Pytorch version problem

[fenss]

Hi, thanks for this work, it helps me a lot.. I'm using Pytorch 1.4-1.5 previously. Everything runs fine. But when I switch to Pytorch 1.7, it returns RuntimeError with:

'lengths' argument should be a 1D CPU int64 tensor, but got 1D cuda:0 Long tensor

Here's full error stack:

File "/my_proj/utils/supar/parsers/biaffine_dependency.py", line 125, in predict return super().predict(Config().update(locals())) File "/my_proj/utils/supar/parsers/parser.py", line 135, in predict preds = self._predict(dataset.loader) File "/opt/conda/lib/python3.8/site-packages/torch/autograd/grad_mode.py", line 26, in decorate_context return func(*args, *kwargs) File "/my_proj/utils/supar/parsers/biaffine_dependency.py", line 187, in _predict s_arc, s_rel = self.model(words, feats) File "/opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py", line 727, in _call_impl result = self.forward(input, kwargs) File "/my_proj/utils/supar/models/dependency.py", line 199, in forward feat_embed = self.feat_embed(feats) File "/opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py", line 727, in _call_impl result = self.forward(*input, **kwargs) File "/my_proj/utils/supar/modules/char_lstm.py", line 66, in forward x = pack_padded_sequence(x, lens[char_mask], True, False) File "/opt/conda/lib/python3.8/site-packages/torch/nn/utils/rnn.py", line 244, in pack_padded_sequence _VF._pack_padded_sequence(input, lengths, batch_first)

Is there any way to fix it?

[yzhangcs]

@fenss I think it's a bug (pytorch/pytorch/issues/47637) of pytorch. Please use pytorch<1.7 right now, or move the tensor lens to cpu as a temporary solution.

[fenss]

Ok, thank you!