Some models using multi head attention from onnx are no longer fusing attention properly after the attention refactor was merged in.