Fix Seq2SeqTrainer crash when BatchEncoding data is None

[iohub]

What does this PR do?

Fix Seq2SeqTrainer crash when evaluate dataset's labels is None.

• Reproduction example:

finetune GLM4 model with customize DataCollator and customize Seq2SeqTrainer according to finetune script : https://github.com/THUDM/GLM-4/blob/main/finetune_demo/finetune.py

class DataCollatorForSeq2Seq(_DataCollatorForSeq2Seq):
    def __call__(self, features, return_tensors=None):
        output_ids = ([feature['output_ids'] for feature in features] if 'output_ids' in features[0].keys() else None)
        if output_ids is not None:
            max_output_length = max(len(out) for out in output_ids)
            if self.pad_to_multiple_of is not None:
                max_output_length = (
                        (
                                max_output_length + self.pad_to_multiple_of - 1) //
                        self.pad_to_multiple_of * self.pad_to_multiple_of
                )
            for feature in features:
                remainder = [self.tokenizer.pad_token_id] * (
                        max_output_length - len(feature['output_ids'])
                )
                if isinstance(feature['output_ids'], list):
                    feature['output_ids'] = feature['output_ids'] + remainder
                else:
                    feature['output_ids'] = np.concatenate(
                        [feature['output_ids'], remainder]
                    ).astype(np.int64)
        return super().__call__(features, return_tensors)

class Seq2SeqTrainer(_Seq2SeqTrainer):
    def prediction_step(
            self,
            model: nn.Module,
            inputs: dict[str, Any],
            prediction_loss_only: bool,
            ignore_keys=None,
            **gen_kwargs,
    ) -> tuple[Optional[float], Optional[torch.Tensor], Optional[torch.Tensor]]:
        if self.args.predict_with_generate:
            output_ids = inputs.pop('output_ids')
        input_ids = inputs['input_ids']
        loss, generated_tokens, labels = super().prediction_step(
            model, inputs, prediction_loss_only, ignore_keys, **gen_kwargs
        )
        generated_tokens = generated_tokens[:, input_ids.size()[1]:]
        labels = output_ids
        return loss, generated_tokens, labels

• Root cause

https://github.com/iohub/transformers/blob/b7672826cad31e30319487af876e608d8af7d37b/src/transformers/data/data_collator.py#L664

        # reintroduce side effects via tokenizer that return respective datatypes for the `return_tensors` argument
        if batch.get("labels", None) is not None:
            if return_tensors == "pt":
                import torch

                batch["labels"] = torch.tensor(batch["labels"], dtype=torch.int64)
            elif return_tensors == "tf":
                import tensorflow as tf

                batch["labels"] = tf.constant(batch["labels"], dtype=tf.int64)
            else:
                batch["labels"] = np.array(batch["labels"], dtype=np.int64)
        else:
            batch["labels"] = None

As the code showed, batch["labels"] = None may cause None value v in BatchEncoding.

    self.data = {k: v.to(device=device) for k, v in self.data.items()}

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

@amyeroberts

[amyeroberts]

Thanks for adding!

It doesn't really make sense for something to be None in the BatchEncoding: this is a class we use to bundle together the outputs of the tokenizer and easily convert to different formats. However, as this is a relatively simple change and it comes from one of our examples and data collators I think it should be OK

[amyeroberts]

@iohub Could you rebase on main to include upstream changes? This should resolve the currently failing CI tests

[iohub]

@amyeroberts Thanks for approval. this branch already rebase on upstream main, and CI tests are success.