Open iMountTai opened 1 year ago
I had the same question. I understand I need to set "attention_pruning_method" : pruning_method
and "dense_pruning_method" : pruning_method:1d_alt
for hybrid pruning.
Is it that for hybrid-filled approach we train the model again with reinitializing the weights for heads which are not completely pruned and this is done separately after hybrid pruning? or Is there a argument I can set in the code for hybrid-filled pruning? @madlag @echarlaix
@madlag