Question about when fuse() is triggered in Conv2d_BN / RepDW / RepDW_Axias

[PinkSlime]

I would like to thank you for proposing such an efficient Mamba-based method, which has been very inspiring for my research. While reading through your code, I noticed that the classes Conv2d_BN, RepDW_Axias, and RepDW all implement a fuse() function for structural re-parameterization.
However, I could not find where this fuse() function is actually invoked during training or inference. Could you please tell me where and when to call the fuse() function to simplify the structure?

[xwmaxwma]

Thank you for your recognition of this work. You can refer to the application of fuse() here:

TinyViM/speed_gpu.py

Line 48 in cff6452

utils.replace_batchnorm(model)