Add support for EncoderDecoderModel

[alvations]

Feature request

There's already support for marian and various LLMs. But sometimes users create their own generic EncoderDecoderModel, e.g.

from transformers import EncoderDecoderModel
from optimum.onnxruntime import ORTModelForSeq2SeqLM

model = EncoderDecoderModel.from_encoder_decoder_pretrained("bert-base-multilingual-cased", "bert-base-multilingual-cased") 

model.save_pretrained("model_dir")

# Should be able to load this, but isn't supported yet.
ort_model = ORTModelForSeq2SeqLM.from_pretrained("model_dir", from_transformers=True)

Motivation

The EncoderDecoderModel is generic enough to cover quite a lot of use-cases but this is might be hard too since it can most probably only cover EncoderDecoder of ORT supported LLMs

Your contribution

Maybe, if there's some guidance on how to do so.

[mht-sharma]

Fixed in #851