[fsmt] deal with -100 indices in decoder ids (#18592)

* [fsmt] deal with -100 indices in decoder ids

Fixes: https://github.com/huggingface/transformers/issues/17945

decoder ids get the default index -100, which breaks the model - like t5 and many other models add a fix to replace -100 with the correct pad index. 

For some reason this use case hasn't been used with this model until recently - so this issue was there since the beginning it seems.

Any suggestions to how to add a simple test here? or perhaps we have something similar already? user's script is quite massive.

* style

src/transformers/models/fsmt/modeling_fsmt.py

@@ -372,6 +372,10 @@ def _check_shapes(shape_1, shape2):
 
 def shift_tokens_right(input_ids, pad_token_id):
     """Shift input ids one token to the right, and wrap the last non pad token (usually <eos>)."""
+
+    # replace possible -100 values in labels by `pad_token_id`
+    input_ids.masked_fill_(input_ids == -100, pad_token_id)
+
     prev_output_tokens = input_ids.clone()
     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()