[PT] detach tensor in get_const_data

nncf/quantization/algorithms/weight_compression/torch_backend.py

@@ -38,6 +38,7 @@
 from nncf.torch.graph.transformations.commands import PTSharedFnInsertionCommand
 from nncf.torch.graph.transformations.commands import PTTargetPoint
 from nncf.torch.model_graph_manager import find_const_node_in_constant_subgraph
+from nncf.torch.model_graph_manager import get_const_data
 from nncf.torch.model_graph_manager import get_const_node
 from nncf.torch.model_graph_manager import get_module_by_name
 from nncf.torch.model_graph_manager import split_const_name
@@ -173,10 +174,8 @@ def get_weight(
     ) -> Tensor:
         weight_node = get_const_node(node_with_weight, weight_port_id, graph)
         weight_name = weight_node.layer_attributes.name
-        module_name, weight_attr_name = split_const_name(weight_name)
-        module = get_module_by_name(module_name, model)
-        weight = getattr(module, weight_attr_name)
-        if weight is None or not isinstance(weight, torch.nn.Parameter):
+        weight = get_const_data(weight_node, model)
+        if weight is None:
             raise nncf.InternalError(f"Could not find a torch.nn.Parameter in the model by name {weight_name}.")
 
         return Tensor(weight)
@@ -222,10 +221,8 @@ def transform_model(
 
             weight_node = get_const_node(wc_params.node_with_weight, wc_params.weight_port_id, graph)
             weight_name = weight_node.layer_attributes.name
-            module_name, weight_attr_name = split_const_name(weight_name)
-            module = get_module_by_name(module_name, model)
-            weight = getattr(module, weight_attr_name)
-            if weight is None or not isinstance(weight, torch.nn.Parameter):
+            weight = get_const_data(weight_node, model)
+            if weight is None:
                 raise nncf.InternalError(f"Could not find a torch.nn.Parameter in the model by name {weight_name}.")
 
             # calculates compressed weights and decompression parameters
@@ -264,7 +261,14 @@ def transform_model(
             packed_tensor = decompressor.pack_weight(compressed_weight.tensor.data)
 
             # sets compressed tensor
+            # TODO:(AlexanderDokuchaev): update set_const_data
             compressed_parameter = torch.nn.Parameter(packed_tensor, requires_grad=False)
+            module_name, weight_attr_name = split_const_name(weight_name)
+            module = get_module_by_name(module_name, model)
+            weight = getattr(module, weight_attr_name)
+            if not isinstance(weight, torch.nn.Parameter):
+                raise nncf.InternalError(f"Weight is not a torch.nn.Parameter in the model by name {weight_name}.")
+
             setattr(module, weight_attr_name, compressed_parameter)
 
             consumer_nodes = graph.get_next_nodes(weight_node)

nncf/torch/model_graph_manager.py

@@ -117,7 +117,7 @@ def get_module_by_name(module_name: str, model: torch.nn.Module) -> torch.nn.Mod
 
 def get_const_data(const_node: NNCFNode, model: NNCFNetwork) -> torch.Tensor:
     """
-    Retrieves a constant tensor associated with a given node.
+    Retrieves a detached constant tensor associated with a given node.
 
     :param const_node: The node associated with const data.
     :param model: The NNCFNetwork object.
@@ -128,8 +128,8 @@ def get_const_data(const_node: NNCFNode, model: NNCFNetwork) -> torch.Tensor:
     module = get_module_by_name(module_name, model)
     data = getattr(module, const_attr_name)
     if isinstance(data, torch.nn.Parameter):
-        return data.data
-    return data
+        return data.data.detach()
+    return data.detach()
 
 
 def get_const_data_on_port(node: NNCFNode, port_id: int, model: NNCFNetwork) -> torch.Tensor:

tests/torch/test_model_graph_manager.py

@@ -239,10 +239,15 @@ def test_get_set_const_data():
     graph = model.nncf.get_graph()
     const_node = graph.get_node_by_name("conv.bias")
 
+    assert model.conv.bias.requires_grad
+
     data = get_const_data(const_node, model)
+    assert not data.requires_grad
     assert torch.all(model.conv.bias.data == data)
+
     set_const_data(torch.ones_like(data), const_node, model)
     assert torch.all(model.conv.bias.data == torch.ones_like(data))
+    assert model.conv.bias.requires_grad
 
 
 @pytest.mark.parametrize(