TPU Out of memory in Pix2Struct ForConditionalGeneration model

I am trying to train Pix2StructForConditionalGeneration model on google colab TPU but I get memory error while training for large patch size and large output token length. Is it possible to parallelize the model across the cores of 1 TPU device? (I am just testing on colab but I have access to google TPU research cloud)