Hyperparameter Tuning QNLI Colab Example using RoBERTa "RuntimeError('CUDA out of memory..."

I started from 10 trials and started lower the number every time I get the error. Can someone help with this error? Would it work if I had two GPUs working?

It seems no matter the number of trials, it crashes anyway at the last one.

Is there a way to free memory after each run?