Fix attention parity for GPT-2 (#8549)

Commit

4 years ago

Fix attention parity for GPT-2 (#8549) * Use persistent softmax to parity with huggingface * fix undirectional mask logic * add test