Issue
I am trying to use transformer model in colab tpu and kaggle tpu. But under same package, same model can’t get built. When I ran the code on colab-tpu, the model successfully build but not on kaggle-tpu. It throws
UnimplementedError: The Conv2D op currently does not support grouped convolutions on the CPU
which is a known issue but if it’s the main cause it should also happen on colab-tpu.
Code to reproduce