prepare_seq2seq_batch makes labels/ decoder_input_ids made later. (#6654)

* broken test

* batch parity

* tests pass

* boom boom

* boom boom

* split out bart tokenizer tests

* fix tests

* boom boom

* Fixed dataset bug

* Fix marian

* Undo extra

* Get marian working

* Fix t5 tok tests

* Test passing

* Cleanup

* better assert msg

* require torch

* Fix mbart tests

* undo extra decoder_attn_mask change

* Fix import

* pegasus tokenizer can ignore src_lang kwargs

* unused kwarg test cov

* boom boom

* add todo for pegasus issue

* cover one word translation edge case

* Cleanup

* doc

examples/seq2seq/README.md

@@ -71,8 +71,8 @@ Summarization Tips:
 (It rarely makes sense to start from `bart-large` unless you are a researching finetuning methods).
 
 **Update 2018-07-18**
-Datasets: `Seq2SeqDataset` should be used for all tokenizers without a `prepare_seq2seq_batch` method. For those who do (like Marian, MBart), `TranslationDataset` should be used.**
-A new dataset is needed to support multilingual tasks.
+Datasets: `LegacySeq2SeqDataset` will be used for all tokenizers without a `prepare_seq2seq_batch` method. Otherwise, `Seq2SeqDataset` will be used.
+Future work/help wanted: A new dataset to support multilingual tasks.
 
 
 ### Command Line Options
@@ -106,7 +106,7 @@ The following command should work on a 16GB GPU:
     --train_batch_size=1 \
     --eval_batch_size=1 \
     --output_dir=xsum_results \
-    --num_train_epochs 1 \
+    --num_train_epochs 6 \
     --model_name_or_path facebook/bart-large
 ```