Add non-recursive module.to_empty option

test/test_nn.py

@@ -12314,6 +12314,39 @@ def forward(self, x):
         m.to_empty(device='meta')
         m(input)
 
+    def test_module_to_empty_non_recursive(self, device):
+        class Layer(nn.Module):
+            def __init__(self, in_features, out_features):
+                super().__init__()
+                self.weight = nn.Parameter(torch.randn(in_features, out_features))
+                self.register_buffer('buf', torch.randn(out_features))
+
+            def forward(self, x):
+                return x @ self.weight + self.buf
+
+        class MyModule(nn.Module):
+            def __init__(self, in_features, out_features):
+                super().__init__()
+                self.weight = nn.Parameter(torch.randn(in_features, out_features))
+                self.register_buffer('buf1', torch.randn(out_features))
+                self.layer = Layer(out_features, out_features)
+
+            def forward(self, x):
+                return self.layer(x @ self.weight + self.buf1)
+
+        with torch.device('meta'):
+            m = MyModule(3, 5)
+
+        m.to_empty(device=device, recurse=False)
+
+        # params/buffers of parent should have been materialized on device
+        self.assertTrue(not m.weight.is_meta)
+        self.assertTrue(not m.buf1.is_meta)
+
+        # parameters/buffers of children submodules should still be on meta
+        for p in (*m.layer.parameters(), *m.layer.buffers()):
+            self.assertTrue(p.is_meta)
+
     @skipMeta
     def test_skip_init(self, device):
         torch.manual_seed(1)

torch/nn/cpp.py

@@ -65,7 +65,7 @@ def __init__(self, cpp_module):
             if not attr.startswith("_"):
                 setattr(self, attr, getattr(self.cpp_module, attr))
 
-    def _apply(self, fn):
+    def _apply(self, fn, recurse=True):
         for param in self.parameters():
             # Tensors stored in modules are graph leaves, and we don't
             # want to create copy nodes, so we have to unpack the data.

torch/nn/modules/module.py

@@ -796,9 +796,10 @@ def set_extra_state(self, state: Any):
             "Please file an issue at https://github.com/pytorch/pytorch/issues/new?template=bug-report.yml "
             "to report this bug.")
 
-    def _apply(self, fn):
-        for module in self.children():
-            module._apply(fn)
+    def _apply(self, fn, recurse=True):
+        if recurse:
+            for module in self.children():
+                module._apply(fn)
 
         def compute_should_use_set_data(tensor, tensor_applied):
             if torch._has_compatible_shallow_copy_type(tensor, tensor_applied):
@@ -1015,17 +1016,19 @@ def bfloat16(self: T) -> T:
         """
         return self._apply(lambda t: t.bfloat16() if t.is_floating_point() else t)
 
-    def to_empty(self: T, *, device: Union[str, device]) -> T:
+    def to_empty(self: T, *, device: Union[str, device], recurse: bool = True) -> T:
         r"""Moves the parameters and buffers to the specified device without copying storage.
 
         Args:
             device (:class:`torch.device`): The desired device of the parameters
                 and buffers in this module.
+            recurse (bool): Whether parameters and buffers of submodules should
+                be recursively moved to the specified device.
 
         Returns:
             Module: self
         """
-        return self._apply(lambda t: torch.empty_like(t, device=device))
+        return self._apply(lambda t: torch.empty_like(t, device=device), recurse=recurse)
 
     @overload
     def to(self: T, device: Optional[Union[int, device]] = ..., dtype: Optional[Union[dtype, str]] = ...,

torch/nn/modules/rnn.py

@@ -193,8 +193,8 @@ def flatten_parameters(self) -> None:
                         self.hidden_size, self.proj_size, self.num_layers,
                         self.batch_first, bool(self.bidirectional))
 
-    def _apply(self, fn):
-        ret = super()._apply(fn)
+    def _apply(self, fn, recurse=True):
+        ret = super()._apply(fn, recurse)
 
         # Resets _flat_weights
         # Note: be v. careful before removing this, as 3rd party device types