Convert SmolVLA to ExecuTorch for inference on Arm CPUs
Introduction
Understand the SmolVLA-ExecuTorch workflow
Set up the SmolVLA ExecuTorch environment
Export, lower, and run the FP32 SmolVLA
Quantize SmolVLA to INT8 and compare it to FP32
Next Steps

  Log an issue

  Fork and edit

  Discuss on Discord

Install the required tools

Install git to clone the ExecuTorch repository, curl to fetch project files, and a C++ build toolchain:

    

        
        
sudo apt update
sudo apt install -y \
    git \
    curl \
    build-essential

The project uses Python 3.12. Install uv and use it to install Python 3.12 independently of the Python version provided by your Ubuntu release:

    

        
        
curl -LsSf https://astral.sh/uv/install.sh | sh
source "$HOME/.local/bin/env"
uv python install 3.12

Confirm that Python 3.12 is available:

    

        
        
python3.12 --version

The output is similar to:

    

        
        Python 3.12.14

Download the project files

Create a working directory and download the project files for this Learning Path:

    

        
        
mkdir -p smolvla-executorch-conversion-work
cd smolvla-executorch-conversion-work
curl -fL --retry 3 \
    -o smolvla-executorch-conversion.tar.gz \
    'https://gitlab.arm.com/learning-code-examples/code-examples/-/archive/main/code-examples-main.tar.gz?path=learning-paths/laptops-and-desktops/smolvla-executorch-conversion'
mkdir -p smolvla-executorch-conversion
tar xfz smolvla-executorch-conversion.tar.gz \
    --strip-components=4 \
    -C smolvla-executorch-conversion
cd smolvla-executorch-conversion

Set up the software environment

The setup.sh script does the following:

  • Creates a project-local virtual environment in .venv
  • Pins ExecuTorch v1.4.1 at commit e4d02f41f7909e8ed5bf4a14ffc520d733453d9f
  • Builds the ExecuTorch and XNNPACK runtime libraries with KleidiAI support and the Python bindings
  • Installs the required Python packages
  • Downloads the pinned SmolVLA checkpoint from Hugging Face

Run the setup script:

    

        
        
./scripts/setup.sh

The script downloads several gigabytes and compiles native libraries. The download can take 30 minutes or longer on systems with a small number of CPU cores.

Activate the virtual environment and resolve repository-local path variables:

    

        
        
source env.sh

Verify the resources

Verify the pinned packages, ExecuTorch revision and Python binding, XNNPACK runtime build with KleidiAI support, and SmolVLA checkpoint files:

    

        
        
python scripts/check_environment.py

The output of a successful check is similar to:

    

        
        Environment OK: aarch64, ExecuTorch e4d02f41
  Python: /path/to/smolvla-executorch-conversion/.venv/bin/python
  ExecuTorch: /path/to/smolvla-executorch-conversion/toolchain/executorch
  Runtime: /path/to/smolvla-executorch-conversion/toolchain/executorch/cmake-out-xnnpack
  Checkpoint: /path/to/smolvla-executorch-conversion/checkpoints/smolvla_base

What you’ve accomplished and what’s next

You’ve downloaded the scripts to convert the model, built the ExecuTorch runtime, and configured your environment.

Next, you’ll export and lower the FP32 SmolVLA for an Arm CPU and validate the converted model against the PyTorch reference.

Back
Next