fix use of mems in Transformer-XL (#4826)

Fixed duplicated memory use in Transformer-XL generation leading to bad predictions and performance.

src/transformers/modeling_transfo_xl.py

@@ -1016,11 +1016,14 @@ class TransfoXLLMHeadModel(TransfoXLPreTrainedModel):
             return self.crit.out_layers[-1]
 
     def prepare_inputs_for_generation(self, input_ids, past, **model_kwargs):
-        inputs = {"input_ids": input_ids}
+        inputs = {}
 
         # if past is defined in model kwargs then use it for faster decoding
         if past:
             inputs["mems"] = past
+            inputs["input_ids"] = input_ids[:, -1].unsqueeze(-1)
+        else:
+            inputs["input_ids"] = input_ids
 
         return inputs