Prepare your model and VS Code workspace

Understand the NX profiling workflow

You’ll run SmallUpscalerModel on an Android device with an Arm neural accelerator (NX) and capture one run with Scenario Runner. Then, you’ll inspect cycle costs and follow an operation to its Python source.

You’ll use the Arm ML Inference Advisor (MLIA) extension for Visual Studio Code (VS Code). The extension presents the results in Model Explorer.

To profile your own Android application, follow the existing-application capture instructions . After you’ve collected a capture, continue with Analyze measured performance in VS Code using your capture and matching VGF model.

Otherwise, continue with the following steps to profile the sample model using Scenario Runner.

Prepare a compatible VGF export

In a new project directory, create an export environment:

    

        
        
python3.12 -m venv export-venv
source export-venv/bin/activate
python -m pip install \
    'executorch[vgf]==1.5.1' \
    'torch==2.14.0+cpu' \
    'torchvision==0.29.0+cpu' \
    --extra-index-url https://download.pytorch.org/whl/cpu

    

Keep this environment active while running the model export steps.

Save the complete SmallUpscalerModel post-training quantization (PTQ) example from Quantize neural upscaling models with ExecuTorch as quantize_and_export_vgf.py in your project directory. Run it in the export environment:

    

        
        
python quantize_and_export_vgf.py

    

The script downloads CIFAR-10, trains for one epoch, applies PTQ, and exports output/output_tag0.tosa.vgf.

Create a profiling workspace

Open the directory containing quantize_and_export_vgf.py in VS Code using File > Open Folder. Open Terminal > New Terminal and use Bash for the remaining commands.

Copy the exported model into the profiling workspace:

    

        
        
export NX_WORKSPACE="$PWD/nx-profile"
export MODEL_VGF="$PWD/output/output_tag0.tosa.vgf"
mkdir -p "$NX_WORKSPACE/workload" "$NX_WORKSPACE/captures"
cp "$MODEL_VGF" "$NX_WORKSPACE/workload/model.vgf"
cd "$NX_WORKSPACE"

    

Your project now contains the following key files and folders:

    

        
        
your-project/
├── quantize_and_export_vgf.py
├── export-venv/
├── output/
│   └── output_tag0.tosa.vgf
└── nx-profile/
    ├── captures/
    └── workload/
        └── model.vgf

    

Keep this terminal open for the preparation and capture commands. Leave the Python source in the project directory.

Prepare the managed environment for the extension

Download the MLIA for VS Code .vsix package from the MLIA for VS Code v0.1.1 release . Then, follow the VSIX installation instructions .

In VS Code:

1. Select MLIA in the Activity Bar.

2. In New Analysis, select Refresh MLIA Environment and wait for setup to finish.

   Initialize the managed MLIA environment

What you’ve accomplished and what’s next

You’ve exported the model and initialized MLIA in VS Code.

Next, you’ll prepare the Android workload.