The different parameter size of mobilenetv3 in 2.7.0 from nightly

Hello, I was training a model using Mobilenet V3 small.
The number of parameters of MobileNetv3Small(alpha=1) in tf=2.7.0 is 1.5M, while it is 0.9M in tf-nightly
The same code was used to load the MobileNetV3Small model.
base_model =keras.applications.MobileNetV3Small(input_shape=(384, 512, 3),include_top=False,weights='imagenet',) base_model.summary()

I saw that some layers were deleted in tf-nightly. What is that? Which one is the correct one?

Some layers was moved with this commit:

So the weights hash changed