Arm Keil MDK v6 Essential and Professional editions are now available to buy as monthly or annual subscriptions from our Developer Tools store. Visit the store to get access to the tools suite today!

  1. Packs
  2. ExecuTorch

ExecuTorch CMSIS Pack

Overview

ExecuTorch is the PyTorch Edge Runtime, enabling efficient on-device AI inference. This pack is the Cortex-M / bare-metal distribution.

This pack provides: - Core Runtime: Program loading and execution - Portable Operators: Platform-independent operator implementations - Quantized Operators: Optimized quantized inference - Ethos-U Backend: Hardware acceleration for Arm Ethos-U NPU - Cortex-M Operators: CMSIS-NN optimized INT8 kernels

Getting Started

Basic Usage

  1. Add the pack to your csolution: ```yaml packs:

    • pack: PyTorch::ExecuTorch ```
  2. Add required components. CMSIS component IDs use a single colon between Cclass / Cgroup / Csub, with a space (not "::") inside the Csub name. ```yaml components: # Core runtime (always required)

    • component: Machine Learning:ExecuTorch:Runtime
    • component: Machine Learning:ExecuTorch:Kernel Utils
    • component: Machine Learning:ExecuTorch:Kernel Registration

    # Optional: TensorPtr / make_tensor_ptr helpers - component: Machine Learning:ExecuTorch:Extension Tensor

    # Optional: Ethos-U NPU delegate. Also requires an NPU driver # component from the ARM::ethos-u-core-driver pack. - component: Machine Learning:ExecuTorch:Backend EthosU ```

CMSIS-NN acceleration is not a backend component: select the individual Cortex-M <op> operator components (see below). All of them except quantize_per_tensor and dequantize_per_tensor additionally require ARM::CMSIS:NN Lib.

  1. Include ExecuTorch headers in your code: cpp #include <executorch/runtime/executor/program.h> #include <executorch/runtime/executor/method.h>

Model Integration

Operators are included as individual CMSIS components, one per op_*.cpp. Add the ones your model uses:

components:
  - component: Machine Learning:ExecuTorch Operators:Portable add
  - component: Machine Learning:ExecuTorch Operators:Portable mul
  # ... other operators used by your model

You can also generate this list automatically: a Model Pack produced from your .pte file declares dependencies on the operators the model needs, and the toolchain resolves them against this pack's components.

Renamed Components

The CMSIS-Pack schema limits component names to 32 characters. Since pack version 1.4.1 three operator components therefore carry shortened names; projects written against the 1.4.0 pack must update them:

Up to 1.4.0 From 1.4.1
Portable max_pool2d_with_indices_backward Portable max_pool2d_indices_bwd
Cortex-M quantized_depthwise_conv2d Cortex-M quantized_dw_conv2d
Cortex-M quantized_transpose_conv2d Cortex-M quantized_tr_conv2d

Memory Requirements

Component Flash (approx) RAM (approx)
Runtime 50-100 KB 4-8 KB
Per Operator 1-10 KB minimal
Ethos-U Backend 20-40 KB 2-4 KB

Actual requirements depend on: - Selected operators - Model complexity - Tensor sizes

Compiler Support

Tested with arm-none-eabi-gcc 14.3 and Arm Compiler for Embedded (AC6) 6.23 on Cortex-M85. LLVM Embedded Toolchain for Arm is expected to work but was not exercised for this release.

AC6 projects must build with -ffp-mode=full. armclang otherwise assumes finite math, which breaks operators that depend on IEEE inf/NaN behaviour (isinf, isnan, gelu).

Dependencies

  • ARM::CMSIS (core headers)
  • ARM::CMSIS-NN (for the Cortex-M <op> CMSIS-NN accelerated operators)
  • ARM::ethos-u-core-driver (for Ethos-U backend)

CMSIS-NN 8.0.0 on Cortex-M55 and Cortex-M85

The ARM::CMSIS-NN 8.0.0 pack does not compile in its default configuration for cores with Helium floating-point support (MVE-F), such as Cortex-M55 and Cortex-M85. Four of its element-wise sources fail with implicit declaration of function 'arm_nn_clamp_mve_f32' (or ..._f16).

Until a fixed CMSIS-NN release is available, enable CMSIS-NN's float support in any project that selects the Cortex-M operators on such a core:

# in the *.csolution.yml or *.cproject.yml
define:
  - ARM_NN_ENABLE_F32: 1
  - ARM_NN_ENABLE_F16: 1

Resources