TFE_DeleteTensorHandle(h1_task0);
  TFE_DeleteTensorHandle(h0_task1);
  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);

  ASSERT_TRUE(TF_GetCode(status) == TF_OK) << TF_Message(status);
  TF_DeleteFunction(fn);

  for (bool async : {false, true, false}) {
    TFE_Executor* old_executor = TFE_ContextGetExecutorForThread(ctx);
    TFE_Executor* executor = TFE_NewExecutor(
        /*is_async=*/async, /*enable_streaming_enqueue=*/true,
        /*in_flight_nodes_limit=*/0);
    TFE_ContextSetExecutorForThread(ctx, 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_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_DeleteExecutor(executor);
  TF_DeleteStatus(status);
}

    TFE_Execute(matmul, &retvals[0], &num_retvals, status);
    CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
    if (state.iterations() >= state.max_iterations && async) {
      TFE_Executor* executor = TFE_ContextGetExecutorForThread(ctx);
      TFE_ExecutorWaitForAllPendingNodes(executor, status);
      ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
      TFE_DeleteExecutor(executor);
    }
  }
  TFE_DeleteOp(matmul);

  TF_DeleteTensor(t);
  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);

}

void TFE_ContextSetExecutorForThread(TFE_Context* ctx, TFE_Executor* executor) {
  tensorflow::unwrap(ctx)->SetExecutorForThread(executor->executor());
}

TFE_Executor* TFE_ContextGetExecutorForThread(TFE_Context* ctx) {
  return new TFE_Executor(&tensorflow::unwrap(ctx)->Executor());
}

void TFE_HostAddressSpace(TFE_Context* ctx, TF_Buffer* buf) {

                                                           TFE_Executor*);

// Returns the Executor for the current thread.
TF_CAPI_EXPORT extern TFE_Executor* TFE_ContextGetExecutorForThread(
    TFE_Context*);

// -----------------------------------------------------------------------------
// Dynamic cluster API.