Repolode
Submit
pytorch

pytorch/executorch

PyTorch's on-device inference stack for running exported models on phones, laptops, browsers, and microcontrollers.

View on GitHubWebsite
Stars
5,107
Forks
1,207
Open issues
1,328
Last commit
6 hours ago

What is it?

ExecuTorch takes a PyTorch model, captures it with torch.export, lowers it to a backend-specific .pte program for the target hardware, and runs that program through C++, Python, Swift/Objective-C, Kotlin/Java, or JavaScript APIs. It covers targets from phones and laptops down to embedded systems and microcontrollers.

Hardware acceleration is handled by partitioning: supported graph regions are delegated to CPU, GPU, NPU, or DSP backends such as XNNPACK, Core ML, Vulkan, Qualcomm, and Arm Ethos-U, while portable CPU kernels handle the rest. The runtime can be trimmed with selective build so that only the operators and delegates a deployment needs are linked. ETDump, ETRecord, and numeric debugging tools support profiling and accuracy checks.

The project also includes LLM, speech, and vision examples and an experimental local OpenAI-compatible server. Models run through the same export flow, but each target needs its own backend configuration and validation of accuracy and performance.

Key features

  • 01Exports PyTorch models via torch.export into .pte programs for on-device execution.
  • 02Delegates supported graph regions to XNNPACK, Core ML, Vulkan, Qualcomm, Arm and other backends.
  • 03Provides C++, Python, Swift, Kotlin/Java and JavaScript runtime APIs.
  • 04Supports selective build to include only the operators a deployment needs.
  • 05Offers ETDump, ETRecord and numeric debugging for profiling and accuracy checks.
  • 06Includes examples for LLMs, speech recognition, speech synthesis and computer vision.

Use cases

  • Running a PyTorch vision or language model inside an Android or iOS app without a server.
  • Deploying a neural network to a microcontroller or DSP with Arm Cortex-M or Ethos-U hardware.
  • Shipping a local LLM or speech model in a desktop or laptop application.
  • Comparing accuracy and latency of one model across several hardware backends.

Quick start

  1. 1

    Install the latest stable Python package (Python 3.11–3.15)

    pip install executorch
  2. 2

    Save the Add example from the README as a script and run it to export, lower to XNNPACK, and execute a model

Things to know

  • Targeting Core ML, Qualcomm or other accelerators needs that backend's own dependencies and a separate export.
  • Some features are experimental or labeled Main/nightly and may change before release.
  • Nightly builds need a separate PyTorch nightly install, and CUDA wheels are Linux only.
  • Prebuilt Python wheels cover only Linux x86-64, Linux AArch64, macOS arm64 and Windows x86-64; other hosts require a source build.

Topics

#on-device-ai#pytorch#edge-ai#inference#mobile#embedded#tinyml#model-deployment#deep-learning#gpu

FAQ

What is executorch?
PyTorch's on-device inference stack for running exported models on phones, laptops, browsers, and microcontrollers. ExecuTorch takes a PyTorch model, captures it with torch.export, lowers it to a backend-specific .pte program for the target hardware, and runs that program through C++, Python, Swift/Objective-C, Kotlin/Java, or JavaScript APIs. It covers targets from phones and laptops down to embedded systems and microcontrollers.
Is executorch free to use? What license does it have?
executorch is open source under the Other license. Check the license terms before using it in commercial products.
Is executorch still maintained?
Actively maintained. The last commit to executorch was on Oct 20, 2026.
What can I use executorch for?
Running a PyTorch vision or language model inside an Android or iOS app without a server. Deploying a neural network to a microcontroller or DSP with Arm Cortex-M or Ethos-U hardware. Shipping a local LLM or speech model in a desktop or laptop application. Comparing accuracy and latency of one model across several hardware backends.
How do I install executorch?
Install the latest stable Python package (Python 3.11–3.15): pip install executorch Save the Add example from the README as a script and run it to export, lower to XNNPACK, and execute a model
Where can I get executorch?
The source code and installation instructions are on GitHub: https://github.com/pytorch/executorch

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects