Executorch does not seem to be able to accelerate Anomaly Transformer inference

I want to use executorch to deploy the anomaly transformer model https://github.com/thuml/Anomaly-Transformer trained with pytorch to the Raspberry Pi 5. When converting the pth model to a pte model, the following warning appears:

/home/taylor/miniconda3/envs/executorch/lib/python3.10/site-packages/torch/export/_unlift.py:75: UserWarning: Attempted to insert a get_attr Node with no underlying reference in the owning GraphModule! Call GraphModule.add_submodule to add the necessary submodule, GraphModule.add_parameter to add the necessary Parameter, or nn.Module.register_buffer to add the necessary buffer
  getattr_node = gm.graph.get_attr(lifted_node)
/home/taylor/miniconda3/envs/executorch/lib/python3.10/site-packages/torch/fx/graph.py:1801: UserWarning: Node original_model_encoder_attn_layers_0_attention_inner_attention_distances target original_model.encoder.attn_layers.0.attention.inner_attention.distances distances of original_model.encoder.attn_layers.0.attention.inner_attention does not reference an nn.Module, nn.Parameter, or buffer, which is what 'get_attr' Nodes typically target
  warnings.warn(
/home/taylor/miniconda3/envs/executorch/lib/python3.10/site-packages/torch/fx/graph.py:1801: UserWarning: Node original_model_encoder_attn_layers_1_attention_inner_attention_distances target original_model.encoder.attn_layers.1.attention.inner_attention.distances distances of original_model.encoder.attn_layers.1.attention.inner_attention does not reference an nn.Module, nn.Parameter, or buffer, which is what 'get_attr' Nodes typically target
  warnings.warn(
/home/taylor/miniconda3/envs/executorch/lib/python3.10/site-packages/torch/fx/graph.py:1801: UserWarning: Node original_model_encoder_attn_layers_2_attention_inner_attention_distances target original_model.encoder.attn_layers.2.attention.inner_attention.distances distances of original_model.encoder.attn_layers.2.attention.inner_attention does not reference an nn.Module, nn.Parameter, or buffer, which is what 'get_attr' Nodes typically target
  warnings.warn(

Then I followed the steps of https://pytorch.org/executorch/stable/getting-started-setup.html to slightly modify the executor_runner.cpp code to record the time, and then run the pte model. The running time is as follows

I 00:00:01.327281 executorch:executor_runner.cpp:84] Model file SWaT.pte is loaded.
I 00:00:01.327354 executorch:executor_runner.cpp:93] Using method forward
I 00:00:01.327406 executorch:executor_runner.cpp:140] Setting up planned buffer 0, size 8239104.
I 00:00:01.346501 executorch:executor_runner.cpp:163] Method loaded.
method load time : 2.39548 ms
I 00:00:01.346714 executorch:executor_runner.cpp:178] Inputs prepared.
input prepared time : 0.102407 ms
I 00:00:14.292156 executorch:executor_runner.cpp:192] Model executed successfully.
model run time : 12945.4 ms
I 00:00:14.292199 executorch:executor_runner.cpp:199] 1 outputs: 
Output 0: tensor(sizes=[100], [
  0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 
  0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 
  0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 
  0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 
  0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 
  0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 
  0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 
  0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 
  0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 
  0., 0., 1.41644e-34, 2.33109e-21, 6.68264e-10, 0.0322331, 0.350365, 1.78738e-08, 4.11204e-23, 2.78298e-42, 
])

You can see that it takes about 13 seconds to infer once with the pte model, while it only takes 0.3 seconds to run the pth model with pytorch. I don't understand why this happens? Could this be the problem with this warning?

cc @mergennachin @cccclai @helunwencser @jackzhxng @byjlw

Reproduce the Anomaly Transformer pth-to-pte conversion using the linked ExecuTorch setup guide and record the reported warnings. Then inspect the modified executor_runner.cpp timing around method loading, input preparation, and execution on Raspberry Pi 5; done means identifying the cause of the 13-second inference time and whether the warnings or conversion affect the output.