Is it Possible to Refactor SeqBoat using HuggingFace without Performance Drop? about seqboat HOT 3 OPEN

Thanks for your interest in our work!
Yes, it is definitely possible to refactor SeqBoat using HuggingFace. The custom FairSeq modules used in this repo are totally based on pytorch, so there should be no technical barrier in refactoring. The tricky part can be how to migrate the design of KV cache from FairSeq style to HF style, which I haven't look into it closely. I am personally quite interested in doing such migration for greater benefits of the research community and also for potentially scaling up SeqBoat in the future, so maybe we can colaborate together on providing a standalone version of SeqBoat in this repo. What do you think?

from seqboat.

Hi @KatarinaYuan, I have created a standalone version of the SeqBoat layer here. It does not support KV caching for now. Please feel free to test it with HuggingFace and see if it works well in your use case : )

from seqboat.

Hi hi, thank you for the quick response and support! I'll try it out first and play around with it for a little bit. I haven't made up my mind for the refactor though because of tight time schedule for at least the recent month. (But I'm very motivated for the work!)

from seqboat.