Lstm inference generate different results in gpu vs cpu

1

I have a lstm model, its inference in cpu different from its inference in gpu, wonder why

2

Depending on the differences you are seeing you might observe the expected errors caused by the limited floating point precision and a different order of operations.