Fix ResourceVariable placement during checkpointing to correctly colocate the

copy of the variable on the same machine. Addresses Issue #20914.

PiperOrigin-RevId: 205317119

tensorflow/python/training/saver.py

@@ -126,7 +126,12 @@ class BaseSaverBuilder(object):
           def f():
             with ops.device(v.device):
               x = v.read_value()
-            with ops.device("/device:CPU:0"):
+            # To allow variables placed on non-CPU devices to be checkpointed,
+            # we copy them to CPU on the same machine first.
+            device_spec = pydev.DeviceSpec().parse_from_string(v.device)
+            device_spec.merge_from(
+                pydev.DeviceSpec().parse_from_string("/device:CPU:0"))
+            with ops.device(device_spec.to_string()):
               return array_ops.identity(x)
           return f
 

tensorflow/python/training/saver_test.py

@@ -174,6 +174,24 @@ class SaverTest(test.TestCase):
   def testResourceBasic(self):
     self.basicSaveRestore(resource_variable_ops.ResourceVariable)
 
+  def testResourceColocation(self):
+    partitioner = partitioned_variables.fixed_size_partitioner(num_shards=2)
+    with ops_lib.device("/job:ps/device:GPU:0"):
+      v = variable_scope.get_variable("v0",
+                                      shape=[10, 2],
+                                      partitioner=partitioner,
+                                      use_resource=True)
+    saver_module.Saver({"v0": v}).build()
+    save_op = None
+    for op in ops_lib.get_default_graph().get_operations():
+      if op.type == "SaveV2":
+        save_op = op
+        break
+    assert save_op is not None
+    for save_inp in save_op.inputs[3:]:
+      # Input to SaveV2 op is placed on CPU of the same device as the Variable.
+      self.assertEqual("/job:ps/device:CPU:0", save_inp.device)
+
   def testResourceVariableReadOpsAddedDeterministically(self):
     graph_defs = []
     num_graphs = 10