add past hidden key states for more efficient language generation & add prepare_inputs for gpt2 and ctrl model

src/transformers/modeling_ctrl.py

@@ -490,6 +490,14 @@ class CTRLLMHeadModel(CTRLPreTrainedModel):
     def get_output_embeddings(self):
         return self.lm_head
 
+    def prepare_inputs_for_generation(self, input_ids, **kwargs):
+        # inputs_ids contain only last token if past is in kwargs and defined
+        input_ids = input_ids[:, -1].unsqueeze(-1) if 'past' in kwargs and kwargs['past'] else input_ids
+
+        inputs = {"input_ids": input_ids}
+        inputs.update(kwargs)
+        return inputs
+
     def forward(
         self,
         input_ids=None,