Arm backend: Stabilize GroupNorm transpose-count

Preserve the unmodified GroupNorm model and select the exact expected
transpose count from its exported output layout before backend
lowering. This supports release and nightly PyTorch while retaining
native GroupNorm regression coverage and numerical comparisons.

Authored with assistance from OpenAI Codex.

Signed-off-by: Sangwon Ha <sangwon.ha@arm.com>
Change-Id: I25aa7f4269fe9676579b43b57119850a8ebac45b

Author: FabulousSuperDude

backends/arm/test/misc/test_transpose_counts.py

@@ -13,6 +13,7 @@
     TosaPipelineFP,
     TosaPipelineINT,
 )
+from executorch.backends.test.harness.stages import StageType
 
 
 InputT = Tuple[Any, ...]
@@ -540,15 +541,6 @@ def forward(self, x: torch.Tensor):
         (torch.randn(1, 2, 8, 8).to(memory_format=torch.channels_last),),
         3,
     ),
-    # Rebaselined 1 -> 2 with the dev20260913 pin, which changed how GroupNorm
-    # decomposes under channels-last. The pipeline still compares against the
-    # TOSA reference model, so the extra transpose is numerically correct, but
-    # nobody has checked whether it is one the backend could still fold away.
-    "groupnorm_channels_last": TransposeCountCase(
-        GroupNormModule(),
-        (torch.randn(1, 4, 4, 4).to(memory_format=torch.channels_last),),
-        2,
-    ),
     "cumsum_rank4_dim3_channels_last": TransposeCountCase(
         CumsumModule(),
         (torch.randn(1, 2, 3, 4).to(memory_format=torch.channels_last), 3),
@@ -582,3 +574,24 @@ def test_transpose_counts_tosa_FP_channels_last(case: TransposeCountCase) -> Non
     pipeline = TosaPipelineFP[InputT](case.module, case.inputs, aten_op=[])
     pipeline.count_tosa_ops({"TRANSPOSE": case.expected_transposes})
     pipeline.run()
+
+
+def test_transpose_counts_tosa_FP_groupnorm_channels_last() -> None:
+    inputs = (torch.randn(1, 4, 4, 4).to(memory_format=torch.channels_last),)
+    pipeline = TosaPipelineFP[InputT](
+        GroupNormModule(), inputs, aten_op="torch.ops.aten.group_norm.default"
+    )
+    pipeline.tester.export()
+    pipeline.pop_stage("export")
+
+    # PyTorch's output layout determines the boundary transposes before lowering.
+    exported_program = pipeline.tester.get_artifact(StageType.EXPORT)
+    output_node = next(n for n in exported_program.graph.nodes if n.op == "output")
+    (output,) = output_node.args[0]
+    expected_transposes = {
+        (0, 1, 2, 3): 1,
+        (0, 2, 3, 1): 2,
+    }[output.meta["val"].dim_order()]
+
+    pipeline.count_tosa_ops({"TRANSPOSE": expected_transposes})
+    pipeline.run()