diff --git a/library/train_util.py b/library/train_util.py
index 62cd145e1..b9b4eecf8 100644
--- a/library/train_util.py
+++ b/library/train_util.py
@@ -2416,6 +2416,7 @@ def add_training_arguments(parser: argparse.ArgumentParser, support_dreambooth:
         "--mixed_precision", type=str, default="no", choices=["no", "fp16", "bf16"], help="use mixed precision / 混合精度を使う場合、その精度"
     )
     parser.add_argument("--full_fp16", action="store_true", help="fp16 training including gradients / 勾配も含めてfp16で学習する")
+    parser.add_argument("--full_bf16", action="store_true", help="bf16 training including gradients / 勾配も含めてbf16で学習する")
     parser.add_argument(
         "--clip_skip",
         type=int,
diff --git a/sdxl_train.py b/sdxl_train.py
index 9cf20252d..06cbc5710 100644
--- a/sdxl_train.py
+++ b/sdxl_train.py
@@ -267,6 +267,14 @@ def fn_recursive_set_mem_eff(module: torch.nn.Module):
         unet.to(weight_dtype)
         text_encoder1.to(weight_dtype)
         text_encoder2.to(weight_dtype)
+    elif args.full_bf16:
+        assert (
+            args.mixed_precision == "bf16"
+        ), "full_bf16 requires mixed precision='bf16' / full_bf16を使う場合はmixed_precision='bf16'を指定してください。"
+        accelerator.print("enable full bf16 training.")
+        unet.to(weight_dtype)
+        text_encoder1.to(weight_dtype)
+        text_encoder2.to(weight_dtype)
 
     # acceleratorがなんかよろしくやってくれるらしい
     if args.train_text_encoder: