[PT2] no use gradient for building graph (#3278)

### Changes

Use `torch.no_grad()` for build graph
Affect: 
- unknown type for `tenosr.T` and `tensor.H` operation, but graph is
still connected
- removed `requires_grad` attribute for tensor, that always false for
no_grad, but it's not used anywhere

### Reason for changes

High memory usage

Author: AlexanderDokuchaev

nncf/experimental/torch2/function_hook/graph/build_graph_mode.py

@@ -212,14 +212,15 @@ def execute_post_hooks(self, outputs: Any, op_meta: OpMeta) -> Any:
         Overload execute_post_hooks to correct registered node for operation.
         Process __get__ function, to detect permute and transpose operation
         and remove node if operation return not tensor.
+
+        :param output: The output of the function.
+        :param op_meta: Metadata for the operation.
+        :return: The modified output after post-hooks.
         """
-        if op_meta.func.__name__ == "__get__":
-            if isinstance(outputs, torch.Tensor):
-                self.process_tensor_attributes(outputs, op_meta)
-            else:
-                # Remove the node corresponding to this operation from the graph, as non-tensor
-                # outputs (like `tensor.shape` or similar) are not relevant for further algorithmic use.
-                self.graph.remove_node(op_meta.extra_info["node_id"])
+        if op_meta.func.__name__ == "__get__" and not isinstance(outputs, torch.Tensor):
+            # Remove the node corresponding to this operation from the graph, as non-tensor
+            # outputs (like `tensor.shape` or similar) are not relevant for further algorithmic use.
+            self.graph.remove_node(op_meta.extra_info["node_id"])
         outputs = super().execute_post_hooks(outputs, op_meta)
         return outputs
 
@@ -368,8 +369,7 @@ def build_graph(model: nn.Module, *args: Any, **kwargs: Any) -> nx.MultiDiGraph:
     :return: A nx.MultiDiGraph where nodes represent operations of model.
     """
     with training_mode_switcher(model, is_training=False):
-        with torch.enable_grad():  # type: ignore
-            # Gradient use to get information about __get__ functions to detect tensor.(T, mT) attributes
+        with torch.no_grad():
             with GraphBuilderMode(model=model, hook_storage=get_hook_storage(model)) as ctx:
                 args, kwargs = ctx.process_model_inputs(args, kwargs)
                 wrapped_forward = cast(ForwardWithHooks, model.forward)

nncf/experimental/torch2/function_hook/graph/graph_utils.py

@@ -43,11 +43,10 @@ def __str__(self) -> str:
 class TensorMeta:
     dtype: torch.dtype
     shape: Tuple[int, ...]
-    requires_grad: bool
 
     @staticmethod
     def from_tensor(tensor: torch.Tensor) -> TensorMeta:
-        return TensorMeta(tensor.dtype, tuple(tensor.shape), tensor.requires_grad)
+        return TensorMeta(tensor.dtype, tuple(tensor.shape))
 
 
 @dataclass

tests/torch2/data/function_hook/graph_visualization/to_pydot_style_full.dot

@@ -3,10 +3,10 @@ rankdir=TB;
 0 [fillcolor="#adadad", fontcolor="#000000", label="{type: input|name: x|dtype: torch.float32|shape: (1, 1, 3, 3)}", shape=record, style="filled,rounded"];
 1 [fillcolor="#ffffff", fontcolor="#000000", label="{type: const|name: conv.weight|dtype: torch.float32|shape: (1, 1, 1, 1)}", shape=record, style="filled,rounded"];
 2 [fillcolor="#ffffff", fontcolor="#000000", label="{type: const|name: conv.bias|dtype: torch.float32|shape: (1,)}", shape=record, style="filled,rounded"];
-3 [fillcolor="#ffd6a5", fontcolor="#000000", label="{type: function_call|op_name: conv/conv2d/0|fn_name: conv2d|args: [\nTensorMeta(dtype=torch.float32, shape=(1, 1, 3, 3), requires_grad=False),\nTensorMeta(dtype=torch.float32, shape=(1, 1, 1, 1), requires_grad=True),\nTensorMeta(dtype=torch.float32, shape=(1,), requires_grad=True),\n(1, 1),\n(0, 0),\n(1, 1),\n1,\n]|kwargs: \{\}}", shape=record, style="filled,rounded"];
+3 [fillcolor="#ffd6a5", fontcolor="#000000", label="{type: function_call|op_name: conv/conv2d/0|fn_name: conv2d|args: [\nTensorMeta(dtype=torch.float32, shape=(1, 1, 3, 3)),\nTensorMeta(dtype=torch.float32, shape=(1, 1, 1, 1)),\nTensorMeta(dtype=torch.float32, shape=(1,)),\n(1, 1),\n(0, 0),\n(1, 1),\n1,\n]|kwargs: \{\}}", shape=record, style="filled,rounded"];
 4 [fillcolor="#ffffff", fontcolor="#000000", label="{type: const|name: __nncf_hooks.post_hooks.conv/conv2d/0__0.0.w|dtype: torch.float32|shape: (1,)}", shape=record, style="filled,rounded"];
-5 [fillcolor="#caffbf", fontcolor="#000000", label="{type: function_call|op_name: conv/post_hook__conv-conv2d-0__0[0]/add/0|fn_name: add|args: [\nTensorMeta(dtype=torch.float32, shape=(1, 1, 3, 3), requires_grad=True),\nTensorMeta(dtype=torch.float32, shape=(1,), requires_grad=True),\n]|kwargs: \{\}}", shape=record, style="filled,rounded"];
-6 [fillcolor="#a0c4ff", fontcolor="#000000", label="{type: function_call|op_name: /relu/0|fn_name: relu|args: [\nTensorMeta(dtype=torch.float32, shape=(1, 1, 3, 3), requires_grad=True),\n]|kwargs: \{\}}", shape=record, style="filled,rounded"];
+5 [fillcolor="#caffbf", fontcolor="#000000", label="{type: function_call|op_name: conv/post_hook__conv-conv2d-0__0[0]/add/0|fn_name: add|args: [\nTensorMeta(dtype=torch.float32, shape=(1, 1, 3, 3)),\nTensorMeta(dtype=torch.float32, shape=(1,)),\n]|kwargs: \{\}}", shape=record, style="filled,rounded"];
+6 [fillcolor="#a0c4ff", fontcolor="#000000", label="{type: function_call|op_name: /relu/0|fn_name: relu|args: [\nTensorMeta(dtype=torch.float32, shape=(1, 1, 3, 3)),\n]|kwargs: \{\}}", shape=record, style="filled,rounded"];
 7 [fillcolor="#adadad", fontcolor="#000000", label="{type: output|name: output|dtype: torch.float32|shape: (1, 1, 3, 3)}", shape=record, style="filled,rounded"];
 0 -> 3  [label="(1, 1, 3, 3)\n0 → 0"];
 1 -> 3  [label="(1, 1, 1, 1)\n0 → 1"];

tests/torch2/data/function_hook/handle_inner_functions/inner_functions_BatchNormModel.dot

@@ -5,7 +5,7 @@ rankdir=TB;
 2 [fillcolor="#ffffff", fontcolor="#000000", label="{type: const|name: bn.bias|dtype: torch.float32|shape: (1,)}", shape=record, style="filled,rounded"];
 3 [fillcolor="#ffffff", fontcolor="#000000", label="{type: const|name: bn.running_mean|dtype: torch.float32|shape: (1,)}", shape=record, style="filled,rounded"];
 4 [fillcolor="#ffffff", fontcolor="#000000", label="{type: const|name: bn.running_var|dtype: torch.float32|shape: (1,)}", shape=record, style="filled,rounded"];
-5 [fillcolor="#ffadad", fontcolor="#000000", label="{type: function_call|op_name: bn/batch_norm/0|fn_name: batch_norm|args: [\nTensorMeta(dtype=torch.float32, shape=(1, 1, 1), requires_grad=False),\nTensorMeta(dtype=torch.float32, shape=(1,), requires_grad=True),\nTensorMeta(dtype=torch.float32, shape=(1,), requires_grad=True),\nTensorMeta(dtype=torch.float32, shape=(1,), requires_grad=False),\nTensorMeta(dtype=torch.float32, shape=(1,), requires_grad=False),\nFalse,\n0.1,\n1e-05,\nTrue,\n]|kwargs: \{\}}", shape=record, style="filled,rounded"];
+5 [fillcolor="#ffadad", fontcolor="#000000", label="{type: function_call|op_name: bn/batch_norm/0|fn_name: batch_norm|args: [\nTensorMeta(dtype=torch.float32, shape=(1, 1, 1)),\nTensorMeta(dtype=torch.float32, shape=(1,)),\nTensorMeta(dtype=torch.float32, shape=(1,)),\nTensorMeta(dtype=torch.float32, shape=(1,)),\nTensorMeta(dtype=torch.float32, shape=(1,)),\nFalse,\n0.1,\n1e-05,\nTrue,\n]|kwargs: \{\}}", shape=record, style="filled,rounded"];
 6 [fillcolor="#adadad", fontcolor="#000000", label="{type: output|name: output|dtype: torch.float32|shape: (1, 1, 1)}", shape=record, style="filled,rounded"];
 0 -> 5  [label="(1, 1, 1)\n0 → 0"];
 1 -> 5  [label="(1,)\n0 → 1"];