diff --git a/mindspore/ops/operations/_quant_ops.py b/mindspore/ops/operations/_quant_ops.py
index ad1c90d19a..f6df84f18e 100644
--- a/mindspore/ops/operations/_quant_ops.py
+++ b/mindspore/ops/operations/_quant_ops.py
@@ -467,7 +467,10 @@ class FakeQuantPerLayer(PrimitiveWithInfer):
         return x_shape
 
     def infer_dtype(self, x_type, min_type, max_type):
-        valid_types = (mstype.float16, mstype.float32)
+        if context.get_context('device_target') == "GPU":
+            valid_types = (mstype.float32,)
+        else:
+            valid_types = (mstype.float16, mstype.float32)
         validator.check_tensor_type_same({"x": x_type}, valid_types, self.name)
         validator.check_tensor_type_same(
             {"min": min_type}, valid_types, self.name)
@@ -521,7 +524,10 @@ class FakeQuantPerLayerGrad(PrimitiveWithInfer):
         return dout_shape
 
     def infer_dtype(self, dout_type, x_type, min_type, max_type):
-        valid_types = (mstype.float16, mstype.float32)
+        if context.get_context('device_target') == "GPU":
+            valid_types = (mstype.float32,)
+        else:
+            valid_types = (mstype.float16, mstype.float32)
         validator.check_tensor_type_same(
             {"dout": dout_type}, valid_types, self.name)
         validator.check_tensor_type_same({"x": x_type}, valid_types, self.name)
@@ -616,7 +622,10 @@ class FakeQuantPerChannel(PrimitiveWithInfer):
         return x_shape
 
     def infer_dtype(self, x_type, min_type, max_type):
-        valid_types = (mstype.float16, mstype.float32)
+        if context.get_context('device_target') == "GPU":
+            valid_types = (mstype.float32,)
+        else:
+            valid_types = (mstype.float16, mstype.float32)
         validator.check_tensor_type_same({"x": x_type}, valid_types, self.name)
         validator.check_tensor_type_same(
             {"min": min_type}, valid_types, self.name)
@@ -670,7 +679,10 @@ class FakeQuantPerChannelGrad(PrimitiveWithInfer):
         return dout_shape
 
     def infer_dtype(self, dout_type, x_type, min_type, max_type):
-        valid_types = (mstype.float16, mstype.float32)
+        if context.get_context('device_target') == "GPU":
+            valid_types = (mstype.float32,)
+        else:
+            valid_types = (mstype.float16, mstype.float32)
         validator.check_tensor_type_same(
             {"dout": dout_type}, valid_types, self.name)
         validator.check_tensor_type_same({"x": x_type}, valid_types, self.name)