[API][Rename] Rename T5NMTInference to T5Inference

[sxjscience]

Description

Incremental decoding for T5 has been supported in https://github.com/dmlc/gluon-nlp/pull/1498. However, the inference model is called T5NMTInference and we may consider to revise it to T5Inference or T5Seq2SeqInference.

What's your opinion on the naming strategy here? @yongyi-wu @szha @leezu @szhengac @barry-jin

[szha]

It seems to be kept consistent with TransformerNMTInference. I'd suggest to keep the naming convention consistent for easier discovery, so any change should happen to both classes. In terms of naming, I think Seq2Seq seems better and more general than NMT.

[sxjscience]

May be we can keep the name for now and later on add aliases.

[yongyi-wu]

@szha was right— I kept it consistent with TransformerNMTInference, so it would be great if we rename them altogether. Personally I prefer T5Seq2seq as it is shorter and more informative.

[sxjscience]

I will leave the issue open and we can change the name / add alias later.

[yongyi-wu]

Hi all, in #1504 TransformerNMTInference and T5NMTInference have been renamed to TransformerInference and T5Inference, respectively. After discussing with @sxjscience, we believe it is critical to highlight these two models are for inference only, so we keep the name short and straightforward. With that, TransformerNMTInference and T5NMTInference remain in #1504, simply inheriting the renamed models. We may need to remove them in the future.