// once (i.e., on the main function side) in running distributed functions.
// This test creates a cluster with two workers, create a variable on the
// second worker, and run a distributed function (VariableAddFunction) whose ops
// span the local and remote workers. If the graph optimization pass is executed
// on both the main function side and the component function side, an error will
// be thrown in the registered graph optimization pass.

  ExpectScalarEq<float>(result_components[0].get(), 3.);
  ExpectScalarEq<float>(result_components[1].get(), 3.);
}

TEST(PARALLEL_DEVICE, TestCollectiveSync) { TestCollective(/*async=*/false); }

// Note that ops on the parallel device currently don't execute
// asynchronously. The test is just that we don't get deadlocks.
TEST(PARALLEL_DEVICE, TestCollectiveAsync) { TestCollective(/*async=*/true); }

    typedef tensorflow::protobuf::RepeatedPtrField<tensorflow::OpDef> OpDefs;
    const OpDefs& ops = op_list.op();
    bool found = std::find_if(ops.begin(), ops.end(),
                              [](const tensorflow::OpDef& op_def) {
                                return op_def.name() == "TestCApi";
                              }) != ops.end();
    EXPECT_TRUE(found);
    TF_DeleteBuffer(op_list_buffer);
  }
}

      ASSERT_EQ(expected[i], data[i]) << "Different data at index " << i;
    }
  }

  std::vector<TF_Output> ToOutput(const std::vector<TF_Operation*> ops) {
    std::vector<TF_Output> out;
    for (auto op : ops) {
      out.push_back({op, 0});
    }
    return out;
  }

  void Define(int num_opers, const std::vector<TF_Operation*>& opers,

  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  if (forward_input) {
    TFE_DeleteTensorHandle(n);
  }

  int num_retvals = 1;
  if (async) {
    // Enqueue dummy ops so we backlog async execution & actually test async.
    // This is usually unnecessary, but we've experienced the occasional test
    // failure when testing async mode with no explicit forwarding.