Do not wrap Tensor.{grad,_base} by default (#60464)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/60464

Fixes https://github.com/szagoruyko/pytorchviz/issues/65

An alternate implementation of this PR would be to remove the
__torch_function__ interposition points for these accessors entirely.
In the end, I decided to opt for extra expressivity.  See
torch.overrides for the criterion on how I decided which accessors
should get the nowrap treatment.

Signed-off-by: Edward Z. Yang <ezyang@fb.com>

Test Plan: Imported from OSS

Reviewed By: albanD

Differential Revision: D29302835

Pulled By: ezyang

fbshipit-source-id: fbe0ac4530a6cc9d6759a3fdf5514d4d7b1f7690

test/test_overrides.py

@@ -540,6 +540,16 @@ class TestTorchFunctionOverride(TestCase):
         with self.assertRaises(TypeError):
             sn1 + s2
 
+    def test_base(self):
+        # https://github.com/szagoruyko/pytorchviz/issues/65
+        class DummyTensor(torch.Tensor):
+            pass
+
+        a = torch.ones(1)
+        c = DummyTensor(a)
+        self.assertTrue(c._is_view())
+        self.assertTrue(c._base is a)
+
 
 def generate_tensor_like_override_tests(cls):
     from torch.testing._internal.generated.annotated_fn_args import annotated_args

torch/_tensor.py

@@ -11,7 +11,7 @@ from torch._namedtensor_internals import (
     unzip_namedshape, single_ellipsis_index, is_ellipsis)
 from torch.overrides import (
     has_torch_function, has_torch_function_unary, has_torch_function_variadic,
-    handle_torch_function)
+    handle_torch_function, get_default_nowrap_functions)
 import torch.utils.hooks as hooks
 
 
@@ -1007,7 +1007,10 @@ class Tensor(torch._C._TensorBase):
 
         with _C.DisableTorchFunction():
             ret = func(*args, **kwargs)
-            return _convert(ret, cls)
+            if func in get_default_nowrap_functions():
+                return ret
+            else:
+                return _convert(ret, cls)
 
     __module__ = 'torch'
 

torch/overrides.py

@@ -233,6 +233,32 @@ def get_ignored_functions() -> Set[Callable]:
     }
 
 
+@functools.lru_cache(None)
+def get_default_nowrap_functions() -> Set[Callable]:
+    """
+    Return public functions that do not wrap in a subclass when invoked by
+    the default ``Tensor.__torch_function__`` that preserves subclasses.  Typically,
+    these functions represent field accesses (i.e., retrieving a Tensor that
+    is stored somewhere on the Tensor) as opposed to computation.  Users of
+    these functions expect object identity to be preserved over multiple accesses
+    (e.g., ``a.grad is a.grad``) which cannot be upheld if we're wrapping on
+    the fly every time (furthermore, the tensor stored here might already be
+    the subclass, in which case wrapping really ought not to happen).
+
+    Not ALL property accessors have this property; for example ``Tensor.T`` actually
+    just creates a new transposed tensor on the fly, and so we SHOULD interpose on
+    these calls (you need to check the implementation of the function to see if
+    this is the case or not).  Additionally, if a property accessor doesn't return a Tensor,
+    it doesn't have to be on this list (though it is harmless if it is).
+    """
+    Tensor = torch.Tensor
+    return {
+        Tensor._base.__get__,
+        Tensor.grad.__get__,
+        Tensor._grad.__get__,
+    }
+
+
 @functools.lru_cache(None)
 def get_testing_overrides() -> Dict[Callable, Callable]:
     """Return a dict containing dummy overrides for all overridable functions