Created by: suchenzang
Following https://github.com/facebookresearch/metaseq/pull/230 and continuing to break down https://github.com/facebookresearch/metaseq/pull/197.
Removed:
- unused
moe_disable_padding
arg - unused
from_pretrained
methods, since we currently depend onload_model_ensemble_and_task
fromcheckpoint_utils
(not great but saving that for another PR) - unused
hub_models
method - unused
return_all_hiddens
arg, and noencoder_states
being returned as a result -
EncoderDecoderModel
class