[Fix]: Convert operator that does specialization to its symbolic counterpart (pytorch#129578)

#### Issue
During conversion, use symbolic operator when exist.

#### Test Plan
`pytest test/export/test_converter.py`
Pull Request resolved: pytorch#129578
Approved by: https://github.com/angelayi

Author: jiashenC

test/export/test_converter.py

@@ -797,7 +797,18 @@ def forward(self, x):
             torch.randn([2, 3, 4]).to(torch.float32),
             torch.randn([2, 3, 4]).to(torch.float64),
         )
-        self._check_equal_ts_ep_converter(func6, inp)
+        ep_list = self._check_equal_ts_ep_converter(func6, inp)
+
+        # TODO: Additional check once dynamic shape is supported.
+        # for ep in ep_list:
+        #     self.assertEqual(
+        #         ep.module()(
+        #             torch.randn([1, 1, 1]).to(torch.int8),
+        #             torch.randn([1, 1, 1]).to(torch.int32),
+        #             torch.randn([1, 1, 1]).to(torch.float32),
+        #             torch.randn([1, 1, 1]).to(torch.float64),
+        #         )[0], 1
+        #     )
 
     def test_prim_tolist(self):
         class Module(torch.nn.Module):

torch/_export/converter.py

@@ -107,6 +107,12 @@ def get_dtype_as_int(tensor):
     "aten::__contains__": operator.contains,
     "prim::dtype": get_dtype_as_int,
     "aten::len": len,
+    # Mapping from specialized op to its symbolic counterpart.
+    # They currently do not have any other overrides.
+    "aten::numel": torch.ops.aten.sym_numel,
+    "aten::size": torch.ops.aten.sym_size,
+    "aten::storage_offset": torch.ops.aten.sym_storage_offset,
+    "aten::stride": torch.ops.aten.sym_stride,
 }
 
 
@@ -497,9 +503,6 @@ def convert_prim_SetAttr(self, node: torch._C.Node):
     def convert_call_function_op(self, node: torch._C.Node):
         target = get_op_overload(node)
 
-        if target is torch.ops.aten.size.int:
-            target = torch.ops.aten.sym_size.int
-
         args, kwargs = self.get_args_kwargs(node, target._schema)
 
         fx_node = self.fx_graph.call_function(target, args, kwargs)