ValueError warning in BartAdapterModel is unnecessary for sequence labeling task/generative task

[dyxohjl666]

Information

Model I am using (Bert, XLNet ...): BartAdapterModel

Language I am using the model on (English, Chinese ...): English

Adapter setup I am using (if any): pffeifer

The problem arises when using:

• the official example scripts: (give details below)

The tasks I am working on is:

• an official GLUE/SQUaD task: mlm

To reproduce

Steps to reproduce the behavior:

1. Follow the official run_mlm.py and use BartAdapterModel as the backbone.
2. Got " ValueError("All examples must have the same number of tokens.")" during training.

I checked the source code in bart/adatper_model.py and modeling_bart.py (in transformers) and found this in BartAdapterModel's forward() function:

            if len(torch.unique(eos_mask.sum(1))) > 1:
                raise ValueError("All examples must have the same number of <eos> tokens.")
            cls_representation = x[eos_mask, :].view(x.size(0), -1, x.size(-1))[:, -1, :]

which is the same with BartForSequenceClassification. I think it's not necessary for a sequence labeling task? When doing masked language model, all the input text are concatenated and then grouped by every 1024 tokens, and it's common to include different number of tokens accordingly.
This error exists even I add a seq2seq_lm_head.

Expected behavior

The mentioned ValueError should be applied only for classification tasks.