Pytorch question about BiLSTM-CRF

pytorch tutorial have a bilstm-crf example。But, it isn’t used minibatch。
when i try to make a minibatch in it。I find that, CRF can’t be minibatch?
And, CRF need run in cpu? it will be so slowly!
aspect these,there are also some questiones below:

  1. how pytorch auto deal variable sequence length?padding a same length?but pytorch is dynamic right?
  2. I don’t konw why,but pytorch is slowly so much and the uses of gpu util
    is little.

may be all these question is caused by (batch and crf )
how can CRF read the batch and deal in gpu?:sob:

I also have the problem to fit the model for the GPU.
Did you fixed it?