Export MobileSAM for Ethos-U85

Prepare the MobileSAM source

Run all commands from the ExecuTorch repository root. Confirm that your Python environment is active and that you’ve sourced examples/arm/arm-scratch/setup_path.sh in the current shell.

Prepare the pinned MobileSAM source in a separate working directory:

    

        
        
python examples/arm/mobilesam_prompt_segmentation_example_ethos_u/model_export/prepare_mobilesam.py \
  --source-dir arm_test/mobilesam_manual/mobile_sam/source

    

The script downloads the pinned MobileSAM revision and applies the patch needed for a configurable image size. It keeps the external source separate from the example source.

Build the host quantized operators

The exporter requires the host shared library that registers the quantized operator out variants. Build the operators before exporting the model:

    

        
        
cmake \
  -S . \
  -B arm_test/mobilesam_manual/quantized_ops_aot \
  -DCMAKE_BUILD_TYPE=Release \
  -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON \
  -DEXECUTORCH_BUILD_KERNELS_QUANTIZED_AOT=ON \
  -DEXECUTORCH_BUILD_XNNPACK=OFF \
  -DPYTHON_EXECUTABLE="$(command -v python)"

cmake --build arm_test/mobilesam_manual/quantized_ops_aot \
  --target quantized_ops_aot_lib --parallel

    

Set the library path for the export command. The find command handles the .so extension on Linux and .dylib on macOS:

    

        
        
export EXECUTORCH_QUANTIZED_OPS_AOT_LIBRARY="$(find \
  arm_test/mobilesam_manual/quantized_ops_aot/kernels/quantized \
  -name 'libquantized_ops_aot_lib.*' -type f -print -quit)"
test -f "$EXECUTORCH_QUANTIZED_OPS_AOT_LIBRARY"

    

Export the ExecuTorch program

Export the example image and positive point prompt for the ethos-u85-256 target:

    

        
        
python examples/arm/mobilesam_prompt_segmentation_example_ethos_u/model_export/export_mobilesam.py \
  --output-path arm_test/mobilesam_manual/export/mobilesam_point_ethos_u85_448.pte \
  --calibration-image examples/models/dinov2/dog.jpg \
  --eval-image examples/models/dinov2/dog.jpg \
  --point 219 193 \
  --mobile-sam-source arm_test/mobilesam_manual/mobile_sam/source \
  --num-calibration-samples 1 \
  --num-eval-samples 1 \
  --num-debug-samples 1 \
  --minimum-fp32-quantized-iou 0.9 \
  --artifact-dir arm_test/mobilesam_manual/export/artifacts \
  --debug-output-dir arm_test/mobilesam_manual/export/debug

    

The first export downloads the pinned MobileSAM checkpoint. The exporter calibrates post-training quantization with the example image. It checks the quantized mask against the floating-point mask, and lowers the graph to Ethos-U85.

Export succeeds when the host mask intersection over union (IoU) is at least 0.9 and the .pte is written to arm_test/mobilesam_manual/export/.

Verify the export artifacts

Check that the exporter created the files needed for the remaining steps:

    

        
        
export_dir=arm_test/mobilesam_manual/export

for artifact in \
  mobilesam_point_ethos_u85_448.pte \
  mobilesam_point_ethos_u85_448.json \
  mobilesam_point_ethos_u85_448_metrics.json \
  mobilesam_point_ethos_u85_448_delegation.txt \
  debug/dog/quantized_mask.png; do
  test -s "$export_dir/$artifact" || {
    echo "Missing export artifact: $export_dir/$artifact" >&2
    exit 1
  }
done

find "$export_dir/artifacts" -type f -print -quit | grep -q . || {
  echo "No TOSA or Vela artifacts found in $export_dir/artifacts" >&2
  exit 1
}

    

The point prompt is embedded in the .pte. The image remains a runtime input and is compiled into the bare-metal application.

What you’ve accomplished and what’s next

You’ve prepared MobileSAM and exported a quantized ExecuTorch program for Ethos-U85.

Next, you’ll build the bare-metal application and run it on the Corstone-320 FVP.