Arm backend: Fix Arm tester issue for inplace ops (pytorch#14625)

Deep-copying the input avoids it getting mutated by the first reference
run.

Author: mansnils

backends/arm/test/ops/test_silu.py

@@ -8,7 +8,6 @@
 
 from typing import Optional, Tuple
 
-import pytest
 import torch
 from executorch.backends.arm.test import common
 from executorch.backends.arm.test.tester.test_pipeline import (
@@ -125,7 +124,6 @@ def test_silu_u85_INT_inplace(test_data: input_t):
 
 @common.parametrize("test_data", Silu.test_data)
 @common.SkipIfNoModelConverter
-@pytest.mark.xfail(reason="MLETORCH-1387: Output differs")
 def test_silu_vgf_FP(test_data: input_t):
     silu_data = (test_data(), False)
     pipeline = VgfPipeline[input_t](
@@ -136,7 +134,6 @@ def test_silu_vgf_FP(test_data: input_t):
 
 @common.parametrize("test_data", Silu.test_data)
 @common.SkipIfNoModelConverter
-@pytest.mark.xfail(reason="MLETORCH-1387: Output differs")
 def test_silu_vgf_FP_inplace(test_data: input_t):
     silu_data = (test_data(), True)
     pipeline = VgfPipeline[input_t](
@@ -147,7 +144,6 @@ def test_silu_vgf_FP_inplace(test_data: input_t):
 
 @common.parametrize("test_data", Silu.test_data)
 @common.SkipIfNoModelConverter
-@pytest.mark.xfail(reason="MLETORCH-1387: Output differs")
 def test_silu_vgf_INT(test_data: input_t):
     silu_data = (test_data(), False)
     pipeline = VgfPipeline[input_t](
@@ -161,7 +157,6 @@ def test_silu_vgf_INT(test_data: input_t):
 
 @common.parametrize("test_data", Silu.test_data)
 @common.SkipIfNoModelConverter
-@pytest.mark.xfail(reason="MLETORCH-1387: Output differs")
 def test_silu_vgf_INT_inplace(test_data: input_t):
     silu_data = (test_data(), True)
     pipeline = VgfPipeline[input_t](

backends/arm/test/tester/arm_tester.py

@@ -430,6 +430,10 @@ def run_method_and_compare_outputs(
         for run_iteration in range(num_runs):
             reference_input = inputs if inputs else next(self.generate_random_inputs())
 
+            # Avoid issues with inplace operators
+            test_input = copy.deepcopy(reference_input)
+            original_input = copy.deepcopy(reference_input)
+
             input_shapes = [
                 generated_input.shape if hasattr(generated_input, "shape") else (1,)
                 for generated_input in reference_input
@@ -444,16 +448,16 @@ def run_method_and_compare_outputs(
                 # Run exported module directly
                 test_outputs, _ = pytree.tree_flatten(
                     self._calculate_reference_output(
-                        exported_program.module(), reference_input
+                        exported_program.module(), test_input
                     )
                 )
             else:
                 # Run lowered model with target
                 test_outputs, _ = pytree.tree_flatten(
-                    test_stage.run_artifact(reference_input)
+                    test_stage.run_artifact(test_input)
                 )
 
-            logger.info(f"\n      Input: {reference_input}")
+            logger.info(f"\n      Input: {original_input}")
             logger.info(f"\n Ref output: {reference_outputs}")
             logger.info(f"\nTest output: {test_outputs}")