diff --git a/egs/librispeech/ASR/.test.sh.swp b/egs/librispeech/ASR/.test.sh.swp
index a7ce6b001..020fd2130 100644
Binary files a/egs/librispeech/ASR/.test.sh.swp and b/egs/librispeech/ASR/.test.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 df2c13238..af9842993 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 914d621e3..38caa2b3e 100755
--- a/egs/librispeech/ASR/pruned_transducer_stateless_d2v_v2/train.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless_d2v_v2/train.py
@@ -1009,6 +1009,8 @@ 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.
+            if scaler._scale.item() < 1.0e-05:
+                continue
             scaler.scale(loss).backward()
             if params.multi_optim and batch_idx % params.accum_grads == 0:
                 set_batch_count(model, params.batch_idx_train)