Does block_mask work during inference?

[ki-lw]

In the _recache_after_switch function, there is a process for generating block_mask. However, it seems that this block_mask is not actually used during inference in CausalWanModel.

Am I misunderstanding something here, or is there a part of the implementation I might have missed where block_mask comes into play?

[ki-lw]

When kv_cache is not None!

[zlenyk]

I also noticed that. For causal inference it always uses KV-cache, which means that block mask is essentially ignored and we are running bi-directional attention on recached frames. I'm not sure that's intended. Paper also isn't clear about what should happen.