Fix Bart Shift (#9135)

* correct mistake in order

* fix tensor copy

* clone tensor correctly

src/transformers/models/bart/modeling_bart.py

@@ -72,9 +72,10 @@ def shift_tokens_right(input_ids: torch.Tensor, pad_token_id: int):
     # replace possible -100 values in labels by `pad_token_id`
     prev_output_tokens.masked_fill_(prev_output_tokens == -100, pad_token_id)
 
-    index_of_eos = (input_ids.ne(pad_token_id).sum(dim=1) - 1).unsqueeze(-1)
-    prev_output_tokens[:, 0] = input_ids.gather(1, index_of_eos).squeeze()
-    prev_output_tokens[:, 1:] = input_ids[:, :-1]
+    index_of_eos = (prev_output_tokens.ne(pad_token_id).sum(dim=1) - 1).unsqueeze(-1)
+    decoder_start_tokens = prev_output_tokens.gather(1, index_of_eos).squeeze()
+    prev_output_tokens[:, 1:] = prev_output_tokens[:, :-1].clone()
+    prev_output_tokens[:, 0] = decoder_start_tokens
 
     return prev_output_tokens