Abstract: Network slicing latency is often considered as a key metric for comparing different network slicing resource allocation schemes. However, deploying different slicing schemes and measuring delay in real networks incurs prohibitively high costs. So we propose a model based on graph neural network (GNN) and multi-head attention mechanism to predict the network slicing latency. It can model the real network for machine learning, learn the traffic state of the nodes in the network, and obtain the network slicing latency by using the improved GNN algorithm with multi-head attention mechanism. The results show that the model can adapt to dynamic scenarios where the number of slices and nodes changes, and the prediction results have high accuracy, which proves that this model can be well applied to the monitoring and maintenance of operational networks.