diff --git a/mindnlp/core/npu/amp/__init__.py b/mindnlp/core/npu/amp/__init__.py
index de8bea8a6..34b9bfbaa 100644
--- a/mindnlp/core/npu/amp/__init__.py
+++ b/mindnlp/core/npu/amp/__init__.py
@@ -1,8 +1,9 @@
 from .autocast_mode import autocast, custom_bwd, custom_fwd
-
+from .grad_scaler import GradScaler
 
 __all__ = [
     "autocast",
     "custom_bwd",
     "custom_fwd",
+    "GradScaler"
 ]
\ No newline at end of file
diff --git a/mindnlp/core/npu/amp/grad_scaler.py b/mindnlp/core/npu/amp/grad_scaler.py
new file mode 100644
index 000000000..07adf1524
--- /dev/null
+++ b/mindnlp/core/npu/amp/grad_scaler.py
@@ -0,0 +1,38 @@
+from typing_extensions import deprecated
+
+from mindnlp import core
+
+# We need to keep this unused import for BC reasons
+from ...amp.grad_scaler import OptState  # noqa: F401
+
+
+__all__ = ["GradScaler"]
+
+
+class GradScaler(core.amp.GradScaler):
+    r"""
+    See :class:`torch.amp.GradScaler`.
+    ``torch.npu.amp.GradScaler(args...)`` is deprecated. Please use ``torch.amp.GradScaler("npu", args...)`` instead.
+    """
+
+    @deprecated(
+        "`torch.npu.amp.GradScaler(args...)` is deprecated. "
+        "Please use `torch.amp.GradScaler('npu', args...)` instead.",
+        category=FutureWarning,
+    )
+    def __init__(
+        self,
+        init_scale: float = 2.0**16,
+        growth_factor: float = 2.0,
+        backoff_factor: float = 0.5,
+        growth_interval: int = 2000,
+        enabled: bool = True,
+    ) -> None:
+        super().__init__(
+            "npu",
+            init_scale=init_scale,
+            growth_factor=growth_factor,
+            backoff_factor=backoff_factor,
+            growth_interval=growth_interval,
+            enabled=enabled,
+        )
\ No newline at end of file
diff --git a/mindnlp/core/ops/creation.py b/mindnlp/core/ops/creation.py
index 91e07f539..f5f89ae59 100644
--- a/mindnlp/core/ops/creation.py
+++ b/mindnlp/core/ops/creation.py
@@ -189,11 +189,13 @@ def empty(*size, dtype=None, device=None, requires_grad=False, pin_memory=False,
             device = 'CPU'
         elif device.lower() == 'npu':
             device = 'Ascend'
-        else:
+        elif device.lower() == 'cuda':
             device = 'GPU'
+        else:
+            device = 'meta'
 
     # To avoid the problem in irecv and recv of using empty.
-    if has_empty and use_pyboost():
+    if device != 'meta':
         out = mindspore.mint.empty(size, dtype=dtype, device=device)
     else:
         out = CTensor(dtype=dtype, shape=size)
diff --git a/mindnlp/utils/torch_proxy.py b/mindnlp/utils/torch_proxy.py
index 81950036e..ae1fccae8 100644
--- a/mindnlp/utils/torch_proxy.py
+++ b/mindnlp/utils/torch_proxy.py
@@ -6,6 +6,8 @@
 import importlib.machinery
 from types import ModuleType
 
+from mindnlp.core.configs import DEVICE_TARGET
+
 TORCH_VERSION = '2.7.1+dev'
 
 class RedirectFinder(importlib.abc.MetaPathFinder):
@@ -19,6 +21,8 @@ def find_spec(self, fullname, path, target=None):
             if fullname == proxy_prefix or fullname.startswith(proxy_prefix + "."):
                 # 计算实际模块名
                 target_name = fullname.replace(proxy_prefix, target_prefix, 1)
+                if DEVICE_TARGET == 'Ascend':
+                    target_name = target_name.replace('cuda', 'npu')
                 try:
                     importlib.import_module(target_name)
                 except Exception as e: