Make sure requires_grad is propagated for all backend

The if statement is not strictly necessary but that avoid having to call this function if we don't need it.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/76256
Approved by: https://github.com/ezyang, https://github.com/soulitzer

test/test_serialization.py

@@ -20,8 +20,8 @@ from torch._utils_internal import get_file_path_2
 from torch._utils import _rebuild_tensor
 from torch.serialization import check_module_version_greater_or_equal
 
-from torch.testing._internal.common_utils import TestCase, IS_WINDOWS, \
-    TEST_DILL, run_tests, download_file, BytesIOContext, TemporaryFileName
+from torch.testing._internal.common_utils import TestCase, IS_WINDOWS, TEST_DILL, \
+    run_tests, download_file, BytesIOContext, TemporaryFileName, parametrize, instantiate_parametrized_tests
 from torch.testing._internal.common_device_type import instantiate_device_type_tests
 from torch.testing._internal.common_dtype import all_types_and_complex_and
 
@@ -948,8 +948,18 @@ class TestSubclassSerialization(TestCase):
         self.assertEqual(new_tensor.elem, my_tensor.elem)
         self.assertEqual(new_tensor.foo, foo_val)
 
+    @parametrize('requires_grad', (True, False))
+    def test_cloned_deepcopy(self, requires_grad):
+        my_tensor = torch.rand(2, requires_grad=requires_grad, device='meta')
+
+        new_tensor = deepcopy(my_tensor)
+
+        self.assertEqual(new_tensor.requires_grad, my_tensor.requires_grad)
+
+
 
 instantiate_device_type_tests(TestBothSerialization, globals())
+instantiate_parametrized_tests(TestSubclassSerialization)
 
 if __name__ == '__main__':
     run_tests()

torch/_tensor.py

@@ -95,6 +95,7 @@ class Tensor(torch._C._TensorBase):
             # does accurate alias tracking; however, the code below
             # doesn't work because of
             # https://github.com/pytorch/pytorch/issues/47442
+            # Update the test in test_serialization if you remove 'meta' from here
             if self.is_sparse or self.device.type in ['lazy', 'xla', 'mlc', 'ort', 'meta', 'hpu'] or \
                     (type(self) is not Tensor and self.data_ptr() == 0):
                 new_tensor = self.clone()
@@ -150,7 +151,8 @@ class Tensor(torch._C._TensorBase):
                         new_tensor = new_tensor.conj_physical()
                     if self.is_neg():
                         new_tensor = new_tensor.neg()
-                    new_tensor.requires_grad = self.requires_grad
+            if self.requires_grad:
+                new_tensor.requires_grad_()
             if self.grad is not None:
                 new_tensor.grad = self.grad.__deepcopy__(memo)