diff --git a/egs/librispeech/ASR/.run_v3.sh.swp b/egs/librispeech/ASR/.run_v3.sh.swp
index a1471bcc9..594962fdb 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 69093f155..cd7d73fdb 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 0c299c744..33bc35195 100755
--- a/egs/librispeech/ASR/pruned_transducer_stateless_d2v_v2/train.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless_d2v_v2/train.py
@@ -1013,6 +1013,7 @@ def train_one_epoch(
                 scaler.scale(loss).backward()
             else:
                 logging.warning(f"Grad scale is small: {cur_grad_scale}")
+
             if params.multi_optim and batch_idx % params.accum_grads == 0:
                 set_batch_count(model, params.batch_idx_train)
                 scheduler_enc.step_batch(params.batch_idx_train)
diff --git a/egs/librispeech/ASR/run_v3.sh b/egs/librispeech/ASR/run_v3.sh
index d46407a42..8259523de 100755
--- a/egs/librispeech/ASR/run_v3.sh
+++ b/egs/librispeech/ASR/run_v3.sh
@@ -36,13 +36,12 @@ else
 		--enable-spec-aug False \
 		--multi-optim True \
 		--world-size 4 \
+		--start-batch 34000 \
 		--num-epochs 30 \
-		--start-epoch 6 \
 		--full-libri 1 \
 		--exp-dir ./pruned_transducer_stateless_d2v_v2/$1 \
 		--max-duration 150 \
-		--freeze-finetune-updates 2000 \
-		--use-fp16 1 \
+		--freeze-finetune-updates 2000 #		--use-fp16 1 \
 		--peak-enc-lr 0.001 \
 		--peak-dec-lr 0.5 \
 		--accum-grads 3 \
@@ -55,3 +54,5 @@ else
 		--context-size 2 \
 		--ctc-loss-scale 0.2 
 fi
+
+#--start-epoch 6 \