I took the example "Out-of-Core Training on MNIST" from
https://reference.wolfram.com/language/tutorial/NeuralNetworksLargeDatasets.html
The only code element I have modified is
results =
NetTrain[lenet, trainingDataFiles, All,
ValidationSet -> testDataFiles, MaxTrainingRounds -> 3,
TargetDevice -> "GPU"]
So, I added 'TargetDevice -> "GPU" ' which is typically not a problem. But in this example I get the following message:

The Stack Trace for NetTrain::interr2

The issue does not occur with "CPU" as TargetDevice. How to solve the problem?