Created by: bingyanghuang

As Brian's patch to fix the multi-instances performance drop problem specially for rnn1 and chinese ner, it added two intermediate outputs for crf-decoding to decrease the memory allocate time. Therefore, it needs to modify the model as following figure: What we did to test this new patch is to retrain a new model for inference, but for your online deployment, retrain a model or add a pass for this "bug fix" issue needed to be discussed.