Add the possibility to freeze the decoder. Note that freezing the decoder will freeze the decoder embed_tokens layer that is by default tied to proj_out layers (not a layer of the decoder). This way, proj_out also gets frozen, it is thus necessary to unfreeze it.
other minor changes:
use set_seedfrom Accelerate that will also set the seed for numpy
Add the possibility to freeze the decoder. Note that freezing the decoder will freeze the decoder
embed_tokens
layer that is by default tied toproj_out
layers (not a layer of the decoder). This way,proj_out
also gets frozen, it is thus necessary to unfreeze it.other minor changes:
set_seed
from Accelerate that will also set the seed for numpy