Is there anyway to empty cache/memory after loading get-2 into our model?

I wonder if there is any strategy to empty cache after loading (Restoring) gpt2 parameters into our model? I am getting OOM error for batch_size and I wonder how if there is anyway to get around that when using massive pre-trained models?

No file, test, or entry point is named. Start by reproducing the OOM while restoring GPT-2 parameters with the reported batch size and inspect the TensorFlow loading path. Done would require a confirmed memory-release approach or a documented way to avoid the OOM for massive pretrained models.

Written by the indexing model from the issue text.