Update the benchmarking script to clone an eager model (#113046)

Summary:
fix pytorch/pytorch#113029 where running a model in eager somehow can change a weight stride

X-link: pytorch/pytorch#113046
Approved by: https://github.com/angelayi

Reviewed By: PaliC

Differential Revision: D51139962

Pulled By: desertfire

fbshipit-source-id: 3ae8a878cd408c9e20e1b5846f0086b540ff17e4

userbenchmark/dynamo/dynamobench/common.py

@@ -1175,7 +1175,10 @@ def load(cls, model, example_inputs, device):
         if key not in cls.cache:
             # Register the output dataclass to pytree
             example_args, example_kwargs = _normalize_bench_inputs(example_inputs)
-            example_outputs = model(*example_args, **example_kwargs)
+            with torch.no_grad():
+                # copy.deepcopy is required to prevent any surprising side-effect,
+                # see https://github.com/pytorch/pytorch/issues/113029
+                example_outputs = copy.deepcopy(model)(*example_args, **example_kwargs)
             _register_dataclass_output_as_pytree(example_outputs)
 
             so_path = torch._export.aot_compile(model, example_args, example_kwargs)