Replace torchvision.io.write_video with PyAV muxing

Fix skip_attention mask shape in vanilla CrossAttention

Add torchcodec, required by newer torchaudio for audio loading

Remove torch.compile calls (unsupported on ZeroGPU per docs)

Fix indentation of debug prints inserted for hang diagnosis

Add flushed debug prints around model loading stages to diagnose hang

Use vanilla/softmax attention instead of xformers (unsupported on Blackwell)

Use vanilla/softmax attention instead of xformers (unsupported on Blackwell)

Disable xformers memory-efficient attention in VAE (unsupported on Blackwell/sm_120 GPUs)

Increase gpu allocation

text

requirements

requirements

requirements

requirements

requirements

requirements

requirements

requirements

requirements

requirements

requirements

requirements

requirements

requirements

requirements

requirements

requirements

requirements

requirements

requirements

requirements

requirements

requirements

requirements

requirements

requirements

add checkpoint download

addapt for zerogpu

addapt for zerogpu

requirements

requirements

requirements

requirements

requirements

requirements

requirements

requirements

first commit

initial commit