Downloads

Programs you run on your own computer. The volunteer program and the Linux installer are for verified users: sign in with an approved account to download them. No API key is needed.

Start

1. Install Python 3.9 or newer, if you do not have it.
2. Download volunteer.py above and open a terminal in the folder it is in.
3. Run it. On Windows the command is usually py instead of python3.

   python3 volunteer.py

   The first time, it offers to install what it needs (numpy and websockets) and a few optional packages that make it faster or let it use more hardware. To install everything without being asked:

   python3 volunteer.py --install

4. It prints the devices it found, connects, and starts working. Leave it running; Ctrl+C stops it.

To check a download, compare its SHA-256 with the one shown above: sha256sum volunteer.py on Linux, shasum -a 256 volunteer.py on macOS, certutil -hashfile volunteer.py SHA256 on Windows.

Choose how much it uses

The first run creates volunteer.ini next to volunteer.py. Open it in any text editor, change a number and save; the running program picks the change up within a few seconds.

[resources]
cpu = 50          ; processor time
memory = 50       ; system memory (RAM)
gpu = 50          ; graphics cards: compute time
gpu_memory = 50   ; ...and memory on the card
tpu = 50
tpu_memory = 50
npu = 50          ; NPUs and other AI accelerators
npu_memory = 50

• 0 turns a resource off completely. With everything at 0 the program stays idle and does not connect.
• One device can have its own limit, using its id from --list-devices: cuda:1 = 20, or cuda:1.memory = 30 for its memory.
• Under [devices], exclude = wgpu:0 skips a device, and only = cuda:0, cpu uses nothing else.
• It runs at low priority by default, so whatever else you are doing comes first.

While it runs you can also type into its terminal: status, set gpu 25, quit.

What it works on

• Models — the grid's main workload. The program holds part or all of a model's weights in your GPU's memory (or system memory on a CPU) and runs the layers it is sent. When one computer holds everything a text generation needs, it produces the whole reply itself and streams the tokens back, which is several times faster than a round trip per word. When a model would not otherwise fit, the grid may ask the program to keep its weights in a compact 8-bit or 4-bit form (it converts them as they arrive), which uses less of your memory at a small cost in accuracy. Each GPU gets its own connection; the CPU is used for models only when there is no GPU.
• Matrix multiplications — single tasks checked against two other volunteers, on whichever of your devices is fastest.

To do only some of them, set work under [general] in volunteer.ini, for example work = models or work = models, boxes.

Hardware it can use

Any CPU works with nothing extra. Everything else needs the library for that hardware; --list-devices shows what was found and suggests what to install for anything it saw but could not use. For running models, a GPU needs PyTorch; the other libraries below speed up matrix multiplications.

Each device is checked for correct answers before it is used, and anything that fails the check is left alone. A card reachable through two libraries is only driven through the faster one.

Commands

python3 volunteer.py                  # start volunteering
python3 volunteer.py --list-devices   # show what it found and how fast each device is
python3 volunteer.py --benchmark      # measure every device again
python3 volunteer.py --install        # install the required and optional packages
python3 volunteer.py --headless       # in the background: no console, lowest priority, logs to ~/.vtmu/volunteer.log
python3 volunteer.py --install-service    # start in the background whenever you log in (and now)
python3 volunteer.py --uninstall-service  # stop doing that
python3 volunteer.py --config FILE    # use a different settings file
python3 volunteer.py --help

Stopping and removing it

Ctrl+C, or quit, stops it; nothing keeps running afterwards. If you set it to start at login, --uninstall-service stops it and undoes that. To remove it completely, delete volunteer.py, volunteer.ini, and the .vtmu folder in your home folder, which holds a random volunteer id and the speed measurements.

Getting credit for your work

Every account has a public volunteer code (it starts with vc_), shown in the volunteering panel on /account and /compute. A copy downloaded while you are signed in already has yours in it; to change it, put it in volunteer.ini or on the command line:

[general]
account = vc_xxxxxxxxxxxx

python3 volunteer.py --account vc_xxxxxxxxxxxx   # overrides the file for this run

The code goes to the grid with the program's other messages and tells it which account to credit. It is not a password and the grid never shows it to other volunteers, but anyone who reads your volunteer.ini could credit work to you; rotate it on the panel if that worries you (the old one stops earning at once). Work the grid accepts earns 25% of what it would have cost a paying user, in credits for mPAKT Compute, up to a daily cap; work for your own account's models, tasks or boxes earns nothing. Over the last 30 days your contribution sets a tier (Bronze, Silver, Gold): a discount on what your boxes cost and bonus credits on card purchases. Without a code, or with one the grid does not know, the program works the same and nothing is credited.

Running other people's web services (Docker)

On Linux with Docker, the volunteer program can also run boxes for mPAKT Compute: small web services from public GitHub repositories, each in its own container. This is on by default (work = models, matmul, boxes in volunteer.ini; remove boxes to turn it off) and happens only if Docker works. The program itself is for verified accounts, so only verified accounts' volunteers run boxes. The Linux installer does not set Docker up for you.

• Use rootless Docker. A rootless Docker daemon runs as your own unprivileged user, so even a container escape would land in an account that can do almost nothing. Install docker-ce-rootless-extras and uidmap, run dockerd-rootless-setuptool.sh install, then systemctl --user enable --now docker. The program finds the socket by itself. Do not add yourself to the docker group for this: that group is equivalent to root. A normal (root) Docker is accepted with a warning. Docker has to be able to enforce memory and process limits, or boxes stay off, and the program says why.
• Each box is locked down. A read-only filesystem, no Linux capabilities, no extra privileges, memory, process and CPU limits (the CPU share you set), its own private temporary space, and no network of its own: it cannot reach this computer, your network or the internet. If its author turned internet on, what it sends travels over the volunteer program's connection to the grid, which sends it onward through Tor; nothing a box does leaves from your address.
• What does use your connection. Docker itself downloads the images a box needs, from Docker Hub and a few well-known public registries, over your connection. So does the program, for the box's own code: it fetches the box's repository from GitHub (the commit's archive from codeload.github.com, or its files from raw.githubusercontent.com) instead of receiving it from the grid, so GitHub sees your IP address. Before anything runs, every file is checked against the size and SHA-256 the grid recorded for that commit; if anything differs, nothing is run and the program asks the grid for the code instead.
• Honest limits. Containers share your computer's kernel, so a serious Docker or kernel bug could still let one out; rootless Docker limits what it could reach, it does not remove the risk. A box can use the CPU and memory it was given while it runs. Up to four boxes run at once and an idle one is stopped after five minutes.

Other ways to volunteer

Websites can let their visitors contribute from the browser with one script tag; see the volunteer script. A browser doing so also contacts GitHub, the npm and PyPI registries and a public content-delivery network from its own connection, as the participation agreement describes.

Running the volunteer program means taking part under the participation agreement, which describes what runs on your device.