ExecuTorch_Convert_Example

The repository demonstrates using the ExecuTorch to convert Pytorch model to .pte file format and run on Himax WE2 chip.

Installation

• The package has been tested in Ubuntu 20.04 LTS environment.
• Create python 3.10 virtual environment

# install python3.10-dev first
sudo apt install python3.10-dev
git clone --recursive https://github.com/HimaxWiseEyePlus/ExecuTorch_Convert_Example
cd ExecuTorch_Convert_Example
# create python 3.10 venv
python3.10 -m venv --without-pip  executorch_env_py_3_10_dev
wget https://bootstrap.pypa.io/get-pip.py
source './executorch_env_py_3_10_dev/bin/activate' 
which python
python --version
which pip
python get-pip.py

• Install requirement package and download Himax vela config ini file

pip install -r requirements.txt
pip install --no-deps git+https://git.gitlab.arm.com/tosa/tosa-reference-model.git@v2025.07.1
# download himax vela config ini
wget https://raw.githubusercontent.com/HimaxWiseEyePlus/ML_FVP_EVALUATION/main/vela/himax_vela.ini

Convert example

• We use MobilenetV2 pretrained model from huyvnphan/PyTorch_CIFAR10 for the examples.
• There are four example.
  • PTQ (Post-Training Quantization) example, it will generate mbv2_cifar10_ptq_quant_himax_ini_vela_4_5_0.pte.

    python ptq_example.py
    

    
  • QAT (Quantization-Aware Training) example, it will generate mbv2_cifar10_qat_himax_ini_vela_4_5_0.pte.

    python qat_example.py
    

    
  • Unstructure pruning + PTQ example, it will generate mbv2_cifar10_global_unstruct_prune_ptq_himax_ini_vela_4_5_0.pte.

    python unstruct_prune_ptq.py
    

    
  • Structure pruning + PTQ example, it will generate mbv2_cifar10_structured_prune_ptq_himax_ini_vela_4_5_0.pte.

    python struct_prune_ptq.py
    

    

Understanding the ExecuTorch Conversion Flow

The conversion flow uses PyTorch 2.0 Export (torch.export) and ExecuTorch's Arm Backend to compile the models for the Himax WE2 NPU (Arm Ethos-U55). The core steps, as demonstrated in the python scripts, are:

1. Model Preparation & Quantization Setup

   • The PyTorch model is placed in evaluation mode.
   • We create an EthosUCompileSpec specifying the target unit (ethos-u55-64) and reference a configuration file himax_vela.ini. You can also add the Vela option at EthosUCompileSpec such like --optimise Size to optimise SRAM size about model usage.
   • The EthosUQuantizer is initialized to annotate the model with the NPU's quantization constraints.

2. Calibration (PTQ) or Training (QAT)

   • For Post-Training Quantization (PTQ), the model undergoes calibration via prepare_pt2e. A representative dataset is passed through the model so it can record activation ranges.
   • For Quantization-Aware Training (QAT), the model is prepared using prepare_qat_pt2e and then fine-tuned through normal training loops.

3. Quantization Conversion

   • The convert_pt2e function converts the internal recorded active ranges or fake quants into actual quantized integer operations.

4. Lowering to ExecuTorch Edge Dialect

   • Ensure your model is exported via torch.export.export.
   • Using EthosUPartitioner, the graph is partitioned and delegated to the Arm Ethos-U backend via to_edge_transform_and_lower.
   • Passes like QuantizeInputs and QuantizeOutputs are applied to enforce integer inputs and outputs.

5. Saving .pte Format

   • Finally, the edge dialect is compiled into an ExecuTorch program with edge.to_executorch().
   • It is saved to disk using save_pte_program.

Run converted pte & Executorch on WE2

• You can reference torch_mb_cls to try run the models on WE2.