Run PyTorch models on Nordic's AXON NPU — showcase models, Jupyter notebooks, and deployment tools for ExecuTorch on the nRF54LM20B.

• Docker
• Nordic sdk-edge-ai library (proprietary — not redistributed here; mount it from your own install path)
• nRF54LM20DK development kit (for on-device testing)

# x86_64 Linux
docker build -t axon-ai:latest docker/

# Apple Silicon Mac
docker build --platform linux/arm64 -t axon-ai:latest docker/

First build is slow — it downloads the NCS toolchain, Zephyr modules, and builds ExecuTorch's C++ extensions. Layers cache for subsequent rebuilds.

docker run --rm -it \
  -v ~/sdk-edge-ai:/opt/sdk-edge-ai:ro \
  -v $(pwd):/workspace/axon-ai \
  -p 8888:8888 \
  axon-ai:latest

Open the http://127.0.0.1:8888/lab?token=... URL printed to the terminal. Try the notebooks in order — 01_getting_started.ipynb through 04_deploy_and_verify.ipynb.

To flash firmware from inside the container, pass your DK's serial devices and the USB bus:

docker run --rm -it \
  -v ~/sdk-edge-ai:/opt/sdk-edge-ai:ro \
  -v $(pwd):/workspace/axon-ai \
  --device=/dev/ttyACM0 --device=/dev/ttyACM1 \
  -v /dev/bus/usb:/dev/bus/usb \
  --device-cgroup-rule='c 189:* rmw' \
  -p 8888:8888 \
  axon-ai:latest

Then run notebooks/04_deploy_and_verify.ipynb for the full flash + inference round-trip.

Docker Desktop on macOS has no USB passthrough. Export + build inside the container, then flash from the host. The build output is in the mounted volume, so the .hex file is accessible from both.

All models are hardware-verified on the nRF54LM20DK. Trained on synthetic data for demonstration — see each model's README for how to use your own dataset.

PyTorch Model
    |  train + quantize to INT8
    v
AxonPartitioner --- delegates supported ops to AXON NPU
    |
    v
TOSA Lowering (shared with ARM Ethos-U backend)
    |
    v
AXON Compiler --- produces command buffers
    |
    v
.pte + headers --- embed in Zephyr firmware
    |
    v
nRF54LM20DK --- AXON NPU executes inference

Read more:

• guides/ARCHITECTURE.md — backend design and delegation flow
• guides/SUPPORTED_OPS.md — full list of AXON-accelerated operations

axon-ai/
├── models/                   # Three showcase models, each with export + firmware
├── notebooks/                # Jupyter tutorials (start here)
├── docker/                   # Dockerfile
├── guides/                   # Architecture + supported ops
└── scripts/                  # Flash + monitor helper

This repository contains only iote.ai's own code. It composes with several third-party projects, none of which are redistributed here — they are fetched, installed, or mounted at build/run time and keep their own licenses.

Apache 2.0 — see LICENSE.

Copyright (c) 2026 iote.ai. Built by iote.ai — info@iote.ai