Model Output Corrupted After Recent Update to ultravox_model.py
#3
by
frznkshn
- opened
Hi, I noticed that after the recent updates to the ultravox_model.py script in the past few days, running the following command:
pipe = transformers.pipeline(model='fixie-ai/ultravox-v0_5-llama-3_2-1b', trust_remote_code=True)
produces output that seems to be a mix of corrupted or scrambled tokens, without any meaningful interpretation. However, if I explicitly set the revision to a previous commit like this:
pipe = pipeline(model='fixie-ai/ultravox-v0_5-llama-3_2-1b',
trust_remote_code=True,
revision='06c7f4eb509f60ce5f03563c0514756ebf357d39')
everything works as expected. Could you please look into this issue?
Thanks for reporting and sorry for the late reply. It turns out the fix worked with the new version of transformers but not the old one. It took me some time to figure out the right way to make it work for both.
Can you try again? The issue should be fixed now.