hf_T5 runs fine in our nightly benchmarks, so there must be something different about your setup. Feel free to submit an issue with more details.
Export should work fine, but you it gives you an FX graph which can’t be torch.save()ed – so you aren’t using it in the intended way. Export is not designed to be a solution for model serialization. It gives you a graph intended to be passed to other frameworks and compiled.