Use DataCollatorForSeq2Seq in run_summarization in all cases (#10856)

Co-authored-by: Eliza <eliza@habanero.tiger.com.pl>

examples/seq2seq/run_summarization.py

@@ -38,7 +38,6 @@ from transformers import (
     HfArgumentParser,
     Seq2SeqTrainer,
     Seq2SeqTrainingArguments,
-    default_data_collator,
     set_seed,
 )
 from transformers.file_utils import is_offline_mode
@@ -466,9 +465,6 @@ def main():
 
     # Data collator
     label_pad_token_id = -100 if data_args.ignore_pad_token_for_loss else tokenizer.pad_token_id
-    if data_args.pad_to_max_length:
-        data_collator = default_data_collator
-    else:
     data_collator = DataCollatorForSeq2Seq(
         tokenizer,
         model=model,