ComputeField Machine

The open-source compute client for a Compute Field platform. It is installed on compute computers, not on the platform server. It connects outbound, processes one leased workload at a time, and clears task data before reuse. Docker Desktop is not required for the native packages.

Install

Ubuntu 24.04 or newer, x86-64, with a working NVIDIA driver:

curl -fsSL https://github.com/ComputeField/machine/releases/latest/download/bootstrap-ubuntu.sh \
  | sudo -E bash

CPU-only Ubuntu 24.04 or newer, x86-64:

curl -fsSL https://github.com/ComputeField/machine/releases/latest/download/bootstrap-ubuntu-cpu.sh \
  | sudo -E bash

macOS with Apple silicon:

curl -fsSL https://github.com/ComputeField/machine/releases/latest/download/bootstrap-macos.sh | bash

From a source checkout:

sudo ./packaging/install-ubuntu.sh       # Ubuntu
sudo ./packaging/install-ubuntu-cpu.sh   # Ubuntu CPU service
./packaging/install-macos.sh             # macOS MPS/CPU

Tagged releases also contain separate GPU and CPU Debian packages and their SHA-256 checksums. Each bootstrap installer verifies its package before installation. The Ubuntu package installs a hardened systemd service; the macOS package installs a launchd agent. Dependencies stay inside the application virtual environment. The first install downloads the pinned PyTorch runtime and can take several minutes; an interrupted upgrade does not replace the previous working environment.

Ubuntu's restricted-user-namespace policy is supported without disabling it: the package provisions a path-specific AppArmor permission for its private, root-owned Bubblewrap executable. Installation also runs the sandbox through a one-shot unit with the production systemd security boundary before reporting success. No global security sysctl is changed.

Pair

Open Machines → Connect machine at https://computefield.net/machines and copy the command shown there. The CLI defaults to the same public HTTPS origin:

computefield-machine pair ABCD-EF12-3456

Use computefield-machine-cpu pair ABCD-EF12-3456 for the CPU package. GPU and CPU packages can coexist on one server. They have separate service users, state directories, credentials, and systemd units, and each needs its own code.

Compare the fingerprint in the terminal and browser, confirm it in the browser, then accept or decline the cross-account workload prompt. --share or --private records the same choice non-interactively. On Ubuntu the command requests sudo, writes the service identity, and restarts the service automatically. macOS also starts its launchd agent immediately after a successful pairing:

computefield-machine status

One account may pair several Machines. Private use is the default. Cross-account work is available in every native packaged installation after explicit owner consent:

computefield-machine sharing enable
computefield-machine sharing disable

Each workload runs in a fresh credential-free OS sandbox. Linux uses namespaces through Bubblewrap 0.9 or newer; macOS uses the built-in Seatbelt facility. Neither route needs BIOS virtualization, Docker Desktop, the Mac App Store, or a paid runtime license. Installers provision and verify the sandbox before reporting success.

Operate

journalctl -u computefield-machine -f          # Ubuntu logs
sudo systemctl stop computefield-machine       # Ubuntu stop
sudo systemctl stop computefield-machine-cpu   # Ubuntu CPU stop
computefield-machine stop                      # macOS stop
computefield-machine unpair --yes              # remove local identity

Unlinking a Machine in the browser immediately revokes its server credential. Installers honor standard proxy and TLS environment variables.

Uninstall or reinstall on Ubuntu

Remove the CPU package and its local pairing, cached workloads, service user, AppArmor profile, and service configuration:

sudo apt-get purge -y computefield-machine-cpu

For the NVIDIA package, use computefield-machine instead. Unlink the old offline card on the Machines page, then run the corresponding public install command above and pair it again. A normal upgrade or reinstall preserves the pairing; purge deliberately removes it.

Develop and release

python3 -m venv .venv
.venv/bin/pip install -r requirements-dev.txt
.venv/bin/pytest -q
.venv/bin/ruff check .
python tools/generate-python-reference.py
docker build --target cpu -t computefield-machine:dev .

Release checks test and audit Python, validate packaging, and build the CPU image. A reviewed release is built explicitly with packaging/build-release.sh VERSION; GitHub automation is intentionally disabled. Upload the generated GPU/CPU Debian packages, macOS source archive, three bootstrap installers, and their SHA-256 files to the matching manual release. This directory is an independent public repository; the private platform repository is not needed to build or run it. The generated Python interface index is in docs/reference/python-api.md.