[Help]: Vevo2

Problem Overview

I have Vevo2 (models from HF RMSnow) working except the new Qwen AR model.

Steps Taken

1. Models (tokenizer, Qwen AR, FM, vocoder) are all working as expected, there are however random sounds, when switching back to the previous custom AR model this is fixed. Maybe you have the updated ar_model.py?

Expected Outcome

I need some help in the token mapping for the new model and implementation of the vocab.json since the input format is different, I suspect this is causing the problems.

Environment Information

• Operating System: Windows 11

• Python Version: 3.12

• Driver & CUDA Version: Cuda 13.

• Error Messages and Logs: None