PyTorch: ExecuTorch comes to Arm Cortex-A, Cortex-M and Ethos-U85 NPU for edge AI inference

PyTorch Foundation officially extended the ExecuTorch runtime to Arm Cortex-A, Cortex-M, and Ethos-U NPU platforms on May 12, 2026. The initiative brings efficient PyTorch inference to resource-constrained edge devices — wearables, smart cameras, IoT systems — while preserving the PyTorch developer experience.

Which Arm hardware does ExecuTorch now support?

The CPU side covers Cortex-A platforms such as Raspberry Pi 5 and Cortex-M microcontrollers. ExecuTorch uses KleidiAI microkernels and Neon architectural features via the XNNPACK backend, delivering significant latency reduction compared to PyTorch eager mode. PyTorch documentation warns: without XNNPACK delegation the runtime may have higher latency than eager mode even though it retains a smaller footprint.

The NPU side targets Ethos-U accelerators, which are typically combined with a Cortex-A or Cortex-M CPU. An example hardware configuration is Ethos-U85 with 256 MAC (multiply-accumulate) units, which ExecuTorch treats as a primary deployment target.

Which models does the tutorial cover?

The documentation demonstrates three scenarios. The OPT-125M transformer model runs on Raspberry Pi 5 via the XNNPACK CPU backend. MobileNetV2 INT8 is deployed to an Ethos-U NPU via the EthosUQuantizer and EthosUCompileSpec API. The MobileNetV2 + LRN case shows how the runtime handles unsupported operators.

What do developers get in the Arm Education repository?

The arm-education/executorch_on_arm_labs GitHub repository contains hands-on labs for CPU inference on Raspberry Pi 5, hardware-accelerated inference with Ethos-U NPU, and Model Explorer Adapters (developed by Arm) for visualizing .pte files, backend partitioning, TOSA representation, and VGF files for the Arm ML SDK Vulkan. Lab content was prepared by Marcelo Rovai (UNIFEI University, Edge AI Foundation) and academic reviewers from IIIT Bangalore.

Position in the edge AI ecosystem

The deployment pipeline converts a PyTorch model through quantization, TOSA (Tensor Operator Set Architecture) representation, and Vela optimization into a .pte artifact. A static computation graph ensures predictable execution on constrained systems. ExecuTorch thus positions the PyTorch ecosystem as a serious player for edge inference, where TensorFlow Lite Micro and ONNX Runtime have previously dominated.