Set up the Zephyr and ExecuTorch development environment

Set up the platform and software

The Arm Corstone SSE-320 FPGA image for MPS4 (FI101) is an FPGA implementation that runs on the MPS4 board. The image includes an Arm Cortex-M85 processor, an Arm Ethos-U85 NPU, and a range of peripheral components. It provides a practical hardware platform for developing and evaluating machine learning applications.

Download the latest Corstone-320 FPGA image and review the platform documentation:

• Arm Corstone SSE-320 with Cortex-M85 and Ethos-U85: Example FPGA (FI101)
• SSE-320 FPGA Image for MPS4 Application Note
• Arm MPS4 FPGA Prototyping Board Technical Reference Manual
• Arm Corstone SSE-320 Example Subsystem Software Programmers Guide

Set up a Zephyr workspace and board target

To set up the Zephyr workspace for the Arm Corstone-320 MPS4 platform, complete the steps in Port Zephyr RTOS and run applications on the Arm Corstone-320 MPS4 platform . Use Zephyr version V4.3.0.

ExecuTorch integration in the Zephyr tree

ExecuTorch is integrated into the Zephyr workspace as an external module located in modules/lib/executorch. The module provides the ExecuTorch runtime, the Arm backend, the Ethos-U delegate, build scripts, and sample applications. You can build the sample applications using the Zephyr build system.

To add ExecuTorch as a Zephyr module, create executorch.yaml in zephyr/submanifests with the following content:

    

        
        
manifest:
  projects:
    - name: executorch
      url: https://github.com/pytorch/executorch
      revision: main
      path: modules/lib/executorch

    

Run the following commands to fetch the ExecuTorch repository and its submodules. The commands place the ExecuTorch source tree in modules/lib/executorch:

    

        
        
west update
cd modules/lib/executorch
git submodule sync
git submodule update --init --recursive
./install_executorch.sh

    

Set up the Arm and Ethos-U toolchain

ExecuTorch includes a setup script that downloads the Arm GNU Toolchain, the Tensor Operator Set Architecture (TOSA) Serialization Library, the Ethos-U Vela graph compiler, and other utilities.  

Run the following commands to download, install, and configure these tools on your system:

    

        
        
./examples/arm/setup.sh --i-agree-to-the-contained-eula
source examples/arm/arm-scratch/setup_path.sh

    

With the development environment and toolchain configured, prepare the model for deployment.

Pre-process the PyTorch model for NPU delegation

The ExecuTorch Ahead-of-Time (AOT) pipeline takes a PyTorch model (a torch.nn.Module) and produces a .pte binary file. The ExecuTorch runtime uses this file for inference.

The following example shows a PyTorch model, add.py, that performs a single addition:

    

        
        
import torch

b = 2

class myModelAdd(torch.nn.Module):
    def __init__(self):
        super().__init__()

    def forward(self, x):
        return x + x + b


ModelUnderTest = myModelAdd()
ModelInputs = (torch.ones(5),)

    

Run the following commands from the modules/lib/executorch directory to quantize the model and export it through the AOT flow using the Ethos-U backend:

    

        
        
source ~/zephyrproject/.venv/bin/activate
python3 -m executorch.backends.arm.scripts.aot_arm_compiler \
  --model_name=examples/arm/example_modules/add.py \
  -t ethos-u85-1024 \
  --delegate \
  --quantize \
  --memory_mode=Sram_Only \
  -o add_u85_1024_sram_only.pte

    

The following are key parameters used in the commands:

The add_u85_1024_sram_only.pte file contains the model graph, quantized weights, and a Vela-compiled command stream. The Ethos-U85 executes the command stream directly.

Verify the model file was created:

    

        
        
ls -la add_u85_1024_sram_only.pte

    

What you’ve accomplished and what’s next

You now have a quantized .pte model file ready for deployment.

Next, you’ll port the hello-executorch sample application to the Corstone-320 MPS4 platform and run inference on the Ethos-U85 NPU.