How can I convert lora for split qkv layer to concatenated one?

[Adenialzz]

I notice that in diffusers, linear projection is split implemented (attn.to_q, 3072*3072*3), while in origin flux repo, it is concatenated, (attn.qkv 3072*9216). I wonder how can I convert lora for diffusers format (split qkv) to concatenated one?

e.g. 3 lora for split qkv: (3072*16, 16*3072) * 3 ===> lora for concatenated qkv: (3072*r, r*9216).

What should I do? sum or concat?