Reproduction issue of stage 2.

[TumCCC]

Hi, authors of FlashVSR. Thanks for contributing so excellent work for the community.

But I still meet some problems when using it on some extremely bad cases, So I would like to fine-tune it.

And thus I am trying to reproduce the training code, but I got some problems during reproducing stage 2, and I hope you can share some details about that:

1. LR_proj module is converted into Causal one for only inference or both inference and training?
2. generate_draft_block_mask in wan_video_dit.py is only for inference? As it can be derivable. Do you use other mask strategy during your training?