Inconsistent inference speed in different runs using osnet_ain_x1_0

[Ahmad-Hammoudeh]

Thanks for your great work, it helped me a lot!

I'm facing a problem when passing a batch of size~32 persons crops to FeatureExtractor object with osnet_ain_x1_0 model. It takes too long to finish (about 0.1 seconds in worst case and 0.05 in best case). In addition to the problem of inconsistent inference speed in different runs.
python version = 3.11.4
pytorch version = 2.0.1
cuda version = release 11.2, V11.2.67
GPU = GTX 1060

what could be the problem?

Thanks in advance.