diff --git a/python/paddle/fluid/dataloader/dataloader_iter.py b/python/paddle/fluid/dataloader/dataloader_iter.py
index f248a1188bbee0d7811a63bcb21aed4a75965626..e47d75326b357fdccced84c9d3bffdfdf17ba068 100644
--- a/python/paddle/fluid/dataloader/dataloader_iter.py
+++ b/python/paddle/fluid/dataloader/dataloader_iter.py
@@ -21,7 +21,6 @@ import logging
 import itertools
 import threading
 import numpy as np
-import multiprocessing
 from collections import namedtuple
 from paddle.fluid.framework import (
     _set_expected_place,
@@ -422,6 +421,8 @@ class _DataLoaderIterMultiProcess(_DataLoaderIterBase):
         self._shutdown = False
 
     def _init_workers(self):
+        import paddle.incubate.multiprocessing as multiprocessing
+
         # multiprocess worker and indice queue list initial as empty
         self._workers = []
         self._worker_status = []
diff --git a/python/paddle/fluid/dataloader/worker.py b/python/paddle/fluid/dataloader/worker.py
index 155208791eacd2f828dbaf1d314617bf7accd8cf..64ec697cf5308ec7d4218d41ad5ea295c6d87696 100644
--- a/python/paddle/fluid/dataloader/worker.py
+++ b/python/paddle/fluid/dataloader/worker.py
@@ -373,21 +373,19 @@ def _worker_loop(
                     out_queue.put((idx, batch, None))
                 batch, structure = _flatten_batch(batch)
                 if use_shared_memory:
-                    # NOTE: In eager mode, Tensor._share_memory has no
-                    # effect, fall back to _array_to_share_memory_tensor
-                    def tensor_share_memory(tensor):
-                        if _in_eager_without_dygraph_check():
-                            return core._array_to_share_memory_tensor(tensor)
-                        return tensor._share_memory()
+
+                    def numpy2lodtensor(arr):
+                        lodtensor = core.Tensor()
+                        lodtensor.set(arr, core.CPUPlace())
+                        return lodtensor
 
                     tensor_list = [
-                        core._array_to_share_memory_tensor(b)
+                        numpy2lodtensor(b)
                         if isinstance(b, np.ndarray)
-                        else tensor_share_memory(b)
+                        else b.value().get_tensor()
                         for b in batch
                     ]
                     out_queue.put((idx, tensor_list, structure))
-                    core._remove_tensor_list_mmap_fds(tensor_list)
                 else:
                     out_queue.put((idx, batch, structure))
     except KeyboardInterrupt:
diff --git a/python/paddle/incubate/__init__.py b/python/paddle/incubate/__init__.py
index 0c5e90c6975e2944403b83af42d303e5dff42622..d5ff9454a8046ea6c2865b8bd523f99e13353216 100644
--- a/python/paddle/incubate/__init__.py
+++ b/python/paddle/incubate/__init__.py
@@ -34,6 +34,7 @@ from . import autograd  # noqa: F401
 from . import autotune  # noqa: F401
 from . import nn  # noqa: F401
 from . import asp  # noqa: F401
+from . import multiprocessing  # noqa: F401
 
 from ..fluid.layers.loss import identity_loss