add deprecation warning to nn stateless functional_call (#87367)

Same as the release version but just for master

Pull Request resolved: https://github.com/pytorch/pytorch/pull/87367
Approved by: https://github.com/albanD, https://github.com/atalman

test/test_stateless.py

@@ -176,6 +176,37 @@ class TestStatelessFunctionalAPI(TestCase):
         self.assertEqual(orig_sn_weight, module.l1.weight)
 
 
+    def test_tied_weights_warns(self):
+        module = MockModule()
+        module.tied_bias = module.l1.bias
+        module.register_buffer("tied_buffer", module.buffer)
+        weight = torch.tensor([[1.0]],)
+        bias = torch.tensor([0.0])
+        buffer = torch.tensor([0.0])
+
+        parameters = {'l1.weight': weight,
+                      'l1.bias': bias,
+                      'buffer': buffer}
+        x = torch.randn(1, 1)
+        self.assertNotWarn(lambda: stateless.functional_call(module, parameters, x))
+
+        # if tied values are the same tensors, shouldn't warn
+        parameters['tied_bias'] = bias
+        parameters['tied_buffer'] = buffer
+        self.assertNotWarn(lambda: stateless.functional_call(module, parameters, x))
+        del parameters['tied_bias']
+        del parameters['tied_buffer']
+
+        with self.assertWarnsOnceRegex(UserWarning, "functional_call was passed multiple values"):
+            parameters['tied_bias'] = torch.tensor([5.0])
+            stateless.functional_call(module, parameters, x)
+        del parameters['tied_bias']
+
+        with self.assertWarnsOnceRegex(UserWarning, "functional_call was passed multiple values"):
+            parameters['tied_buffer'] = torch.tensor([5.0])
+            stateless.functional_call(module, parameters, x)
+
+
     def test_setattr(self):
         class Foo(torch.nn.Module):
             def __init__(self):

torch/nn/utils/stateless.py

@@ -1,3 +1,4 @@
+import warnings
 import contextlib
 from typing import Any, Callable, Dict, Iterator, List, Tuple
 
@@ -35,10 +36,22 @@ def _change_class(module, params_and_buffers) -> None:
     module.__class__ = param_cls
     module._orig_class = cls
 
-def _create_swap_params(params_and_buffers):
+
+def _check_tied_val_already_replaced(old_val, new_val, replaced_tensors_map):
+    if old_val not in replaced_tensors_map:
+        replaced_tensors_map[old_val] = new_val
+    elif replaced_tensors_map[old_val] is not new_val:
+        warnings.warn("functional_call was passed multiple values for tied weights. "
+                      "This behavior is deprecated and will be an error in future versions")
+
+
+def _create_swap_params(params_and_buffers, replaced_tensors_map):
     def _swap_parameters(module, tensor_name: str, full_path: str, tensor: Tensor) -> None:
         # Changes the module class to get a new __getattr__ dunder method
         # that looks for the reparametrized tensor
+        if hasattr(module, tensor_name):
+            old_val = getattr(module, tensor_name)
+            _check_tied_val_already_replaced(old_val, tensor, replaced_tensors_map)
         if hasattr(module, "_attr_to_path"):
             module._attr_to_path[tensor_name] = full_path
         else:
@@ -60,9 +73,10 @@ def _reparametrize_module(
     module: 'torch.nn.Module',
     parameters_and_buffers: Dict[str, Tensor],
 ) -> Iterator[None]:
+    orig_tensors_to_replacements: Dict[Tensor, Tensor] = {}
     for name, tensor in parameters_and_buffers.items():
         _apply_func_submodules(
-            _create_swap_params(parameters_and_buffers),
+            _create_swap_params(parameters_and_buffers, orig_tensors_to_replacements),
             module, name.split("."), name, (tensor,))
     try:
         yield