important thing to declare is the target_spcae_id in different T2T implementation is different, default 4 in 1.0.14, but 0 in 1.6.5
revise the conver model script, now we use conver_t2t_to_fairseq.py to convert the t2t checkpoint into fairseq.pt
Build FairseqDecoder with T2T training checkpoint, I support both the t2t-1.0.14 version and 1.6.5 version