[Seq2Seq] Fix a couple of bugs and clean examples (#7474)

* clean T5

* fix t5 tests

* fix index typo

* fix tf common test

* fix examples

* change positional ordering for Bart and FSTM

* add signature test

* clean docs and add tests

* add docs to encoder decoder

* clean docs

* correct two doc strings

* remove sig test for TF Elektra & Funnel

* fix tf t5 slow tests

* fix input_ids to inputs in tf

* Update src/transformers/modeling_bart.py

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/modeling_bart.py

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* implement lysandre results

* make style

* fix encoder decoder typo

* fix tf slow tests

* fix slow tests

* renaming

* remove unused input

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

src/transformers/modeling_utils.py

@@ -237,8 +237,15 @@ class ModuleUtilsMixin:
                 batch_size, seq_length = input_shape
                 seq_ids = torch.arange(seq_length, device=device)
                 causal_mask = seq_ids[None, None, :].repeat(batch_size, seq_length, 1) <= seq_ids[None, :, None]
+                # in case past_key_values are used we need to add a prefix ones mask to the causal mask
+                if causal_mask.shape[1] < attention_mask.shape[1]:
+                    prefix_seq_len = attention_mask.shape[1] - causal_mask.shape[1]
+                    causal_mask = torch.cat(
+                        [torch.ones((batch_size, seq_length, prefix_seq_len), device=device), causal_mask], axis=-1
+                    )
                 # causal and attention masks must have same type with pytorch version < 1.3
                 causal_mask = causal_mask.to(attention_mask.dtype)
+
                 extended_attention_mask = causal_mask[:, None, :, :] * attention_mask[:, None, None, :]
             else:
                 extended_attention_mask = attention_mask[:, None, None, :]