Jiatong (Julius) Han

They look fairly reasonable as compared with the demonstrations given by the team. But anyone having suggestions is welcome to share here.

What's your command @bluenevus ?

Could @zhengzangw please take a look?

Can @lbwang2006 also provide more information on what you meant by 'needs to 10000 step' with QK norm enabled?

Sure. Before we get to this, could you please let us know if the detail on the error you met and the dockerfile you used?

Thanks. Would you please help me pinpoint where you identified this issue? I saw [this](https://github.com/hpcaitech/Open-Sora/blob/278cd759e58aff49c4349e102584e7d1c15045bf/opensora/schedulers/iddpm/gaussian_diffusion.py#L738) but failed to locate where channel selection happened.

Check this out: https://github.com/hpcaitech/Open-Sora/blob/476b6dc79720e5d9ddfb3cd589680b2308871926/docs/acceleration.md

I suspect the `num_frames` or `resolution` of videos in your training corpus **exceeds** the one specified in your config, hence leading to the bucket 'overflow'. It would be great if...

Is this what you are finding? https://github.com/hpcaitech/Open-Sora/blob/476b6dc79720e5d9ddfb3cd589680b2308871926/opensora/models/layers/blocks.py#L79

My intuition is that it relates to baseline comparisons or competition-based evaluations.