Cortex-M backend: Minimize scope of cmsis_nn dependency.

backends/cortex_m/TARGETS

@@ -20,12 +20,23 @@ python_library(
     ],
 )
 
+python_library(
+    name = "cmsis_nn",
+    srcs = [
+        "library/__init__.py",
+        "library/cmsis_nn.py",
+    ],
+    deps = [
+        "fbsource//third-party/cmsis-nn:cmsis_nn_py",
+    ],
+)
+
 python_library(
     name = "target_config",
     srcs = [
         "target_config.py",
     ],
     deps = [
-        "fbsource//third-party/cmsis-nn:cmsis_nn_py",
+        ":cmsis_nn",
     ],
 )

backends/cortex_m/library/__init__.py

@@ -0,0 +1,4 @@
+# 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.

backends/cortex_m/library/cmsis_nn.py

@@ -0,0 +1,279 @@
+# 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.
+
+
+from __future__ import annotations
+
+from types import ModuleType
+from typing import Any, cast, ClassVar, Sequence, TYPE_CHECKING
+
+_cmsis_nn: ModuleType | None = None
+_cmsis_nn_import_error: ModuleNotFoundError | None = None
+
+
+class _EnumValue:
+    def __init__(self, enum_name: str, name: str, value: int) -> None:
+        self._enum_name = enum_name
+        self.name = name
+        self.value = value
+
+    def __repr__(self) -> str:
+        return f"<{self._enum_name}.{self.name}: {self.value}>"
+
+    def __str__(self) -> str:
+        return f"{self._enum_name}.{self.name}"
+
+
+class Backend:
+    MVE: ClassVar[Backend]
+    DSP: ClassVar[Backend]
+    SCALAR: ClassVar[Backend]
+
+    name: str
+    value: int
+
+
+Backend.MVE = cast(Backend, _EnumValue("Backend", "MVE", 0))
+Backend.DSP = cast(Backend, _EnumValue("Backend", "DSP", 1))
+Backend.SCALAR = cast(Backend, _EnumValue("Backend", "SCALAR", 2))
+
+
+class CortexM:
+    M0: ClassVar[CortexM]
+    M0PLUS: ClassVar[CortexM]
+    M3: ClassVar[CortexM]
+    M4: ClassVar[CortexM]
+    M7: ClassVar[CortexM]
+    M23: ClassVar[CortexM]
+    M33: ClassVar[CortexM]
+    M35P: ClassVar[CortexM]
+    M55: ClassVar[CortexM]
+    M85: ClassVar[CortexM]
+
+    name: str
+    value: int
+
+
+CortexM.M0 = cast(CortexM, _EnumValue("CortexM", "M0", 0))
+CortexM.M0PLUS = cast(CortexM, _EnumValue("CortexM", "M0PLUS", 1))
+CortexM.M3 = cast(CortexM, _EnumValue("CortexM", "M3", 2))
+CortexM.M4 = cast(CortexM, _EnumValue("CortexM", "M4", 3))
+CortexM.M7 = cast(CortexM, _EnumValue("CortexM", "M7", 4))
+CortexM.M23 = cast(CortexM, _EnumValue("CortexM", "M23", 5))
+CortexM.M33 = cast(CortexM, _EnumValue("CortexM", "M33", 6))
+CortexM.M35P = cast(CortexM, _EnumValue("CortexM", "M35P", 7))
+CortexM.M55 = cast(CortexM, _EnumValue("CortexM", "M55", 8))
+CortexM.M85 = cast(CortexM, _EnumValue("CortexM", "M85", 9))
+
+
+class DataType:
+    A8W4: ClassVar[DataType]
+    A8W8: ClassVar[DataType]
+    A16W8: ClassVar[DataType]
+
+    name: str
+    value: int
+
+
+DataType.A8W4 = cast(DataType, _EnumValue("DataType", "A8W4", 0))
+DataType.A8W8 = cast(DataType, _EnumValue("DataType", "A8W8", 1))
+DataType.A16W8 = cast(DataType, _EnumValue("DataType", "A16W8", 2))
+
+
+if not TYPE_CHECKING:
+    try:
+        import cmsis_nn as _real_cmsis_nn  # type: ignore[import-not-found, import-untyped]
+    except ModuleNotFoundError as exc:
+        if exc.name != "cmsis_nn":
+            raise
+        _cmsis_nn_import_error = exc
+    else:
+        _cmsis_nn = _real_cmsis_nn
+        Backend = _real_cmsis_nn.Backend
+        CortexM = _real_cmsis_nn.CortexM
+        DataType = _real_cmsis_nn.DataType
+
+
+def _missing_dependencies_error() -> ModuleNotFoundError:
+    return ModuleNotFoundError(
+        "Cortex-M backend dependencies are not installed. "
+        "Install by running `examples/arm/setup.sh --i-agree-to-the-contained-eula`, "
+        "or pip install from the CMSIS-NN repo."
+    )
+
+
+def _require_cmsis_nn() -> ModuleType:
+    if _cmsis_nn is None:
+        raise _missing_dependencies_error() from _cmsis_nn_import_error
+    return _cmsis_nn
+
+
+def resolve_backend(cpu: CortexM) -> Backend:
+    return _require_cmsis_nn().resolve_backend(cpu)
+
+
+def convolve_wrapper_buffer_size(
+    backend: Backend,
+    data_type: DataType,
+    *,
+    input_nhwc: Sequence[int],
+    filter_nhwc: Sequence[int],
+    output_nhwc: Sequence[int],
+    padding_hw: Sequence[int],
+    stride_hw: Sequence[int],
+    dilation_hw: Sequence[int],
+    input_offset: int = 0,
+    output_offset: int = 0,
+    activation_min: int = -128,
+    activation_max: int = 127,
+) -> int:
+    return _require_cmsis_nn().convolve_wrapper_buffer_size(
+        backend,
+        data_type,
+        input_nhwc=input_nhwc,
+        filter_nhwc=filter_nhwc,
+        output_nhwc=output_nhwc,
+        padding_hw=padding_hw,
+        stride_hw=stride_hw,
+        dilation_hw=dilation_hw,
+        input_offset=input_offset,
+        output_offset=output_offset,
+        activation_min=activation_min,
+        activation_max=activation_max,
+    )
+
+
+def depthwise_conv_wrapper_buffer_size(
+    backend: Backend,
+    data_type: DataType,
+    *,
+    input_nhwc: Sequence[int],
+    filter_nhwc: Sequence[int],
+    output_nhwc: Sequence[int],
+    padding_hw: Sequence[int],
+    stride_hw: Sequence[int],
+    dilation_hw: Sequence[int],
+    ch_mult: int,
+    input_offset: int = 0,
+    output_offset: int = 0,
+    activation_min: int = -128,
+    activation_max: int = 127,
+) -> int:
+    return _require_cmsis_nn().depthwise_conv_wrapper_buffer_size(
+        backend,
+        data_type,
+        input_nhwc=input_nhwc,
+        filter_nhwc=filter_nhwc,
+        output_nhwc=output_nhwc,
+        padding_hw=padding_hw,
+        stride_hw=stride_hw,
+        dilation_hw=dilation_hw,
+        ch_mult=ch_mult,
+        input_offset=input_offset,
+        output_offset=output_offset,
+        activation_min=activation_min,
+        activation_max=activation_max,
+    )
+
+
+def fully_connected_buffer_size(
+    backend: Backend,
+    data_type: DataType,
+    *,
+    filter_nhwc: Sequence[int],
+) -> int:
+    return _require_cmsis_nn().fully_connected_buffer_size(
+        backend,
+        data_type,
+        filter_nhwc=filter_nhwc,
+    )
+
+
+def transpose_conv_buffer_size(
+    backend: Backend,
+    data_type: DataType,
+    *,
+    input_nhwc: Sequence[int],
+    filter_nhwc: Sequence[int],
+    output_nhwc: Sequence[int],
+    padding_hw: Sequence[int],
+    stride_hw: Sequence[int],
+    dilation_hw: Sequence[int],
+    padding_offsets_hw: Sequence[int] = (0, 0),
+    input_offset: int = 0,
+    output_offset: int = 0,
+    activation_min: int = -128,
+    activation_max: int = 127,
+) -> int:
+    return _require_cmsis_nn().transpose_conv_buffer_size(
+        backend,
+        data_type,
+        input_nhwc=input_nhwc,
+        filter_nhwc=filter_nhwc,
+        output_nhwc=output_nhwc,
+        padding_hw=padding_hw,
+        stride_hw=stride_hw,
+        dilation_hw=dilation_hw,
+        padding_offsets_hw=padding_offsets_hw,
+        input_offset=input_offset,
+        output_offset=output_offset,
+        activation_min=activation_min,
+        activation_max=activation_max,
+    )
+
+
+def transpose_conv_reverse_conv_buffer_size(
+    backend: Backend,
+    data_type: DataType,
+    *,
+    input_nhwc: Sequence[int],
+    filter_nhwc: Sequence[int],
+    padding_hw: Sequence[int],
+    stride_hw: Sequence[int],
+    dilation_hw: Sequence[int] = (1, 1),
+    padding_offsets_hw: Sequence[int] = (0, 0),
+    input_offset: int = 0,
+    output_offset: int = 0,
+    activation_min: int = -128,
+    activation_max: int = 127,
+) -> int:
+    return _require_cmsis_nn().transpose_conv_reverse_conv_buffer_size(
+        backend,
+        data_type,
+        input_nhwc=input_nhwc,
+        filter_nhwc=filter_nhwc,
+        padding_hw=padding_hw,
+        stride_hw=stride_hw,
+        dilation_hw=dilation_hw,
+        padding_offsets_hw=padding_offsets_hw,
+        input_offset=input_offset,
+        output_offset=output_offset,
+        activation_min=activation_min,
+        activation_max=activation_max,
+    )
+
+
+def avgpool_buffer_size(
+    backend: Backend,
+    data_type: DataType,
+    *,
+    dim_dst_width: int,
+    ch_src: int,
+) -> int:
+    return _require_cmsis_nn().avgpool_buffer_size(
+        backend,
+        data_type,
+        dim_dst_width=dim_dst_width,
+        ch_src=ch_src,
+    )
+
+
+def __getattr__(name: str) -> Any:
+    return getattr(_require_cmsis_nn(), name)
+
+
+def __dir__() -> list[str]:
+    cmsis_names = set() if _cmsis_nn is None else set(dir(_cmsis_nn))
+    return sorted(set(globals()) | cmsis_names)

backends/cortex_m/passes/BUCK

@@ -1,6 +1,7 @@
 load("@fbcode_macros//build_defs:build_file_migration.bzl", "fbcode_target", "non_fbcode_target")
 # Copyright (c) Meta Platforms, Inc. and affiliates.
 # All rights reserved.
+# 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.
@@ -40,6 +41,7 @@ fbcode_target(_kind = runtime.python_library,
     deps=[
         "//caffe2:torch",
         "//executorch/backends/arm/_passes:passes",
+        "//executorch/backends/cortex_m:cmsis_nn",
         "//executorch/backends/cortex_m:target_config",
         "//executorch/backends/cortex_m/ops:ops",
         "//executorch/backends/cortex_m/passes:passes_utils",

backends/cortex_m/passes/aten_to_cortex_m_pass.py

@@ -8,12 +8,12 @@
 import math
 from typing import cast, Optional
 
-import cmsis_nn  # type: ignore[import-not-found, import-untyped]
 import executorch.backends.cortex_m.ops.operators  # noqa
 import executorch.exir as exir
 import torch
 import torch.fx
 from executorch.backends.arm._passes.arm_pass_utils import get_first_fake_tensor
+from executorch.backends.cortex_m.library import cmsis_nn
 
 from executorch.backends.cortex_m.passes.passes_utils import (
     build_activation_lut,

backends/cortex_m/passes/scratch_buffer_sizes.py

@@ -6,11 +6,11 @@
 from collections.abc import Callable
 from typing import Any, cast
 
-import cmsis_nn  # type: ignore[import-not-found, import-untyped]
 import executorch.backends.cortex_m.ops.operators  # noqa
 
 import torch
 import torch.fx
+from executorch.backends.cortex_m.library import cmsis_nn
 
 from executorch.exir.dialects._ops import ops as exir_ops
 

backends/cortex_m/target_config.py

@@ -1,5 +1,6 @@
 # Copyright (c) Meta Platforms, Inc. and affiliates.
 # All rights reserved.
+# 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.
@@ -10,7 +11,7 @@
 from enum import auto, Enum
 from typing import Optional
 
-import cmsis_nn  # type: ignore[import-not-found, import-untyped]
+from executorch.backends.cortex_m.library import cmsis_nn
 
 
 class CortexM(Enum):

backends/cortex_m/test/misc/test_cmsis_pybind.py

@@ -1,5 +1,4 @@
 # Copyright 2026 Arm Limited and/or its affiliates.
-# All rights reserved.
 #
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
@@ -11,7 +10,7 @@
 
 def _import_cmsis_nn():
     try:
-        return importlib.import_module("cmsis_nn")
+        return importlib.import_module("executorch.backends.cortex_m.library.cmsis_nn")
     except Exception as exc:
         pytest.fail(f"Failed to resolve cmsis_nn: {exc}")
 

backends/cortex_m/test/misc/test_target_config.py

@@ -1,12 +1,13 @@
 # Copyright (c) Meta Platforms, Inc. and affiliates.
 # All rights reserved.
+# 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.
 
-import cmsis_nn  # type: ignore[import-not-found, import-untyped]
 import pytest
 
+from executorch.backends.cortex_m.library import cmsis_nn
 from executorch.backends.cortex_m.target_config import CortexM, CortexMTargetConfig