Training with GPU will always load data

I tried to train a retrosynthesis model using OpenNMT-py.
And using the CPU during training is successful before, but this would have to reduce the batch size a lot which would affect the results .
When I want to use GPU during training, the model will always load data. Although I tried reduce the batch_size.
Does anyone else have this problem?
I am new to open nmt. Please can anybody help me.