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 8961359e4..d88c03f4e 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 d3ff4f0e0..72304341d 100755
--- a/egs/librispeech/ASR/pruned_transducer_stateless_d2v_v2/train.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless_d2v_v2/train.py
@@ -1086,6 +1086,12 @@ def train_one_epoch(
                     f"grad_scale is too small, exiting: {cur_grad_scale}"
                 )
 
+        if params.batch_idx_train > 4000 and loss > 300:
+            wb.log({"valid/loss": 10000})
+            raise RunteimError(
+                    f"divergence... exiting: loss={loss}"
+                )
+
         if batch_idx % (params.log_interval*params.accum_grads) == 0:
             if params.multi_optim:
                 cur_enc_lr = scheduler_enc.get_last_lr()[0]