diff --git a/egs/librispeech/ASR/.run_v3.sh.swp b/egs/librispeech/ASR/.run_v3.sh.swp
index c769b208b..0db172118 100644
Binary files a/egs/librispeech/ASR/.run_v3.sh.swp and b/egs/librispeech/ASR/.run_v3.sh.swp differ
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless_d2v_v2/.train.py.swp b/egs/librispeech/ASR/pruned_transducer_stateless_d2v_v2/.train.py.swp
index c9557d85a..483165d63 100644
Binary files a/egs/librispeech/ASR/pruned_transducer_stateless_d2v_v2/.train.py.swp and b/egs/librispeech/ASR/pruned_transducer_stateless_d2v_v2/.train.py.swp differ
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless_d2v_v2/train.py b/egs/librispeech/ASR/pruned_transducer_stateless_d2v_v2/train.py
index be454c04d..013d7c60b 100755
--- a/egs/librispeech/ASR/pruned_transducer_stateless_d2v_v2/train.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless_d2v_v2/train.py
@@ -1009,7 +1009,7 @@ def train_one_epoch(
 
             # NOTE: We use reduction==sum and loss is computed over utterances
             # in the batch and there is no normalization to it so far.
-            logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            scaler.scale(loss).backward()
 
             if params.multi_optim and batch_idx % params.accum_grads == 0:
                 set_batch_count(model, params.batch_idx_train)