Alignment heads?

#1
by ostegm - opened

When converting, I see the v2 model has specific alignment heads. Did those need to be changed in v3?

good catch! i just copied them from over from generation_config.json

Im not super familiar with the ctranslate2 code base, but it seems the alignment heads also need to be set in ctranslate2? without adding alignment heads in the conversion script, do you expect the model to perform as expected?

It looks like a PR is up on the ctranslate2 codebase:
https://github.com/OpenNMT/CTranslate2/pull/1530

oh yeah i am following that PR, looks like it just merged. also need the is_multilingual fix from the pr (which breaks language detection in this model). trying to update ct2 & run a new conversion

fwiw i didnt really test it much, but noticed timestamps were more accurate after updating config.json with the alignment heads post-conversion

Sign up or log in to comment