Could someone explain to me the operation of efficient attention？

[CHDyshli]

def relu_linear_att(self, qkv: torch.Tensor) -> torch.Tensor:
        B, _, H, W = list(qkv.size())
        if qkv.dtype == torch.float16:
            qkv = qkv.float()
        qkv = torch.reshape(
            qkv,
            (
                B,
                -1,
                3 * self.dim,
                H * W,
            ),
        )
        q, k, v = (
            qkv[:, :, 0 : self.dim],
            qkv[:, :, self.dim : 2 * self.dim],
            qkv[:, :, 2 * self.dim :],
        )
        # lightweight linear attention
        q = self.kernel_func(q)
        k = self.kernel_func(k)

        # linear matmul
        trans_k = k.transpose(-1, -2)

        v = F.pad(v, (0, 0, 0, 1), mode="constant", value=1)
        vk = torch.matmul(v, trans_k)
        out = torch.matmul(vk, q)
        out = out[:, :, :-1] / (out[:, :, -1:] + self.eps)

        out = torch.reshape(out, (B, -1, H, W))
        return out

In the above segment, could someone explain why we need the pad operation?

        v = F.pad(v, (0, 0, 0, 1), mode="constant", value=1)
        vk = torch.matmul(v, trans_k)
        out = torch.matmul(vk, q)
        out = out[:, :, :-1] / (out[:, :, -1:] + self.eps)