Learn how to deploy ExecuTorch on Arm CPUs and NPUs for effective, lightweight AI inference at the edge.

Start course
0% complete

Code

GitHub repository

What you'll learn

Practical Edge Linux, mobile, and embedded NPU deployment workflows that turn PyTorch models into lightweight ExecuTorch applications.

  • Deploy transformer and convolutional neural network models using ExecuTorch.
  • Compare PyTorch and ExecuTorch performance at the edge.
  • Explain how backend delegate choices affect inference latency.
  • Apply quantization techniques to reduce model size and deployment computation.
  • Analyze models using Model Explorer and Arm adapters.

Who this course is for

  • Developers who want to deploy ExecuTorch on Arm CPUs and NPUs for effective, lightweight inference at the edge.

Prerequisites

  • A macOS (Apple silicon) or Ubuntu development machine; Ubuntu under WSL is supported.
  • A Raspberry Pi 5 running 64-bit Raspberry Pi OS Trixie for Modules 1 and 2.
  • Android smartphone or emulation via Android Studio for Module 3.
  • Ubuntu LTS development machine or cloud instance for Module 4.
  • Basic Python experience and a working understanding of AI and machine learning concepts.

Modules

Module 0 Introduction and setupIntroduce ExecuTorch, edge AI, the course workflow, hardware requirements, and environment setup across CPU and NPU labs. Module 1 Compare Transformer Inference on Arm CPUsExport, lower, inspect, and benchmark decoder-only and encoder-based transformers with ExecuTorch and XNNPACK. Module 2 CNN inference on Arm Cortex-A and optimize with INT8 quantizationDeploy MobileNetV2 on Arm CPUs and investigate XNNPACK delegation, INT8 quantization, latency reduction, and classification behavior. Module 3 Deploy INT8 semantic segmentation in an Android appPrepare, validate, package, and run a quantized semantic-segmentation artifact with ExecuTorch and XNNPACK on Android. Module 4 Deploy MobileNetV2 with ExecuTorch and Arm Ethos-U accelerationQuantize a vision model for Ethos-U, compare delegation, validate it on FVP, and connect raw output to a classification result. Next steps Further learning and next stepsContinue with LLMs, vision, Android, iOS, neural graphics, and physical Ethos-U systems.

License

Unless otherwise noted, learning content in this course is available under CC BY-SA 4.0 and Arm-authored code under BSD-3-Clause-Clear . Third-party code and linked repositories remain under their stated licenses.