Bart-CNN (#3059)

`generate` code that produces 99% identical summarizations to fairseq on CNN test data, with caching.

src/transformers/configuration_bart.py

@@ -26,7 +26,7 @@ _bart_large_url = "https://s3.amazonaws.com/models.huggingface.co/bert/facebook/
 BART_PRETRAINED_CONFIG_ARCHIVE_MAP = {
     "bart-large": _bart_large_url,
     "bart-large-mnli": _bart_large_url,  # fine as same
-    "bart-cnn": None,  # not done
+    "bart-large-cnn": "https://s3.amazonaws.com/models.huggingface.co/bert/facebook/bart-large-cnn/config.json",
 }
 
 
@@ -59,6 +59,7 @@ class BartConfig(PretrainedConfig):
         classifier_dropout=0.0,
         output_past=False,
         num_labels=3,
+        bos_token_id=0,
         **common_kwargs
     ):
         r"""
@@ -67,12 +68,16 @@ class BartConfig(PretrainedConfig):
                 config = BartConfig.from_pretrained('bart-large')
                 model = BartModel(config)
         """
-        super().__init__(num_labels=num_labels, output_past=output_past, pad_token_id=pad_token_id, **common_kwargs)
-
+        super().__init__(
+            num_labels=num_labels,
+            output_past=output_past,
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            **common_kwargs,
+        )
         self.vocab_size = vocab_size
         self.d_model = d_model  # encoder_embed_dim and decoder_embed_dim
         self.eos_token_id = eos_token_id
-
         self.encoder_ffn_dim = encoder_ffn_dim
         self.encoder_layers = self.num_hidden_layers = encoder_layers
         self.encoder_attention_heads = encoder_attention_heads