[PT FE] Extract signature for FX scenario (#27014)

### Details:
 - *Extract signature from `forward`*
 - *Remove `gpt2` from hf tests as it was moved to `test_llm.py`*

### Tickets:
 - *ticket-id*

src/bindings/python/src/openvino/frontend/pytorch/fx_decoder.py

@@ -6,6 +6,7 @@
 
 import logging
 import torch
+import inspect
 
 from openvino.frontend.pytorch.py_pytorch_frontend import _FrontEndPytorchDecoder as Decoder
 from openvino.frontend.pytorch.py_pytorch_frontend import _Type as DecoderType
@@ -77,6 +78,10 @@ def __init__(self, pt_module, fx_gm=None, nodes=None, mark_node_callback=None, i
             if not input_types or len(input_types) == 0:
                 self.input_types = found_types
 
+            if hasattr(pt_module, "forward"):
+                input_params = inspect.signature(pt_module.forward).parameters
+                self._input_signature = list(input_params)
+
         elif issubclass(type(pt_module), torch.fx.Node):
 
             self._nodes = nodes  # passed from outer context

tests/layer_tests/py_frontend_tests/test_torch_decoder.py

@@ -667,3 +667,19 @@ def test_pytorch_decoder_can_convert_scripted_function():
     scripted = torch.jit.script(f)
     model = convert_model(scripted, input=[Type.f32, Type.f32])
     assert model is not None
+
+
+@pytest.mark.precommit
+def test_pytorch_fx_decoder_extracts_signature():
+    from openvino.frontend.pytorch.fx_decoder import TorchFXPythonDecoder
+
+    class TestModel(torch.nn.Module):
+        def forward(self, a, b):
+            return a["x"] + a["y"] + b
+
+    example = ({"x": torch.tensor(1), "y": torch.tensor(2)}, torch.tensor(3))
+    em = torch.export.export(TestModel(), example)
+    nc_decoder = TorchFXPythonDecoder(em.module())
+    assert nc_decoder.get_input_signature_name(0) == "a"
+    assert nc_decoder.get_input_signature_name(1) == "b"
+    assert nc_decoder._input_signature == ["a", "b"]

tests/model_hub_tests/pytorch/hf_transformers_models

@@ -79,7 +79,6 @@ git,microsoft/git-large-coco,xfail,Trace error: We don't have an op for aten::fu
 glpn,vinvino02/glpn-nyu
 gpt_bigcode,bigcode/tiny_starcoder_py
 gpt_neo,EleutherAI/gpt-neo-2.7B
-gpt2,openai-community/gpt2
 gptsan-japanese,Tanrei/GPTSAN-japanese,xfail,Unsupported op aten::index_put_ prim::TupleConstruct prim::TupleUnpack
 graphormer,clefourrier/graphormer-base-pcqm4mv2,xfail,Load error: GraphormerForGraphClassification.forward() missing 6 required positional arguments: 'input_edges' 'attn_bias' 'in_degree' 'out_degree' 'spatial_pos' and 'attn_edge_type'
 grounding-dino,IDEA-Research/grounding-dino-base,xfail,Trace error: op->outputs().size() == 1 INTERNAL ASSERT FAILED

tests/model_hub_tests/pytorch/test_hf_transformers.py

@@ -520,13 +520,20 @@ def load_model_with_default_class(name, **kwargs):
                                            ("bert-base-uncased", "bert"),
                                            ("google/flan-t5-base", "t5"),
                                            ("google/tapas-large-finetuned-wtq", "tapas"),
-                                           ("gpt2", "gpt2"),
                                            ("openai/clip-vit-large-patch14", "clip"),
                                            ])
     @pytest.mark.precommit
     def test_convert_model_precommit(self, name, type, ie_device):
         self.run(model_name=name, model_link=type, ie_device=ie_device)
 
+    @pytest.mark.parametrize("name,type", [("bert-base-uncased", "bert"),
+                                           ("openai/clip-vit-large-patch14", "clip"),
+                                           ])
+    @pytest.mark.precommit
+    def test_convert_model_precommit_export(self, name, type, ie_device):
+        self.mode = "export"
+        self.run(model_name=name, model_link=type, ie_device=ie_device)
+
     @pytest.mark.parametrize("type,name,mark,reason",
                              get_models_list(os.path.join(os.path.dirname(__file__), "hf_transformers_models")))
     @pytest.mark.nightly

tests/model_hub_tests/pytorch/torch_utils.py

@@ -68,28 +68,14 @@ def convert_model_impl(self, model_obj):
             from packaging import version
 
             model_obj.eval()
+            graph = None
             if isinstance(self.example, dict):
-                from openvino.frontend.pytorch.fx_decoder import TorchFXPythonDecoder
-
-                # need to infer before export to initialize everything, otherwise it will be initialized with FakeTensors
                 pt_res = model_obj(**self.example)
-                graph = export(model_obj, tuple(), self.example)
-                if version.parse(torch.__version__) >= version.parse("2.2"):
-                    from torch._decomp import get_decompositions
-                    from openvino.frontend.pytorch.torchdynamo.decompositions import get_export_decomposition_list
-                    decomp = get_decompositions(get_export_decomposition_list())
-                    graph = graph.run_decompositions(decomp_table=decomp)
-
-                gm = graph.module()
-                print(gm.code)
-
-                decoder = TorchFXPythonDecoder(gm)
-                decoder._input_signature = list(self.example.keys())
-                ov_model = convert_model(decoder, verbose=True)
+                graph = export(model_obj, args=tuple(), kwargs=self.example)
             else:
                 pt_res = model_obj(*self.example)
                 graph = export(model_obj, self.example)
-                ov_model = convert_model(graph, verbose=True)
+            ov_model = convert_model(graph, verbose=True)
 
             if isinstance(pt_res, dict):
                 for i, k in enumerate(pt_res.keys()):