[Feature] Causal attention

[betterze]

Motivation

Does the method support causal attention? Although the code includes a causal flag, it appears to be unused.

https://github.com/hao-ai-lab/FastVideo/blob/main/fastvideo/attention/backends/video_sparse_attn.py#L206

Related resources

No response

[Edenzzzz]

No it doesn't

[betterze]

Here is an example of adding causal mask to wan https://github.com/tianweiy/CausVid/blob/master/causvid/models/wan/causal_model.py#L414.

Is there possible to combine block causal mask with sparse attention? How can I add the block causal mask?

[cucNaifuXue]

Same question

[SolitaryThinker]

We've added support for causal attention based on self-forcing: https://github.com/hao-ai-lab/FastVideo/pull/766