what is wrong with my flops calculation of dm_nfnet_f0 model?

[serser]

import timm
import torch
from thop import profile

model = timm.create_model('dm_nfnet_f0')
x = torch.randn(1, 3, 224, 224)

flops, params = profile(model, inputs=(x,), verbose=False)

print('flops = %.4fM' % (flops / 1e6))
print("param size = %.4fMB" % (params / 1e6))

which gives output,

flops = 22.3761M
param size = 21.8461MB

It doesn't seem to match dm_nfnet_f0 - 83.464 results and NFNet-F0 as shown in https://arxiv.org/pdf/2102.06171v1.pdf