OOM Error when trying to save t5-large using save_strategy="epoch"

I am trying to finetune t5-large for summarization on xsum. I followed all the steps like:

  • max_input_length=512,
  • max_target_length=64,
  • batch_size=2,
  • AdaFactor instead of AdamW,
  • XLA_USE_BF16=true,
  • initialized all models and datasets in the global scope instead of _mp_fn.

I want to know what I am doing wrong. I am using Kaggle and 16gb of RAM should be sufficient for this according to T5 Finetuning tips.
Please guide me

@valhalla @sgugger @sshleifer