downcast the scores after upcasting to prevent runtime errors

this change fixes https://github.com/huggingface/transformers/issues/41238

custom_generate/generate.py

@@ -379,7 +379,8 @@ def _group_beam_search(
             next_token_scores = next_token_scores.expand_as(next_token_scores_processed)
 
             if output_scores:
-                processed_score[batch_group_indices] = next_token_scores_processed
+                processed_score[batch_group_indices] = next_token_scores_processed.to(processed_score.dtype)
+
 
             # reshape for beam search
             next_token_scores = next_token_scores.view(