This issue seems to be a bug in the serverless environment v4. The serverless instance does not clean up no-use models timely, which leads to insufficient space for new model training. To force resetting the serveless instance, neither the "New Session" nor the "Terminate" functions work perfectly. We have to reconfigure the serverless environment by changing the memory, either from 16GB to 32GB or from 32GB to 16GB. This forces the serveless instance to be fully reset, then rerunning the model training won't generate the same error again. But this is just a temporary solution, because after several runs of model training, the same error will occur again. Hope the Serverless team can fix this bug ASAP.