Avoid dummy token in PLD to optimize performance (#29445)

src/transformers/generation/candidate_generator.py

@@ -302,8 +302,8 @@ class PromptLookupCandidateGenerator(CandidateGenerator):
                 break
 
         if chosen_ids is None or len(chosen_ids) == 0:
-            # Need to make a dummy tensor to avoid errors
-            chosen_ids = torch.zeros((1), dtype=torch.long, device=input_ids.device)
+            # In case we didn't find a match return the input sequence unchanged, reverts back to autoregressive decoding
+            return input_ids, None
 
         # Now need extend input_ids with chosen_ids
         chosen_ids = chosen_ids.unsqueeze(0)