Accuracy calculation failure

[sia-cerebras]

I see that here in evaluate.py accuracy is calculated in two different ways. And there is an assert statements to make sure they match.
In my tests, these assertion sometimes fails. Why is this assertion there? and what does it failure signify?

Thank you,