Arm backend: Add Ethos-U FVP tests for MLPerf Tiny models

backends/arm/MODELS.md

@@ -1,12 +1,17 @@
 <!-- Copyright 2025-2026 Arm Limited and/or its affiliates. -->
 # The following file contains all models that have been confirmed to be functional and tested for the Arm backend:
+# Note: Deep AutoEncoder requires manual Linear+BatchNorm1d fusion as the quantizer does not yet support this pattern.
+# Note: DS CNN requires AvgPool2d workaround for Ethos-U55 due to stride > 3 limitation.
 - Conformer
+- Deep AutoEncoder
 - Deit Tiny
 - DeepLab v3 (DL3)
+- DS CNN
 - Inception v3 (IC3)
 - Llama
 - Gemma3n
 - Long Short-Term Memory (LSTM)
+- MobileNet V1 0.25
 - MobileNet v2 (MV2)
 - MobileNet v3 (MV3)
 - Some popular torch.nn.functional models (NN functional)
@@ -16,6 +21,7 @@
 - Neural Super Sampler (NSS)
 - Phi-3
 - ResNet 18
+- ResNet-8
 - Wav2Letter (W2L)
 - Stable Diffusion:
     * CLIP Text Encoder (CLIP Text with Projection)

backends/arm/test/models/test_deep_autoencoder.py

@@ -0,0 +1,123 @@
+# Copyright 2026 Arm Limited and/or its affiliates.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+"""Ethos-U FVP tests for the MLPerf Tiny anomaly detection Deep AutoEncoder."""
+
+from typing import Tuple
+
+import pytest
+import torch
+import torch.nn as nn
+from executorch.backends.arm.test import common
+from executorch.backends.arm.test.tester.test_pipeline import (
+    EthosU55PipelineINT,
+    EthosU85PipelineINT,
+    TosaPipelineFP,
+    TosaPipelineINT,
+)
+
+from executorch.examples.models.mlperf_tiny import DeepAutoEncoderModel
+from torch.nn.utils.fusion import fuse_linear_bn_eval
+
+
+def _fuse_linear_bn(mod: nn.Module) -> nn.Module:
+    """Fuse Linear + BatchNorm1d pairs in the model.
+
+    The TOSA quantizer does not annotate linear+batch_norm patterns, so we fold
+    the BatchNorm1d into the preceding Linear before export.
+    TODO: Remove once the quantizer supports linear+bn.
+
+    """
+    if not isinstance(mod, nn.Sequential):
+        for name, child in mod.named_children():
+            setattr(mod, name, _fuse_linear_bn(child))
+        return mod
+    new_layers = []
+    layers = list(mod)
+    i = 0
+    while i < len(layers):
+        if (
+            isinstance(layers[i], nn.Linear)
+            and i + 1 < len(layers)
+            and isinstance(layers[i + 1], nn.BatchNorm1d)
+        ):
+            new_layers.append(fuse_linear_bn_eval(layers[i], layers[i + 1]))  # type: ignore[type-var, arg-type]
+            i += 2
+        else:
+            new_layers.append(_fuse_linear_bn(layers[i]))
+            i += 1
+    return nn.Sequential(*new_layers)
+
+
+_wrapper = DeepAutoEncoderModel()
+model = _fuse_linear_bn(_wrapper.get_eager_model())
+model_inputs = _wrapper.get_example_inputs()
+input_t = Tuple[torch.Tensor]
+
+quant_test_data = {
+    "per_channel_quantization=true": True,
+    "per_channel_quantization=false": False,
+}
+
+
+def test_deep_autoencoder_tosa_FP():
+    pipeline = TosaPipelineFP[input_t](
+        model,
+        model_inputs,
+        aten_op=[],
+        exir_op=[],
+        use_to_edge_transform_and_lower=True,
+    )
+    pipeline.run()
+
+
+@common.parametrize("per_channel_quantization", quant_test_data)
+def test_deep_autoencoder_tosa_INT(per_channel_quantization):
+    pipeline = TosaPipelineINT[input_t](
+        model,
+        model_inputs,
+        aten_op=[],
+        exir_op=[],
+        use_to_edge_transform_and_lower=True,
+        per_channel_quantization=per_channel_quantization,
+        atol=0.25,
+        qtol=1,
+        frobenius_threshold=None,
+        cosine_threshold=None,
+    )
+    pipeline.run()
+
+
+@pytest.mark.slow
+@common.XfailIfNoCorstone300
+@common.parametrize("per_channel_quantization", quant_test_data)
+def test_deep_autoencoder_u55_INT(per_channel_quantization):
+    pipeline = EthosU55PipelineINT[input_t](
+        model,
+        model_inputs,
+        aten_ops=[],
+        exir_ops=[],
+        use_to_edge_transform_and_lower=True,
+        per_channel_quantization=per_channel_quantization,
+        atol=0.25,
+        qtol=1,
+    )
+    pipeline.run()
+
+
+@pytest.mark.slow
+@common.XfailIfNoCorstone320
+@common.parametrize("per_channel_quantization", quant_test_data)
+def test_deep_autoencoder_u85_INT(per_channel_quantization):
+    pipeline = EthosU85PipelineINT[input_t](
+        model,
+        model_inputs,
+        aten_ops=[],
+        exir_ops=[],
+        use_to_edge_transform_and_lower=True,
+        per_channel_quantization=per_channel_quantization,
+        atol=0.25,
+        qtol=1,
+    )
+    pipeline.run()

backends/arm/test/models/test_ds_cnn.py

@@ -0,0 +1,97 @@
+# Copyright 2026 Arm Limited and/or its affiliates.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+"""Ethos-U FVP tests for the MLPerf Tiny Keyword Spotting DS-CNN model."""
+
+from typing import Tuple
+
+import pytest
+import torch
+import torch.nn as nn
+from executorch.backends.arm.test import common
+from executorch.backends.arm.test.tester.test_pipeline import (
+    EthosU55PipelineINT,
+    EthosU85PipelineINT,
+    TosaPipelineFP,
+    TosaPipelineINT,
+)
+
+from executorch.examples.models.mlperf_tiny import DSCNNKWSModel
+
+_wrapper = DSCNNKWSModel()
+model = _wrapper.get_eager_model()
+# TODO: Remove once a pass decomposes large-stride AvgPool2d.
+# Replace AvgPool2d(24,5) with AdaptiveAvgPool2d(1) so the
+# DecomposeAdaptiveAvgPool2dPass can break it into stride-1
+# pools that satisfy the Ethos-U55 stride <= 3 constraint.
+model.pool = nn.AdaptiveAvgPool2d(output_size=1)  # type: ignore[assignment]
+model_inputs = _wrapper.get_example_inputs()
+input_t = Tuple[torch.Tensor]
+
+quant_test_data = {
+    "per_channel_quantization=true": True,
+    "per_channel_quantization=false": False,
+}
+
+
+def test_ds_cnn_tosa_FP():
+    pipeline = TosaPipelineFP[input_t](
+        model,
+        model_inputs,
+        aten_op=[],
+        exir_op=[],
+        use_to_edge_transform_and_lower=True,
+    )
+    pipeline.run()
+
+
+@common.parametrize("per_channel_quantization", quant_test_data)
+def test_ds_cnn_tosa_INT(per_channel_quantization):
+    pipeline = TosaPipelineINT[input_t](
+        model,
+        model_inputs,
+        aten_op=[],
+        exir_op=[],
+        use_to_edge_transform_and_lower=True,
+        per_channel_quantization=per_channel_quantization,
+        atol=0.25,
+        qtol=1,
+        frobenius_threshold=None,
+        cosine_threshold=None,
+    )
+    pipeline.run()
+
+
+@pytest.mark.slow
+@common.XfailIfNoCorstone300
+@common.parametrize("per_channel_quantization", quant_test_data)
+def test_ds_cnn_u55_INT(per_channel_quantization):
+    pipeline = EthosU55PipelineINT[input_t](
+        model,
+        model_inputs,
+        aten_ops=[],
+        exir_ops=[],
+        use_to_edge_transform_and_lower=True,
+        per_channel_quantization=per_channel_quantization,
+        atol=0.25,
+        qtol=1,
+    )
+    pipeline.run()
+
+
+@pytest.mark.slow
+@common.XfailIfNoCorstone320
+@common.parametrize("per_channel_quantization", quant_test_data)
+def test_ds_cnn_u85_INT(per_channel_quantization):
+    pipeline = EthosU85PipelineINT[input_t](
+        model,
+        model_inputs,
+        aten_ops=[],
+        exir_ops=[],
+        use_to_edge_transform_and_lower=True,
+        per_channel_quantization=per_channel_quantization,
+        atol=0.25,
+        qtol=1,
+    )
+    pipeline.run()

backends/arm/test/models/test_mobilenet_v1_025.py

@@ -0,0 +1,93 @@
+# Copyright 2026 Arm Limited and/or its affiliates.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+"""Ethos-U FVP tests for the MLPerf Tiny Visual Wake Words MobileNetV1 (width
+0.25).
+"""
+
+from typing import Tuple
+
+import pytest
+import torch
+from executorch.backends.arm.test import common
+from executorch.backends.arm.test.tester.test_pipeline import (
+    EthosU55PipelineINT,
+    EthosU85PipelineINT,
+    TosaPipelineFP,
+    TosaPipelineINT,
+)
+
+from executorch.examples.models.mlperf_tiny import MobileNetV1025Model
+
+_wrapper = MobileNetV1025Model()
+model = _wrapper.get_eager_model()
+model_inputs = _wrapper.get_example_inputs()
+input_t = Tuple[torch.Tensor]
+
+quant_test_data = {
+    "per_channel_quantization=true": True,
+    "per_channel_quantization=false": False,
+}
+
+
+def test_mobilenet_v1_025_tosa_FP():
+    pipeline = TosaPipelineFP[input_t](
+        model,
+        model_inputs,
+        aten_op=[],
+        exir_op=[],
+        use_to_edge_transform_and_lower=True,
+    )
+    pipeline.run()
+
+
+@common.parametrize("per_channel_quantization", quant_test_data)
+def test_mobilenet_v1_025_tosa_INT(per_channel_quantization):
+    pipeline = TosaPipelineINT[input_t](
+        model,
+        model_inputs,
+        aten_op=[],
+        exir_op=[],
+        use_to_edge_transform_and_lower=True,
+        per_channel_quantization=per_channel_quantization,
+        atol=0.25,
+        qtol=1,
+        frobenius_threshold=None,
+        cosine_threshold=None,
+    )
+    pipeline.run()
+
+
+@pytest.mark.slow
+@common.XfailIfNoCorstone300
+@common.parametrize("per_channel_quantization", quant_test_data)
+def test_mobilenet_v1_025_u55_INT(per_channel_quantization):
+    pipeline = EthosU55PipelineINT[input_t](
+        model,
+        model_inputs,
+        aten_ops=[],
+        exir_ops=[],
+        use_to_edge_transform_and_lower=True,
+        per_channel_quantization=per_channel_quantization,
+        atol=0.25,
+        qtol=1,
+    )
+    pipeline.run()
+
+
+@pytest.mark.slow
+@common.XfailIfNoCorstone320
+@common.parametrize("per_channel_quantization", quant_test_data)
+def test_mobilenet_v1_025_u85_INT(per_channel_quantization):
+    pipeline = EthosU85PipelineINT[input_t](
+        model,
+        model_inputs,
+        aten_ops=[],
+        exir_ops=[],
+        use_to_edge_transform_and_lower=True,
+        per_channel_quantization=per_channel_quantization,
+        atol=0.25,
+        qtol=1,
+    )
+    pipeline.run()