fix bart loss masking (#9131)

src/transformers/models/bart/modeling_bart.py

@@ -70,6 +70,11 @@ def shift_tokens_right(input_ids: torch.Tensor, pad_token_id: int):
     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]
+
+    assert pad_token_id is not None, "self.model.config.pad_token_id has to be defined."
+    # replace possible -100 values in labels by `pad_token_id`
+    prev_output_tokens.masked_fill_(prev_output_tokens == -100, pad_token_id)
+
     return prev_output_tokens