diff --git a/paddle/fluid/operators/array_operator.h b/paddle/fluid/operators/array_operator.h
index bcf4c53947b780ba85e7dab2003345f8f319db9a..19b90d360201e928287d7ac2e2257f305f65daef 100644
--- a/paddle/fluid/operators/array_operator.h
+++ b/paddle/fluid/operators/array_operator.h
@@ -52,7 +52,8 @@ class ArrayOp : public framework::OperatorBase {
     size_t offset;
     if (platform::is_gpu_place(i_tensor.place()) ||
         platform::is_xpu_place(i_tensor.place()) ||
-        platform::is_npu_place(i_tensor.place())) {
+        platform::is_npu_place(i_tensor.place()) ||
+        platform::is_custom_place(i_tensor.place())) {
       // FIXME: Avoid copy from GPU to CPU
       framework::Tensor t;
       framework::TensorCopy(i_tensor, platform::CPUPlace(), dev_ctx, &t);
diff --git a/paddle/fluid/operators/reduce_ops/reduce_op.h b/paddle/fluid/operators/reduce_ops/reduce_op.h
index 2c275dcfc0a12efd5beae4cbb17e8ce261e3b6d4..ec3cf1908c5b5d430f510a9068d3ea7a979d680c 100644
--- a/paddle/fluid/operators/reduce_ops/reduce_op.h
+++ b/paddle/fluid/operators/reduce_ops/reduce_op.h
@@ -566,7 +566,8 @@ class ReduceOp : public framework::OperatorWithKernel {
       PADDLE_ENFORCE_EQ(
           platform::is_gpu_place(ctx.GetPlace()) ||
               platform::is_npu_place(ctx.GetPlace()) ||
-              platform::is_mlu_place(ctx.GetPlace()),
+              platform::is_mlu_place(ctx.GetPlace()) ||
+              platform::is_custom_place(ctx.GetPlace()),
           true,
           platform::errors::InvalidArgument(
               "float16 can only be used on GPU or NPU or MLU place"));
diff --git a/paddle/fluid/pybind/pybind.cc b/paddle/fluid/pybind/pybind.cc
index a49e11aba685153f57c2e33ff7ef45fe84b5937b..fef27358498e0922cf22ecb87ff51326518d1013 100644
--- a/paddle/fluid/pybind/pybind.cc
+++ b/paddle/fluid/pybind/pybind.cc
@@ -2801,6 +2801,7 @@ All parameter, weight, gradient are variables in Paddle.
       .def("_equals", &IsSamePlace<platform::Place, platform::IPUPlace>)
       .def("_equals", &IsSamePlace<platform::Place, platform::CUDAPinnedPlace>)
       .def("_equals", &IsSamePlace<platform::Place, platform::MLUPlace>)
+      .def("_equals", &IsSamePlace<platform::Place, platform::CustomPlace>)
       .def("is_gpu_place",
            [](platform::Place &self) { return platform::is_gpu_place(self); })
       .def("is_cpu_place",
diff --git a/python/paddle/device/__init__.py b/python/paddle/device/__init__.py
index 929a1c2d77fb4c3fae3a16a685bd7e7e1766e79a..4fcf9c5d21b263639fbd72173c0c1b3d4c57819e 100644
--- a/python/paddle/device/__init__.py
+++ b/python/paddle/device/__init__.py
@@ -349,6 +349,10 @@ def get_device():
     elif isinstance(place, core.MLUPlace):
         device_id = place.get_device_id()
         device = 'mlu:' + str(device_id)
+    elif isinstance(place, core.CustomPlace):
+        device_id = place.get_device_id()
+        device_type = place.get_device_type()
+        device = device_type + ':' + str(device_id)
     else:
         raise ValueError("The device specification {} is invalid".format(place))
 
diff --git a/python/paddle/fluid/tests/unittests/op_test.py b/python/paddle/fluid/tests/unittests/op_test.py
index ba694f535308391cc14b36ae6faf89b2fdd126bf..860a72193e2e42e007dd3cebec0e264325f94197 100644
--- a/python/paddle/fluid/tests/unittests/op_test.py
+++ b/python/paddle/fluid/tests/unittests/op_test.py
@@ -1685,7 +1685,8 @@ class OpTest(unittest.TestCase):
         # Currently not support ParallelExecutor on XPUPlace.
         if not paddle.is_compiled_with_xpu(
         ) and not paddle.is_compiled_with_npu(
-        ) and not paddle.is_compiled_with_mlu():
+        ) and not paddle.is_compiled_with_mlu() and not isinstance(
+                place, core.CustomPlace):
             self.check_inplace_output_with_place(place,
                                                  no_check_set=no_check_set,
                                                  inplace_atol=inplace_atol)