Is causal attention actually used in CausalDiT?

[Dingrui-Wang]

Thank you for your amazing work! @seonghyeonye @zchuning

In you code regarding CausalWanSelfAttention, the related implementation is here

I am confused about the _process_clean_image_only function which will only be called if kv_cache is None, but during inference, kv_cache won't be none after initial timestep, so in this case, the normal self.attn instead of self.causal_attn will be used

Is this designed on purpose? If there is no issue about this, then the name CausalWanSelfAttention can be a bit confusing since no causal attention is really used during inference.

Thank you very much in advance!