What happened? Command-A speeds for IK are higher than mainline, especially since they broke something recently to knock it back from 15t/s to 12t/s. I still see 15s in YALS which uses an older commit ...

I’m currently trying to run the inference_i2v.py script with the visual tokenizer set to "Cosmos-Tokenizer-DV4x8x8". However, I ran into an issue where VocabTranslation._vocab.video_tokens ends up ...