Closed matteodelv closed 4 years ago
Unfortunately, I don't think we'll be able to get to this anytime soon. Maybe one of the other ALBERT implementations will have it?
https://github.com/kamalkraj/ALBERT-TF2.0 has support for multi-GPU pre-training
Is multi-GPU support expected to be implemented anytime soon? What are the conceptual changes that need to be done in order to make ALBERT train/fine-tune on multiple GPUs? Should
TPUEstimator
be replaced byEstimator
, together with a distributed strategy such asMirroredStrategy
? Thank you so much!