Issue of NaN values when running predict on CPU
#6
by himj98 - opened
When forcing the model to run on cpu on a MacBook M2, we get NaN values when running the predict function. Note that when we use the default device = 'mps' we get the required response correctly. Also this happens only for the ms-marco-miniLM-L-6-v2, note that for ms-marco-miniLM-L-12-v2 we get the right responses too.
i second that
Same here, same MacBook Air M2
same problem here. to be 100% sure it couldn't be avoidable i tried with both sentence-transformers and transformers but it doesn't work. Anyone found a fix?