search

LoicGrobol / zeldarose

Train transformer-based models.
https://zeldarose.readthedocs.io
Other
28 stars 3 forks source link

Performance regression with transformers 0.41 #85

Open LoicGrobol opened 3 weeks ago

LoicGrobol commented 3 weeks ago

Observed in training Breton-French models. Could be related to something on our side or in one of the many dependences. Regression testing per-version and then bisection per commit will probably be needed to pinpoint the source.

LoicGrobol commented 2 weeks ago

After intensive bisection, my conclusion is that the problem is not (only) on our side, but comes from a incompatibility that occurs with transformers 0.41 but not with earlier version. Pinpointing the exact source of error will be much more time-consuming however so I'll just pin the upper bound of that dependency and leave that issue open as a todo for the future.

LoicGrobol commented 2 weeks ago

Ooooh, possibly https://github.com/huggingface/transformers/issues/31348