Allennlp train uses all cpu resources (when trained with gpu)

AllenNLP version: 0.9.0

When my model is trained with GPU, I can see from the top command that the program uses up all the CPU resources like 3000%.
Is it by default or anything I need to set?
(My model is revised from an LSTM-CRF model)

Similarly, I directly run the example ner.jsonnet from allennlp, it also took 1600% CPU resources: