diff --git a/fine_tune.py b/fine_tune.py
index eb652742..741e9c85 100644
--- a/fine_tune.py
+++ b/fine_tune.py
@@ -224,8 +224,9 @@ def train(args):
         args.max_train_steps = args.max_train_epochs * math.ceil(
             len(train_dataloader) / accelerator.num_processes / args.gradient_accumulation_steps
         )
-        accelerator.print(f"override steps. steps for {args.max_train_epochs} epochs is / 指定エポックまでのステップ数: {args.max_train_steps}")
-
+        accelerator.print(
+            f"override steps. steps for {args.max_train_epochs} epochs is / 指定エポックまでのステップ数: {args.max_train_steps}"
+        )
     # データセット側にも学習ステップを送信
     train_dataset_group.set_max_train_steps(args.max_train_steps)
 
diff --git a/library/train_util.py b/library/train_util.py
index dbe5a61c..61c83624 100644
--- a/library/train_util.py
+++ b/library/train_util.py
@@ -3975,7 +3975,11 @@ def prepare_accelerator(args: argparse.Namespace):
         if args.mixed_precision.lower() == "fp16":
             deepspeed_plugin.deepspeed_config['fp16']['initial_scale_power'] = 0
         if args.full_fp16 or args.fp16_master_weights_and_gradients:
-            deepspeed_plugin.deepspeed_config['fp16_master_weights_and_gradients'] = True
+            if args.offload_optimizer_device == "cpu":
+                deepspeed_plugin.deepspeed_config['fp16']['fp16_master_weights_and_grads'] = True
+                print("[DeepSpeed] full fp16 enable.")
+            else:
+                print("full fp16, fp16_master_weights_and_grads currently only supported using ZeRO-Offload with DeepSpeedCPUAdam.")
 
     accelerator = Accelerator(
         gradient_accumulation_steps=args.gradient_accumulation_steps,
diff --git a/sdxl_train.py b/sdxl_train.py
index 54902b87..6ffb1bba 100644
--- a/sdxl_train.py
+++ b/sdxl_train.py
@@ -363,8 +363,9 @@ def train(args):
         args.max_train_steps = args.max_train_epochs * math.ceil(
             len(train_dataloader) / accelerator.num_processes / args.gradient_accumulation_steps
         )
-        accelerator.print(f"override steps. steps for {args.max_train_epochs} epochs is / 指定エポックまでのステップ数: {args.max_train_steps}")
-
+        accelerator.print(
+            f"override steps. steps for {args.max_train_epochs} epochs is / 指定エポックまでのステップ数: {args.max_train_steps}"
+        )
     # データセット側にも学習ステップを送信
     train_dataset_group.set_max_train_steps(args.max_train_steps)
 
diff --git a/train_db.py b/train_db.py
index 58536555..c336a1c1 100644
--- a/train_db.py
+++ b/train_db.py
@@ -193,8 +193,9 @@ def train(args):
         args.max_train_steps = args.max_train_epochs * math.ceil(
             len(train_dataloader) / accelerator.num_processes / args.gradient_accumulation_steps
         )
-        accelerator.print(f"override steps. steps for {args.max_train_epochs} epochs is / 指定エポックまでのステップ数: {args.max_train_steps}")
-
+        accelerator.print(
+            f"override steps. steps for {args.max_train_epochs} epochs is / 指定エポックまでのステップ数: {args.max_train_steps}"
+        )
     # データセット側にも学習ステップを送信
     train_dataset_group.set_max_train_steps(args.max_train_steps)