Fix ernie_gen and CRF bug (#5057)

* add ernie_gen

* optimize ernie_gen

* optimize ernie_gen

* optimize ernie_gen code

* fix crf bug

* add ernie_gen __init__.py

* modify nlp version

* fix ernie_gen predict

PaddleNLP/examples/text_generation/ernie-gen/predict.py

@@ -129,7 +129,6 @@ def predict():
                 map(post_process, vocab.to_tokens(predict_ids)))
             print("source :%s\ntarget :%s\npredict:%s\n" %
                   (source_sentence, tgt_sentence, predict_ids))
-        break
 
 
 if __name__ == "__main__":

PaddleNLP/paddlenlp/__init__.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-__version__ = '2.0.0a6'
+__version__ = '2.0.0a7'
 
 from . import data
 from . import datasets

PaddleNLP/paddlenlp/layers/crf.py

@@ -230,10 +230,10 @@ class LinearChainCrf(nn.Layer):
         return self._seq_index[:length]
 
     def _get_batch_seq_index(self, batch_size, length):
-        if self._batch_seq_index is None or length > self._batch_seq_index.shape[1]:
+        if self._batch_seq_index is None or length + 2 > self._batch_seq_index.shape[
+                1]:
             self._batch_seq_index = paddle.cumsum(
-                paddle.ones([batch_size, length + 2], "int64"),
-                axis=1) - 1
+                paddle.ones([batch_size, length + 2], "int64"), axis=1) - 1
         if self.with_start_stop_tag:
             return self._batch_seq_index[:, :length + 2]
         else: