FP16 / FP32 get different sentence vector

[NickChiang1215]

I using same model same input sentence but get different sentence vector by setting FP16 / FP32

[houshangwu]

I also had the same problem

[numb3r3]

@houshangwu Thanks for pointing it out. Some layers/operators result in inconsistent output on FP16. To fix this problem, the mix-precision approach could work. Please keep an eye on our progress, we are working on mix-precision inference support.

[houshangwu]

@houshangwu感谢您指出。某些层/运算符会导致 FP16 上的输出不一致。为了解决这个问题，混合精度方法可以工作。请密切关注我们的进展，我们正在努力支持混合精度推理。

Thank you for your reply. I will continue to follow your progress.