diff --git a/library/train_util.py b/library/train_util.py
index a20edbe1..676652e9 100644
--- a/library/train_util.py
+++ b/library/train_util.py
@@ -3964,6 +3964,8 @@ def prepare_accelerator(args: argparse.Namespace):
             zero3_init_flag=args.zero3_init_flag, zero3_save_16bit_model=args.zero3_save_16bit_model,
         )
         deepspeed_plugin.deepspeed_config['train_micro_batch_size_per_gpu'] = args.train_batch_size
+        deepspeed_plugin.deepspeed_config['train_batch_size'] = \
+            args.train_batch_size * args.gradient_accumulation_steps * int(os.environ['WORLD_SIZE'])
 
     accelerator = Accelerator(
         gradient_accumulation_steps=args.gradient_accumulation_steps,
diff --git a/sdxl_train.py b/sdxl_train.py
index e8680828..ef3ead38 100644
--- a/sdxl_train.py
+++ b/sdxl_train.py
@@ -391,6 +391,12 @@ def train(args):
 
     if args.deepspeed:
         # Wrapping model for DeepSpeed
+        import deepspeed
+        if args.offload_optimizer_device is not None:
+            accelerator.print('[DeepSpeed] start to manually build cpu_adam.')
+            deepspeed.ops.op_builder.CPUAdamBuilder().load()
+            accelerator.print('[DeepSpeed] building cpu_adam done.')
+
         class DeepSpeedModel(torch.nn.Module): 
             def __init__(self, unet, text_encoder) -> None:
                 super().__init__()