After the weight refactor, some tie weight might not have gone as expected. If the lm head was removed, then we are kinda fucked when using from_pretrained.

[ArthurZucker]

What does this PR do?

Tie weights

from transformers import AutoModel, MvpModel, MvpForConditionalGeneration
from transformers import AutoformerModel

MvpForConditionalGeneration._tied_weights_keys = {"model.shared.weight":"lm_head.weight", "model.decoder.embed_tokens.weight": "model.shared.weight", "model.encoder.embed_tokens.weight":"lm_head.weight"}
model = MvpForConditionalGeneration.from_pretrained("RUCAIBox/mvp")

MvpModel._tied_weights_keys = {"model.shared.weight":"lm_head.weight", "model.decoder.embed_tokens.weight": "model.shared.weight", "model.encoder.embed_tokens.weight":"lm_head.weight"}
model = MvpForConditionalGeneration.from_pretrained("RUCAIBox/mvp")

There are a few issues, I think we are gonna put the tie weight keys into a weight converter -> properly copies them even if the source key is unexpected?

model = UMT5EncoderModel.from_pretrained('google/umt5-xxl')