executor);
  TF_DeleteStatus(status);
}

    ASSERT_TRUE(TF_GetCode(status) == TF_OK) << TF_Message(status);
    EXPECT_EQ(*reinterpret_cast<tensorflow::int32*>(TF_TensorData(r)), 42);
    TFE_ContextSetExecutorForThread(ctx, old_executor);
    TFE_ExecutorWaitForAllPendingNodes(executor, status);
    ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
    TFE_DeleteExecutor(executor);
    TFE_DeleteExecutor(old_executor);
    TFE_DeleteTensorHandle(h);

  }
  for (auto* h : handles_task2) {
    TFE_DeleteTensorHandle(h);
  }

  TFE_DeleteOp(matmul);

  TFE_Executor* executor = TFE_ContextGetExecutorForThread(ctx);
  TFE_ExecutorWaitForAllPendingNodes(executor, status);
  ASSERT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);
  TFE_DeleteExecutor(executor);
  if (func) {
    TFE_ContextRemoveFunction(ctx, "MatMulFunction", status);
  }

    TFE_ExecutorWaitForAllPendingNodes(executor, status);
    EXPECT_NE(TF_OK, TF_GetCode(status));
    TF_SetStatus(status, TF_OK, "");
    retvals[0] = nullptr;
    TFE_Execute(matmul2, &retvals[0], &num_retvals, status);
    EXPECT_NE(TF_OK, TF_GetCode(status));
    TFE_ExecutorClearError(executor);
    TFE_ExecutorWaitForAllPendingNodes(executor, status);

    execution_state_ = ExecutionState::kShuttingDown;
  }
  start_execute_.notify_one();
}

void DeviceThread::AsyncWait(TF_Status* status) {
  tensorflow::mutex_lock l(execution_mutex_);
  TFE_ExecutorWaitForAllPendingNodes(executor_.get(), status);
  TFE_ExecutorClearError(executor_.get());
}

void DeviceThread::Run() {
  while (true) {
    {
      tensorflow::mutex_lock l(execution_mutex_);

TF_CAPI_EXPORT extern TFE_Executor* TFE_NewExecutor(
    bool is_async, bool enable_streaming_enqueue, int in_flight_nodes_limit);

// Deletes the eager Executor without waiting for enqueued nodes. Please call
// TFE_ExecutorWaitForAllPendingNodes before calling this API if you want to
// make sure all nodes are finished.
TF_CAPI_EXPORT extern void TFE_DeleteExecutor(TFE_Executor*);

// Returns true if the executor is in async mode.

  EXPECT_EQ(sum, 6.0);

  TFE_DeleteTensorHandle(h0);
  TFE_DeleteTensorHandle(h1);
  TFE_DeleteTensorHandle(h2);

  TFE_Executor* executor = TFE_ContextGetExecutorForThread(ctx);
  TFE_ExecutorWaitForAllPendingNodes(executor, status);
  ASSERT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);
  TFE_DeleteExecutor(executor);
  TFE_ContextRemoveFunction(ctx, "AddVariablesFunction", status);
  TFE_DeleteContext(ctx);

}

void TFE_DeleteExecutor(TFE_Executor* executor) { delete executor; }

bool TFE_ExecutorIsAsync(TFE_Executor* executor) {
  return executor->executor()->Async();
}

void TFE_ExecutorWaitForAllPendingNodes(TFE_Executor* executor,
                                        TF_Status* status) {
  status->status = executor->executor()->WaitForAllPendingNodes();
}