diff --git a/paddle/fluid/imperative/amp_auto_cast.cc b/paddle/fluid/imperative/amp_auto_cast.cc
index 149202468be6c6bec833f100adfd4100c520f8f3..dd00b75666dc4ba576a7b011d3bb017f2e9c29d1 100644
--- a/paddle/fluid/imperative/amp_auto_cast.cc
+++ b/paddle/fluid/imperative/amp_auto_cast.cc
@@ -209,7 +209,9 @@ inline bool NeedCast(const std::shared_ptr<VarType>& var) {
   auto data_type = GetDataType<VarType>(var);
   if (paddle::platform::is_gpu_place(place) ||
       paddle::platform::is_cuda_pinned_place(place) ||
-      paddle::platform::is_xpu_place(place)) {
+      paddle::platform::is_xpu_place(place) ||
+      paddle::platform::is_npu_place(place) ||
+      paddle::platform::is_npu_pinned_place(place)) {
     // CudaPinndePlace is added for varbase created by dataloader
     if (data_type == paddle::framework::proto::VarType::FP32 ||
         data_type == paddle::framework::proto::VarType::FP16 ||
diff --git a/paddle/fluid/pybind/op_function_generator.h b/paddle/fluid/pybind/op_function_generator.h
index 9e86e3df8a6884ec1b75b8525ad858ff8f2e233c..d8750c1d6c115a6de8a493cac4ccadbd47bc10fd 100644
--- a/paddle/fluid/pybind/op_function_generator.h
+++ b/paddle/fluid/pybind/op_function_generator.h
@@ -88,6 +88,7 @@ std::map<std::string, std::set<std::string>> op_ins_map = {
     {"nce",
      {"Input", "Label", "Weight", "Bias", "SampleWeight", "CustomDistProbs",
       "CustomDistAlias", "CustomDistAliasProbs"}},
+    {"check_finite_and_unscale", {"X", "Scale", "FloatStatus"}},
 };
 
 // NOTE(zhiqiu): Like op_ins_map.
diff --git a/python/paddle/fluid/dygraph/amp/auto_cast.py b/python/paddle/fluid/dygraph/amp/auto_cast.py
index 191661b7bf9d5a5b2877f22ebe6d2ec9124f2f96..a449bdf0a189613b611b0cf7034c828e235e259c 100644
--- a/python/paddle/fluid/dygraph/amp/auto_cast.py
+++ b/python/paddle/fluid/dygraph/amp/auto_cast.py
@@ -271,14 +271,19 @@ def amp_guard(enable=True,
             "current_tracer is None, maybe it is not in imperative mode.")
 
     # check device_type:
-    # NOTE: Now, amp only support gpu for float16 and bfloat16, xpu for float16.
+    # NOTE: Now, amp only support gpu for float16 and bfloat16, xpu for float16, npu for float16.
     # Maybe we will support cpu for bfloat16.
     if enable and not (tracer._expected_place.is_gpu_place() or
-                       tracer._expected_place.is_xpu_place()):
+                       tracer._expected_place.is_xpu_place() or
+                       tracer._expected_place.is_npu_place()):
         warnings.warn(
-            'amp_guard can only be enabled on CUDAPlace and XPUPlace, current place is %s, so it makes no effect.'
+            'amp_guard can only be enabled on CUDAPlace, XPUPlace, and NPUPlace, current place is %s, so it makes no effect.'
             % tracer._expected_place)
         enable = False
+    # For npu:
+    if tracer._expected_place.is_npu_place() and (dtype == 'bfloat16'):
+        warnings.warn('NPUPlace only support float16 amp.')
+        enable = False
     # For xpu:
     if tracer._expected_place.is_xpu_place() and (dtype == 'bfloat16'):
         warnings.warn('XPUPlace only support float16 amp.')
diff --git a/python/paddle/fluid/dygraph/amp/loss_scaler.py b/python/paddle/fluid/dygraph/amp/loss_scaler.py
index f7c2d6be574c4e17fa5ce6fa44ca4ecc55a5eb95..3ca4c7dca76d2ac6157196bc9d25d3f4a0df3a83 100644
--- a/python/paddle/fluid/dygraph/amp/loss_scaler.py
+++ b/python/paddle/fluid/dygraph/amp/loss_scaler.py
@@ -105,9 +105,10 @@ class AmpScaler(object):
                 "current_tracer is None, maybe it is not in imperative mode.")
 
         if enable and not (tracer._expected_place.is_gpu_place() or
-                           tracer._expected_place.is_xpu_place()):
+                           tracer._expected_place.is_xpu_place() or
+                           tracer._expected_place.is_npu_place()):
             warnings.warn(
-                'AmpScaler can only be enabled on CUDAPlace and XPUPlace, current place is %s, so it makes no effect.'
+                'AmpScaler can only be enabled on CUDAPlace, XPUPlace and NPUPlace, current place is %s, so it makes no effect.'
                 % tracer._expected_place)
             enable = False
 
@@ -286,14 +287,28 @@ class AmpScaler(object):
                     ) and (param._grad_ivar().dtype == core.VarDesc.VarType.FP32
                            )
             ]
-        if len(param_grads_fp16):
-            _C_ops.check_finite_and_unscale(param_grads_fp16, self._scale,
-                                            param_grads_fp16,
-                                            self._temp_found_inf_fp16)
-        if len(param_grads_fp32):
-            _C_ops.check_finite_and_unscale(param_grads_fp32, self._scale,
-                                            param_grads_fp32,
-                                            self._temp_found_inf_fp32)
+        if core.is_compiled_with_npu():
+            float_status = _C_ops.alloc_float_status()
+            _C_ops.clear_float_status(float_status, float_status)
+
+            if len(param_grads_fp16):
+                _C_ops.check_finite_and_unscale(param_grads_fp16, self._scale,
+                                                float_status, param_grads_fp16,
+                                                self._temp_found_inf_fp16)
+            if len(param_grads_fp32):
+                _C_ops.check_finite_and_unscale(param_grads_fp32, self._scale,
+                                                float_status, param_grads_fp32,
+                                                self._temp_found_inf_fp32)
+        else:
+            if len(param_grads_fp16):
+                _C_ops.check_finite_and_unscale(param_grads_fp16, self._scale,
+                                                param_grads_fp16,
+                                                self._temp_found_inf_fp16)
+            if len(param_grads_fp32):
+                _C_ops.check_finite_and_unscale(param_grads_fp32, self._scale,
+                                                param_grads_fp32,
+                                                self._temp_found_inf_fp32)
+
         if len(param_grads_fp16) and len(param_grads_fp32):
             self._found_inf = self._temp_found_inf_fp16 or self._temp_found_inf_fp32
         elif len(param_grads_fp16):