Installation

Three roles, three lock files. Every dependency is pinned with a hash in requirements/*.lock (generated with scripts/lock_requirements.sh); installing with --require-hashes guarantees the same stack on every machine. The direct dependencies and their minimum versions are in the .in files and follow the software table of Appendix B.

Component	Minimum supported	Pinned in the lock files (thesis, Appendix B, Table B.1)
Operating system	Raspberry Pi OS 64-bit (devices), Ubuntu 22.04 LTS (edge server)	Bookworm 64-bit on the devices (Python 3.11), Ubuntu 22.04 LTS
Python	3.11	3.11
TensorFlow (DNN, Chapter 4)	≥ 2.8.0, CPU	2.15.1
PyTorch (LLM, Chapters 5 and 6)	2.1.2, CPU	2.1.2 (+cpu on x86_64)
Transformers	4.36	4.46.3 (Qwen2 needs 4.37, Llama 3.2 needs 4.43, HeadSplit 4.40)
scikit-learn	≥ 1.0.0, edge server only	1.9.1
Paramiko (SSH)	≥ 2.11, edge server only	5.1.0
Flask (web application)	≥ 2.2.0, edge server only	3.1.3
NTP	client (devices), server (edge server)	chrony, configured by provisioning
psutil, PyYAML, jsonschema, NumPy		7.2.2, 6.0.3, 4.26.0, 1.26.4

Edge server

Ubuntu 22.04 LTS, x86_64. The provisioning script does all of this (and the network services); by hand:

sudo apt-get install -y python3.11 python3.11-venv git rsync iperf3
git clone https://github.com/Dimitrios-Kafetzis/Distributed-Inference-Testbed.git
cd Distributed-Inference-Testbed
python3.11 -m venv .venv
.venv/bin/pip install --require-hashes -r requirements/edge_server.lock \
    --extra-index-url https://download.pytorch.org/whl/cpu   # the CPU build of PyTorch lives there
cp config.example.yaml config.yaml      # then edit
cp nodes.example.json nodes.json        # then edit, or manage through NNPM
.venv/bin/python -m src.config          # shows the effective configuration
.venv/bin/python -m gui.app             # NNPM on nnpm_host:nnpm_port

requirements/edge_server.lock contains TensorFlow and PyTorch (CPU builds): the edge server profiles its own capacity and runs the partitioned-inference service, so it needs the same frameworks as the devices. The x86_64 CPU build of PyTorch (torch==2.1.2+cpu) is published on the PyTorch index, not on PyPI, hence the --extra-index-url (with uv, add --index-strategy unsafe-best-match). With a package mirror in place, replace it with --no-index --find-links mirror/wheels/x86_64.

Edge devices

Raspberry Pi OS Bookworm 64-bit (uname -m must print aarch64). scripts/provision/provision_device.sh <node id> installs everything from the mirror; by hand, on the device:

sudo apt-get install -y python3 python3-venv python3-pip chrony iproute2 iperf3 rsync git
mkdir -p ~/testbed && cd ~/testbed
git clone https://github.com/Dimitrios-Kafetzis/Distributed-Inference-Testbed.git code
python3 -m venv venv
venv/bin/pip install --require-hashes -r code/requirements/edge_device.lock       # TensorFlow
venv/bin/pip install --require-hashes -r code/requirements/edge_device_llm.lock   # PyTorch and Transformers (optional)

The directory layout the edge server expects is given by ed_root in config.yaml (default /home/pi/testbed): code/, venv/, models/, data/, results/, run/. The devices need no copy of config.yaml: the edge server passes every setting as TESTBED_* environment variables and command-line arguments.

Optional, for the INA260 power monitor: venv/bin/pip install -r code/requirements/power.in (the Adafruit Blinka and INA260 packages are platform-specific and therefore not in the lock files) and I2C enabled on the board.

Privileges: tc and the heterogeneity module (cpufreq and cgroup writes) run with sudo -n; provisioning installs a sudoers rule that allows exactly those two commands for the device user. Nothing else on a device needs root.

Development machine

Any Linux machine with Python 3.11 (this is also what the continuous-integration workflow does):

uv venv .venv --python 3.11
uv pip install --python .venv/bin/python --require-hashes -r requirements/dev.lock
.venv/bin/pytest -m "not slow"            # fast suite: fakes, dry runs, local mode
.venv/bin/pytest -m slow                  # real TensorFlow and PyTorch models on the CPU
.venv/bin/ruff check . && .venv/bin/mypy && .venv/bin/shellcheck -x $(git ls-files '*.sh')

requirements/dev.lock includes the edge-server stack plus pytest, ruff, mypy and shellcheck. The whole pipeline runs without hardware in local mode:

.venv/bin/python -m src.orchestration.experiment run --scenario examples/scenarios/local_demo.yaml --local
.venv/bin/python -m gui.app --local        # NNPM with simulated nodes

Updating the pins

Edit the .in files, then scripts/lock_requirements.sh (needs uv), which resolves each lock file for its platform (x86_64-manylinux_2_28 for the edge server and development, aarch64-manylinux_2_28 for the devices) with hashes. Rebuild the mirror afterwards (scripts/provision/build_mirror.sh wheels and manifest).