Arm backend: Change W2L "input type" for Ethos-U55 test (#13839)

The W2L model includes a conv1d layer with an unsupported stride length
(160) when the input_type argument to the model's constructor is set to
the default value ("waveform"). Change the input_type parameter to
"power_spectrum" as a workaround to not include this unsupported layer.

Still the test case will not pass because there is lack of memory
(SRAM). Change the pytest.mark.skip decorator to the U55 test to reflect
the new expected reason for failure.

Co-authored-by: Martin Lindström <[email protected]>

Author: martinlsm

backends/arm/test/models/test_w2l_arm.py

@@ -36,7 +36,6 @@ class TestW2L(unittest.TestCase):
     input_frames = 400
     num_features = 1
 
-    w2l = models.Wav2Letter(num_features=num_features).eval()
     model_example_inputs = get_test_inputs(batch_size, num_features, input_frames)
 
     all_operators = [
@@ -45,11 +44,17 @@ class TestW2L(unittest.TestCase):
         "executorch_exir_dialects_edge__ops_aten_relu_default",
     ]
 
+    @staticmethod
+    def create_model(input_type: str = "waveform"):
+        return models.Wav2Letter(
+            num_features=TestW2L.num_features, input_type=input_type
+        ).eval()
+
 
 @pytest.mark.slow  # about 3min on std laptop
 def test_w2l_tosa_FP():
     pipeline = TosaPipelineFP[input_t](
-        TestW2L.w2l,
+        TestW2L.create_model(),
         TestW2L.model_example_inputs,
         aten_op=[],
         exir_op=TestW2L.all_operators,
@@ -62,7 +67,7 @@ def test_w2l_tosa_FP():
 @pytest.mark.flaky
 def test_w2l_tosa_INT():
     pipeline = TosaPipelineINT[input_t](
-        TestW2L.w2l,
+        TestW2L.create_model(),
         TestW2L.model_example_inputs,
         aten_op=[],
         exir_op=TestW2L.all_operators,
@@ -74,12 +79,14 @@ def test_w2l_tosa_INT():
 @pytest.mark.slow
 @common.XfailIfNoCorstone300
 @pytest.mark.xfail(
-    reason="MLETORCH-1009: Wav2Letter fails on U55 due to unsupported conditions",
-    strict=False,
+    reason="Wav2Letter fails on U55 due to insufficient memory",
+    strict=True,
 )
 def test_w2l_u55_INT():
     pipeline = EthosU55PipelineINT[input_t](
-        TestW2L.w2l,
+        # Use "power_spectrum" variant because the default ("waveform") has a
+        # conv1d layer with an unsupported stride size.
+        TestW2L.create_model("power_spectrum"),
         TestW2L.model_example_inputs,
         aten_ops=[],
         exir_ops=[],
@@ -94,7 +101,7 @@ def test_w2l_u55_INT():
 @pytest.mark.skip(reason="Intermittent timeout issue: MLETORCH-856")
 def test_w2l_u85_INT():
     pipeline = EthosU85PipelineINT[input_t](
-        TestW2L.w2l,
+        TestW2L.create_model(),
         TestW2L.model_example_inputs,
         aten_ops=[],
         exir_ops=[],
@@ -108,7 +115,7 @@ def test_w2l_u85_INT():
 @pytest.mark.slow
 def test_w2l_vgf_INT():
     pipeline = VgfPipeline[input_t](
-        TestW2L.w2l,
+        TestW2L.create_model(),
         TestW2L.model_example_inputs,
         aten_op=[],
         exir_op=TestW2L.all_operators,
@@ -121,7 +128,7 @@ def test_w2l_vgf_INT():
 @common.SkipIfNoModelConverter
 def test_w2l_vgf_FP():
     pipeline = VgfPipeline[input_t](
-        TestW2L.w2l,
+        TestW2L.create_model(),
         TestW2L.model_example_inputs,
         aten_op=[],
         exir_op=TestW2L.all_operators,