How to Understand "Destruction of Locality"?

[vitrun]

Thank you for your outstanding work! With respect to the core contribution of the paper, which focuses on 'preserving structural integrity,' could you please clarify the following question?"

The image gives the impression that all kernels of a 2D convolution are forcibly pieced together into a two-dimensional matrix, leading to what's referred to as the "destruction of locality."

However, looking at the implementation of peft definition of lora weights:

        self.lora_A[adapter_name] = nn.Conv2d(self.in_features, r, kernel_size, stride, padding, bias=False)
        self.lora_B[adapter_name] = nn.Conv2d(r, self.out_features, (1, 1), (1, 1), bias=False)

and get_delta_weight to merge, the 2D convolution is achieved using two low-rank convolutional kernels, without forcibly reshaping multiple kernels into a matrix. Therefore, doesn't this mean there's no issue of "adjacent elements within the kernel being separated in the reshaped matrix, disrupting the spatial locality inherent in the original convolutional space"?