4

Setting Up ExecuTorch¶

This tutorial walks you through an end-to-end example of configuring your environment for ExecuTorch, installing ExecuTorch, exporting your model, and finally building and running a runtime.

What you will learn

How to set up your environment to work on ExecuTorch
How to generate a sample ExecuTorch program
How to build and run an ExecuTorch runtime

Prerequisites

Understand the concepts defined in ExecuTorch Overview

Supported Host Environments¶

We have tested these instructions on the following systems, although they should also work on other systems with similar environments.

Linux (x86_64)

CentOS 8 or later
Ubuntu 20.04.6 LTS or later
RHEL 8 or later
Windows Subsystem for Linux running any of the above

macOS (x86_64/M1/M2)

Big Sur or later

The most critical requirements are:

The ability to install a recent version of conda, described below.
g++ version 8 or higher, clang++ version 8 or higher, or another C++17-compatible toolchain that supports GNU C-style statement expressions (({ ... }) syntax).

Note that the cross-compilable core runtime code supports a wider range of toolchains, down to C++11. See the Runtime Overview for portability details.

Set up Your Environment¶

Before you can start working with ExecuTorch, you’ll need to set up your environment. This is an important step to ensure that everything runs smoothly and efficiently. We recommend using conda to create and manage your virtual environment. Conda is a package management system and environment manager for Python and other programming languages, which is built on top of the Python package manager pip, and provides a more convenient and flexible way to manage packages and environments. In this section, you will set up your conda environment and install the required dependencies.

Follow these steps:

If you do not have it already, install conda on your machine by following the steps in the conda installation guide.

Clone the executorch repository:

git clone https://github.com/pytorch/executorch.git

Update the submodules:

cd executorch
git submodule sync
git submodule update --init

Create and activate your conda environment:

conda create -yn executorch python=3.10.0
conda activate executorch

Or alternatively use a Python virtual environment:

python3 -m venv .executorch
source .executorch/bin/activate

Install Cmake
```
conda install cmake
```
Alternatively:
```
pip install cmake
```
Install ExecuTorch and dependencies:
```
./install_requirements.sh
```
Expose FlatBuffers compiler:

ExecuTorch uses flatc to export models and builds it from sources at third-party/flatbuffers. Make it available by adding it to the $PATH environment variable, as prompted by the previous step, or exporting as $FLATC_EXECUTABLE enironment variable. Run ./build/install_flatc.sh to make sure flatc is installed correctly.

You have successfully set up your environment to work with ExecuTorch. The next step is to generate a sample ExecuTorch program.

Generate a Sample ExecuTorch program¶

After you set up your environment, you are ready to convert your programs into ExecuTorch programs. You will need to use torch.export and the executorch.exir to export your program. Then, you can save your program as a .pte file, which is the file extension ExecuTorch expects. To demonstrate how to do it, we will generate an ExecuTorch program file from an nn.Module.

You can generate an ExecuTorch program by using a sample script or by using the Python interpreter.

We have created the export.py script that demonstrates a simple model export to flatbuffer. This script is available in the pytorch/executorch repository.

To generate a sample program, complete the following steps:

Run the export.py script:

python3 -m examples.portable.scripts.export --model_name="add"

This command has created a add.pte file that contains your sample program, which adds its inputs multiple times.

Alternatively, you can use a Python interpreter to perform similar steps, this time creating a mul.pte program file that multiplies its inputs:

import executorch.exir as exir
from executorch.exir.tests.models import Mul
m = Mul()
open("mul.pte", "wb").write(exir.capture(m, m.get_random_inputs()).to_edge().to_executorch().buffer)

In this step, you learned how you can export your PyTorch program to an ExecuTorch program. You can apply the same principle to your own PyTorch programs.

The next step is to run your program by setting up Buck2 and building an executor_runner.

Building a Runtime¶

After you have exported your program, you are almost ready to run it. The next step involves using Buck2 to build a runtime.

Buck2 is an open-source build system that enables developers to manage project dependencies easily and efficiently. We will use Buck2 to build the executor_runner, a sample wrapper for the ExecuTorch runtime which includes all the operators and backends.

You will need the following prerequisits for this section:

The zstd command line tool — install by running
```
pip3 install zstd
```
Version 2023-07-18 of the buck2 commandline tool — you can download a prebuilt archive for your system from the Buck2 repo. Note that the version is important, and newer or older versions may not work with the version of the buck2 prelude used by the ExecuTorch repo.

Complete the following steps:

Ensure that Git has fetched and updated the submodules. This is necessary whenever the commit hash of a submodule changes. Therefore, you need to periodically sync your submodules with upstream:
```
cd executorch
git submodule sync
git submodule update --init
```
Install ExecuTorch and dependencies:
```
./install_requirements.sh
```

Configure Buck2 by decompressing with the following command (filename depends on your system):

# For example, buck2-x86_64-unknown-linux-musl.zst or buck2-aarch64-apple-darwin.zst
zstd -cdq buck2-DOWNLOADED_FILENAME.zst > /tmp/buck2 && chmod +x /tmp/buck2

You may want to copy the buck2 binary into your $PATH so you can run it as buck2.

Build a binary:

/tmp/buck2 build //examples/portable/executor_runner:executor_runner --show-output

The --show-output flag prints the path to the executable if you want to run it directly.

Now that you have built our sample programs, you can proceed to run them.

Run Your Program¶

After you build your program, you are ready to run it. We will use the buck run command to run our program.

Run the binary:

To run the add.pte program:

/tmp/buck2 run //examples/portable/executor_runner:executor_runner -- --model_path add.pte

Alternatively, you can execute the binary directly from the --show-output path shown in the build step. For example, you can run the following command for the add.pte program:

./buck-out/.../executor_runner --model_path add.pte

Next Steps¶

Congratulations! You have successfully exported, built, and run your first ExecuTorch program. Now that you have a basic understanding of how ExecuTorch works, you can start exploring its advanced features and capabilities. Here is a list of sections you might want to read next:

Exporting a model
Using EXIR for advanced exports
Review more advanced examples in the executorch/examples directory