diff --git a/egs/librispeech/ASR/pruned_transducer_stateless/train.py b/egs/librispeech/ASR/pruned_transducer_stateless/train.py
index b625ed3ff2..33b23038cf 100755
--- a/egs/librispeech/ASR/pruned_transducer_stateless/train.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless/train.py
@@ -457,9 +457,6 @@ def load_checkpoint_if_available(
         if "cur_epoch" in saved_params:
             params["start_epoch"] = saved_params["cur_epoch"]
 
-        if "cur_batch_idx" in saved_params:
-            params["cur_batch_idx"] = saved_params["cur_batch_idx"]
-
     return saved_params
 
 
@@ -674,13 +671,7 @@ def maybe_log_weights(tag: str):
                 global_step=params.batch_idx_train,
             )
 
-    cur_batch_idx = params.get("cur_batch_idx", 0)
-
     for batch_idx, batch in enumerate(train_dl):
-        if batch_idx < cur_batch_idx:
-            continue
-        cur_batch_idx = batch_idx
-
         params.batch_idx_train += 1
         batch_size = len(batch["supervisions"]["text"])
 
@@ -728,7 +719,6 @@ def maybe_log_weights(tag: str):
             params.batch_idx_train > 0
             and params.batch_idx_train % params.save_every_n == 0
         ):
-            params.cur_batch_idx = batch_idx
             save_checkpoint_with_global_batch_idx(
                 out_dir=params.exp_dir,
                 global_batch_idx=params.batch_idx_train,
@@ -738,7 +728,6 @@ def maybe_log_weights(tag: str):
                 sampler=train_dl.sampler,
                 rank=rank,
             )
-            del params.cur_batch_idx
             remove_checkpoints(
                 out_dir=params.exp_dir,
                 topk=params.keep_last_k,
@@ -893,13 +882,14 @@ def remove_short_and_long_utt(c: Cut):
     valid_cuts += librispeech.dev_other_cuts()
     valid_dl = librispeech.valid_dataloaders(valid_cuts)
 
-    scan_pessimistic_batches_for_oom(
-        model=model,
-        train_dl=train_dl,
-        optimizer=optimizer,
-        sp=sp,
-        params=params,
-    )
+    if params.start_batch <= 0:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            params=params,
+        )
 
     for epoch in range(params.start_epoch, params.num_epochs):
         fix_random_seed(params.seed + epoch)
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless2/train.py b/egs/librispeech/ASR/pruned_transducer_stateless2/train.py
index 46d2cb86d4..4d290e39fa 100755
--- a/egs/librispeech/ASR/pruned_transducer_stateless2/train.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless2/train.py
@@ -503,9 +503,6 @@ def load_checkpoint_if_available(
         if "cur_epoch" in saved_params:
             params["start_epoch"] = saved_params["cur_epoch"]
 
-        if "cur_batch_idx" in saved_params:
-            params["cur_batch_idx"] = saved_params["cur_batch_idx"]
-
     return saved_params
 
 
@@ -724,13 +721,7 @@ def train_one_epoch(
 
     tot_loss = MetricsTracker()
 
-    cur_batch_idx = params.get("cur_batch_idx", 0)
-
     for batch_idx, batch in enumerate(train_dl):
-        if batch_idx < cur_batch_idx:
-            continue
-        cur_batch_idx = batch_idx
-
         params.batch_idx_train += 1
         batch_size = len(batch["supervisions"]["text"])
 
@@ -765,7 +756,6 @@ def train_one_epoch(
             params.batch_idx_train > 0
             and params.batch_idx_train % params.save_every_n == 0
         ):
-            params.cur_batch_idx = batch_idx
             save_checkpoint_with_global_batch_idx(
                 out_dir=params.exp_dir,
                 global_batch_idx=params.batch_idx_train,
@@ -777,7 +767,6 @@ def train_one_epoch(
                 scaler=scaler,
                 rank=rank,
             )
-            del params.cur_batch_idx
             remove_checkpoints(
                 out_dir=params.exp_dir,
                 topk=params.keep_last_k,
@@ -944,7 +933,7 @@ def remove_short_and_long_utt(c: Cut):
     valid_cuts += librispeech.dev_other_cuts()
     valid_dl = librispeech.valid_dataloaders(valid_cuts)
 
-    if not params.print_diagnostics:
+    if params.start_batch <= 0 and not params.print_diagnostics:
         scan_pessimistic_batches_for_oom(
             model=model,
             train_dl=train_dl,
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/train.py b/egs/librispeech/ASR/pruned_transducer_stateless3/train.py
index ff9b8d8087..914b9b5eba 100755
--- a/egs/librispeech/ASR/pruned_transducer_stateless3/train.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless3/train.py
@@ -1058,14 +1058,15 @@ def run(rank, world_size, args):
     # It's time consuming to include `giga_train_dl` here
     #  for dl in [train_dl, giga_train_dl]:
     for dl in [train_dl]:
-        scan_pessimistic_batches_for_oom(
-            model=model,
-            train_dl=dl,
-            optimizer=optimizer,
-            sp=sp,
-            params=params,
-            warmup=0.0 if params.start_epoch == 0 else 1.0,
-        )
+        if params.start_batch <= 0:
+            scan_pessimistic_batches_for_oom(
+                model=model,
+                train_dl=dl,
+                optimizer=optimizer,
+                sp=sp,
+                params=params,
+                warmup=0.0 if params.start_epoch == 0 else 1.0,
+            )
 
     scaler = GradScaler(enabled=params.use_fp16)
     if checkpoints and "grad_scaler" in checkpoints:
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless4/train.py b/egs/librispeech/ASR/pruned_transducer_stateless4/train.py
index 893a6a749a..325b013230 100755
--- a/egs/librispeech/ASR/pruned_transducer_stateless4/train.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless4/train.py
@@ -525,9 +525,6 @@ def load_checkpoint_if_available(
         if "cur_epoch" in saved_params:
             params["start_epoch"] = saved_params["cur_epoch"]
 
-        if "cur_batch_idx" in saved_params:
-            params["cur_batch_idx"] = saved_params["cur_batch_idx"]
-
     return saved_params
 
 
@@ -757,13 +754,7 @@ def train_one_epoch(
 
     tot_loss = MetricsTracker()
 
-    cur_batch_idx = params.get("cur_batch_idx", 0)
-
     for batch_idx, batch in enumerate(train_dl):
-        if batch_idx < cur_batch_idx:
-            continue
-        cur_batch_idx = batch_idx
-
         params.batch_idx_train += 1
         batch_size = len(batch["supervisions"]["text"])
 
@@ -805,7 +796,6 @@ def train_one_epoch(
             params.batch_idx_train > 0
             and params.batch_idx_train % params.save_every_n == 0
         ):
-            params.cur_batch_idx = batch_idx
             save_checkpoint_with_global_batch_idx(
                 out_dir=params.exp_dir,
                 global_batch_idx=params.batch_idx_train,
@@ -818,7 +808,6 @@ def train_one_epoch(
                 scaler=scaler,
                 rank=rank,
             )
-            del params.cur_batch_idx
             remove_checkpoints(
                 out_dir=params.exp_dir,
                 topk=params.keep_last_k,
@@ -993,7 +982,7 @@ def remove_short_and_long_utt(c: Cut):
     valid_cuts += librispeech.dev_other_cuts()
     valid_dl = librispeech.valid_dataloaders(valid_cuts)
 
-    if not params.print_diagnostics:
+    if params.start_batch <= 0 and not params.print_diagnostics:
         scan_pessimistic_batches_for_oom(
             model=model,
             train_dl=train_dl,
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/train.py b/egs/librispeech/ASR/pruned_transducer_stateless5/train.py
index 44abdcd490..3bfe22155f 100755
--- a/egs/librispeech/ASR/pruned_transducer_stateless5/train.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless5/train.py
@@ -550,9 +550,6 @@ def load_checkpoint_if_available(
         if "cur_epoch" in saved_params:
             params["start_epoch"] = saved_params["cur_epoch"]
 
-        if "cur_batch_idx" in saved_params:
-            params["cur_batch_idx"] = saved_params["cur_batch_idx"]
-
     return saved_params
 
 
@@ -782,13 +779,7 @@ def train_one_epoch(
 
     tot_loss = MetricsTracker()
 
-    cur_batch_idx = params.get("cur_batch_idx", 0)
-
     for batch_idx, batch in enumerate(train_dl):
-        if batch_idx < cur_batch_idx:
-            continue
-        cur_batch_idx = batch_idx
-
         params.batch_idx_train += 1
         batch_size = len(batch["supervisions"]["text"])
 
@@ -834,7 +825,6 @@ def train_one_epoch(
             params.batch_idx_train > 0
             and params.batch_idx_train % params.save_every_n == 0
         ):
-            params.cur_batch_idx = batch_idx
             save_checkpoint_with_global_batch_idx(
                 out_dir=params.exp_dir,
                 global_batch_idx=params.batch_idx_train,
@@ -847,7 +837,6 @@ def train_one_epoch(
                 scaler=scaler,
                 rank=rank,
             )
-            del params.cur_batch_idx
             remove_checkpoints(
                 out_dir=params.exp_dir,
                 topk=params.keep_last_k,
@@ -1025,7 +1014,7 @@ def remove_short_and_long_utt(c: Cut):
     valid_cuts += librispeech.dev_other_cuts()
     valid_dl = librispeech.valid_dataloaders(valid_cuts)
 
-    if not params.print_diagnostics:
+    if params.start_batch <= 0 and not params.print_diagnostics:
         scan_pessimistic_batches_for_oom(
             model=model,
             train_dl=train_dl,
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless6/train.py b/egs/librispeech/ASR/pruned_transducer_stateless6/train.py
index 596f8f7d93..a4595211c3 100755
--- a/egs/librispeech/ASR/pruned_transducer_stateless6/train.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless6/train.py
@@ -507,9 +507,6 @@ def load_checkpoint_if_available(
         if "cur_epoch" in saved_params:
             params["start_epoch"] = saved_params["cur_epoch"]
 
-        if "cur_batch_idx" in saved_params:
-            params["cur_batch_idx"] = saved_params["cur_batch_idx"]
-
     return saved_params
 
 
@@ -763,13 +760,7 @@ def train_one_epoch(
 
     tot_loss = MetricsTracker()
 
-    cur_batch_idx = params.get("cur_batch_idx", 0)
-
     for batch_idx, batch in enumerate(train_dl):
-        if batch_idx < cur_batch_idx:
-            continue
-        cur_batch_idx = batch_idx
-
         params.batch_idx_train += 1
         batch_size = len(batch["supervisions"]["text"])
 
@@ -811,7 +802,6 @@ def train_one_epoch(
             params.batch_idx_train > 0
             and params.batch_idx_train % params.save_every_n == 0
         ):
-            params.cur_batch_idx = batch_idx
             save_checkpoint_with_global_batch_idx(
                 out_dir=params.exp_dir,
                 global_batch_idx=params.batch_idx_train,
@@ -824,7 +814,6 @@ def train_one_epoch(
                 scaler=scaler,
                 rank=rank,
             )
-            del params.cur_batch_idx
             remove_checkpoints(
                 out_dir=params.exp_dir,
                 topk=params.keep_last_k,
@@ -999,7 +988,7 @@ def remove_short_and_long_utt(c: Cut):
     valid_cuts += librispeech.dev_other_cuts()
     valid_dl = librispeech.valid_dataloaders(valid_cuts)
 
-    if not params.print_diagnostics:
+    if params.start_batch <= 0 and not params.print_diagnostics:
         scan_pessimistic_batches_for_oom(
             model=model,
             train_dl=train_dl,