[examples (seq2seq)] fix preparing decoder_input_ids for T5 (#5994)

examples/seq2seq/finetune.py

@@ -14,7 +14,7 @@ import torch
 from torch.utils.data import DataLoader
 
 from lightning_base import BaseTransformer, add_generic_args, generic_train
-from transformers import MBartTokenizer, get_linear_schedule_with_warmup
+from transformers import MBartTokenizer, T5ForConditionalGeneration, get_linear_schedule_with_warmup
 
 
 try:
@@ -131,8 +131,14 @@ class SummarizationModule(BaseTransformer):
     def _step(self, batch: dict) -> Tuple:
         pad_token_id = self.tokenizer.pad_token_id
         source_ids, source_mask, target_ids = batch["input_ids"], batch["attention_mask"], batch["decoder_input_ids"]
-        decoder_input_ids = target_ids[:, :-1].contiguous()  # Why this line?
-        lm_labels = target_ids[:, 1:].clone()  # why clone?
+
+        if isinstance(self.model, T5ForConditionalGeneration):
+            decoder_input_ids = self.model._shift_right(target_ids)
+            lm_labels = target_ids
+        else:
+            decoder_input_ids = target_ids[:, :-1].contiguous()  # Why this line?
+            lm_labels = target_ids[:, 1:].clone()  # why clone?
+
         outputs = self(source_ids, attention_mask=source_mask, decoder_input_ids=decoder_input_ids, use_cache=False)
 
         if self.hparams.label_smoothing == 0: