Install ExecuTorch and download the model on the Linux host

Clone ExecuTorch 1.4

Keep a source checkout because you’ll cross-compile the Android runtime and the standalone Llama runner:

    

        
        
cd $HOME
git clone --branch release/1.4 --recursive https://github.com/pytorch/executorch.git
cd executorch
git submodule update --init --recursive

    

Create and activate a Python virtual environment:

    

        
        
python3.12 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip setuptools wheel

    

To reduce memory pressure during builds, set the following variable:

    

        
        
export CMAKE_BUILD_PARALLEL_LEVEL=2

    

Fix the PyTorch version mismatch

Always verify torch.__version__ before you start a long native build.

If you build the host envionment with PyTorch 2.14.0+cpu on the release/1.4 branch, the build breaks with the following message:

    

        
        
ATen/core/Tensor.h:70:37: error: 'C10_LIFETIMEBOUND' does not name a type

    

To avoid this, pin the expected version:

    

        
        
python -m pip uninstall -y torch
python -m pip install "torch==2.13.0+cpu" --index-url https://download.pytorch.org/whl/cpu
python -c "import torch; print(torch.__version__)"

    

The output is similar to:

    

        
        2.13.0+cpu

        
    

Then, clean and rebuild the host package:

    

        
        
python install_executorch.py --clean
export CMAKE_BUILD_PARALLEL_LEVEL=2
./install_executorch.sh --use-pt-pinned-commit

    

Request access to Llama 3.2 1B Instruct

The model that you’ll use in this workflow comes from the gated Hugging Face repository meta-llama/Llama-3.2-1B-Instruct .

To request access:

1. Sign in to your Hugging Face account.
2. Open the Llama 3.2 1B Instruct model page.
3. Review and accept the Meta Llama license, then submit the access request.
4. Wait until Hugging Face confirms that your account can access the repository before continuing.

Access approval is associated with the Hugging Face account that submitted the request.

Install and authenticate the Hugging Face CLI

Install the Hugging Face Hub CLI in the active Python virtual environment:

    

        
        
python -m pip install --upgrade huggingface_hub
hf version

    

Sign in with the same account that you used to access the model. The command prompts you to authenticate through a browser or with a Hugging Face user access token:

    

        
        
hf auth login
hf auth whoami

    

Confirm that hf auth whoami displays the account that has access to the gated model.

Download Llama 3.2 1B Instruct

Keep the model outside the source tree:

    

        
        
hf download \
  meta-llama/Llama-3.2-1B-Instruct \
  --include "original/*" \
  --local-dir ~/Llama-3.2-1B-Instruct

    

Verify that the three files required by the export are present:

    

        
        
test -f ~/Llama-3.2-1B-Instruct/original/consolidated.00.pth && echo "Checkpoint OK"
test -f ~/Llama-3.2-1B-Instruct/original/params.json && echo "Parameters OK"
test -f ~/Llama-3.2-1B-Instruct/original/tokenizer.model && echo "Tokenizer OK"

    

The example measured run used:

• consolidated.00.pth at about 2.4 GB
• params.json
• tokenizer.model at about 2.1 MB

What you’ve accomplished and what’s next

You’ve now prepared the ExecuTorch source and Python environment. You’ve also requested and gained access to the Llama 3.2 1B Instruct model.

Next, you’ll export the Vulkan PTE.