ExecuTorch is an end-to-end framework for on-device AI inference and training, targeting mobile, embedded, and microcontroller platforms. It enables developers to leverage PyTorch models across diverse hardware, from smartphones to resource-constrained systems, powering Meta's on-device AI features.

How It Works

ExecuTorch provides a unified toolchain for model authoring, conversion, debugging, and deployment. It achieves portability by supporting a wide range of operating systems and hardware accelerators (CPU, NPU, DSP) through integrations like Vulkan and XNNPACK. The lightweight runtime is designed for high performance and efficient utilization of device capabilities.

Quick Start & Requirements

• Install/Run: Refer to the Step by Step Tutorial or use the provided Colab Notebook.
• Prerequisites: Platform support includes iOS, Mac, Android, Linux, and microcontrollers. Hardware acceleration is available for Apple, Arm, Cadence, MediaTek, OpenVINO, Qualcomm, Vulkan, and XNNPACK.
• LLM Examples: Specific instructions are available for Llama, Qwen 3, Phi-4-mini, and Llava.

Highlighted Details

• Powers on-device AI across Meta's products (Facebook, Instagram, Meta Quest, etc.).
• Supports diverse AI models: LLMs, Computer Vision, ASR, TTS.
• Extensive platform and hardware acceleration support.
• Unified PyTorch toolchain for simplified development.

Maintenance & Community

• Active community engagement via Discussion Board and Discord.
• Contributions are welcomed; refer to Contributing Guidelines.

Licensing & Compatibility

• BSD licensed.
• Compatible with commercial and closed-source applications.

Limitations & Caveats

The framework is actively developed, and while it supports a broad range of platforms, specific hardware integrations or model optimizations may still be evolving. Users should consult the documentation for the latest compatibility and performance details.