gpumesh

Borrow your friends' GPUs. A distributed compute mesh that lets you share GPU power across machines on your network — with one decorator, one CLI command, or a Python API.

Quickstart · Docs · Contributing · Issues

What is gpumesh?

gpumesh turns multiple machines into a single, unified compute pool. Start a coordinator on one machine, join workers from other machines (laptops, desktops, servers — anything with Python), and run code across all of them as if they were one device.

Use cases:

• Hyperparameter search across multiple GPUs
• Data preprocessing sharded across machines
• Model training on a pool of consumer GPUs
• Any embarrassingly parallel workload

Key features

• @mesh / @accelerate decorators — mark a function and it runs on the pool; no job system, no ceremony
• .map() — spread one call across every connected machine at once
• Smart routing — calls go to a mesh worker when one is alive, to your own machine when none is
• Graceful fallback — mesh unreachable? your code runs locally and returns the same value
• Any return value — numpy arrays, torch tensors, DataFrames — exactly what your function returned
• Fault tolerance — workers survive sleep, WiFi drops, and coordinator restarts; dead workers' tasks are re-queued
• Benchmark scoring — every worker gets a relative compute score; the scheduler routes work to the strongest hardware
• Memory-aware scheduling — VRAM is tracked; a payload's gpu_memory_mb hint keeps a task off a busy worker
• Live radar — gpumesh radar discovers nearby devices on your network; no config needed
• Isolated execution — every task runs in its own subprocess; a crashing task can't take down a worker
• Token security — all API calls require a token; rate-limited, timing-safe verification
• Jupyter support — %%mesh cell magic wraps every function in a cell automatically

[!CAUTION] gpumesh runs code you send it — there is no sandbox. A worker executes arbitrary Python as the OS user that started it, and results are deserialized by the submitter — trust runs both ways. Use it only with machines and people you trust. See SECURITY.md and THREAT_MODEL.md.

Quick demo

from gpumesh import GPUMesh, accelerate

mesh = GPUMesh("http://coordinator:8000", token=TOKEN)

@accelerate(mesh)
def train(lr, epochs):
    return {"accuracy": 0.95}

result = train(lr=0.01, epochs=100)         # one mesh worker (or local if none)

results = train.map([                        # spread across all mesh devices
    {"lr": 0.01, "epochs": 100},
    {"lr": 0.05, "epochs": 200},
])

Installation

pip install gpumesh
pip install gpumesh[gpu]       # GPU detection + CUDA benchmarks
pip install gpumesh[tunnel]    # ngrok for public URLs
pip install gpumesh[sysinfo]   # System info (psutil)
pip install gpumesh[notebook]  # DataFrame support (pandas)
pip install gpumesh[ui]        # Setup wizard (rich + questionary)
pip install gpumesh[all]       # Everything above

Or with Docker:

docker pull k4-labs/gpumesh:3.0.0

Requires: Python 3.9+, cloudpickle (auto-installed). PyTorch is optional (needed for GPU detection).

Verify the install:

$ gpumesh --version
gpumesh 3.0.0 (3.11.9, Windows)

Quickstart

One machine is enough. gpumesh serve joins your own CPU/GPU to the pool by default, so a mesh of one runs exactly the same code path as a mesh of ten — same scheduler, same wire format, same results. Get that working first; adding a second machine is one flag and one command.

All terminal output below is real, copied from actual runs, with addresses and hostnames replaced by placeholders.

1. Generate a token

python -c "import secrets; print(secrets.token_urlsafe(32))"

Do this once and reuse the value. It is not decoration: the token is the only thing standing between anyone who can reach the port and code execution as you. Short, memorable tokens are guessable, and the coordinator's hash is a single SHA-256 round with no KDF — the entropy of the token is the whole defence.

2. Start a coordinator

pip install gpumesh
gpumesh serve --port 8000 --token $TOKEN

[gpumesh] config saved to ~/.gpumesh/config.json
[OK] Coordinator listening on 127.0.0.1:8000
   Token: <your token>

   Join from THIS machine:   gpumesh join http://127.0.0.1:8000 --token <your token>
   Other machines CANNOT reach this coordinator (bound to 127.0.0.1 only).
   That is the default now: a worker executes code as the user who started it.
   To let other machines join, re-run with: gpumesh serve --host 0.0.0.0 --port 8000 --token <your token>
   (or set GPUMESH_HOST=0.0.0.0). Read the warning it prints before you do.
   Ctrl+C to stop
[OK] Self-check: server is up (tested on 127.0.0.1 only)
[OK] Self-worker started — this machine's CPU/GPU is part of the pool
   (disable with --no-self-worker)
[worker] device=CPU Intel64 Family 6 Model 154 Stepping 3, GenuineIntel score=0.373 GFLOP/s
[mesh] worker joined: laptop-a (Intel64 Family 6 Model 154 Stepping 3, GenuineIntel, score=0.373)
[worker] joined mesh as 0afecb2dfd44

You now have a working mesh. Confirm it in another terminal:

$ gpumesh workers

  WORKERS
  ------------------------------------------------------
  CPU  0afecb2d    laptop-a          score=0.373     [alive]

Prefer a guided wizard? Run gpumesh setup.

3. Run a distributed task

Save as demo.py:

from gpumesh import mesh          # auto-connects using the saved config

@mesh
def train(lr, epochs):
    return {"lr": lr, "accuracy": round(0.90 + lr, 3)}

print(train(lr=0.01, epochs=100))                              # one worker
print(train.map([{"lr": 0.01, "epochs": 100},                  # every worker
                 {"lr": 0.05, "epochs": 200}]))

$ python demo.py
[gpumesh] connected to coordinator at http://127.0.0.1:8000
{'lr': 0.01, 'accuracy': 0.91}
[{'lr': 0.01, 'accuracy': 0.91}, {'lr': 0.05, 'accuracy': 0.95}]

That is the whole workflow. Kill the coordinator and run it again — you get the identical output, computed locally. Your script does not break because the mesh went away.

More runnable examples, all verified: examples/.

4. Add a second machine

The coordinator binds to loopback, so nothing outside your machine can reach it. That is the default, and opening it up is a decision rather than an oversight — a worker runs whatever code the coordinator hands it, as the OS user who started that worker. Restart with --host 0.0.0.0 when you mean it:

gpumesh serve --host 0.0.0.0 --port 8000 --token $TOKEN

   Join from THIS machine:   gpumesh join http://127.0.0.1:8000 --token <your token>
   Join from ANOTHER machine: gpumesh join http://192.0.2.10:8000 --token <your token>
   (127.0.0.1 always works locally; the LAN IP is for other machines)

  !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
  !! NETWORK-EXPOSED COORDINATOR
  !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
   Bound to:      0.0.0.0:8000 (not loopback — other machines can connect)
   Tasks run as:  alex
   On device:     CPU Intel64 Family 6 Model 154 Stepping 3, GenuineIntel

   Anyone who can reach this port AND has the token can run
   arbitrary code on this machine as alex.
   Only do this on a network you trust, and keep the token secret.
   Loopback-only (the default) is: gpumesh serve --port 8000
  !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!

   Reachability of http://192.0.2.10:8000 from other machines is not tested here — a worker joining confirms it.

Copy the Join from ANOTHER machine line — that is the exact command the second machine needs.

# on machine B
pip install gpumesh
gpumesh join http://192.0.2.10:8000 --token $TOKEN

[worker] device=CPU Intel64 Family 6 Model 154 Stepping 3 score=0.310 GFLOP/s
[worker] joined mesh as 7c1e04ab93f2

$ gpumesh workers

  WORKERS
  ------------------------------------------------------
  CPU  0afecb2d    laptop-a          score=0.373     [alive]
  CPU  7c1e04ab    laptop-b          score=0.310     [alive]

Workers never die. They survive laptop sleep, WiFi drops and coordinator restarts, and reconnect on their own.

Windows: run gpumesh serve as Administrator so the firewall rule is added for you. Without it, workers on other machines may be blocked.

If the worker reports a timeout instead, it could not reach that address — see Troubleshooting.

examples/second_machine.py reads your live mesh and prints the exact command, filled in with your address, port and token.

Prefer the CLI?

Submit a script instead of decorating a function. (The examples/ files ship with the repository, not the pip package — git clone first, or point the command at any script of your own that reads a JSON payload on stdin.)

$ gpumesh submit examples/grid_search.py --payloads examples/payloads.json --wait
[OK] Submitted job c3f81cc498f0
Job: examples/grid_search.py (c3f81cc498f0)

  #################### 100%  6/6 done  (10s)

  Status: finished
  Counts: {'done': 6}

  ✓ task 463a58e92490 [done]  cost=1.0  worker=0afecb2dfd44
    result: {"lr": 0.01, "epochs": 100, "l2": 0.0, "val_accuracy": 0.948, "weights": [0.2271, -0.2757, 0.0019]}
  ✓ task 7fa79a95725f [done]  cost=10.0  worker=0afecb2dfd44
    result: {"lr": 0.3, "epochs": 1000, "l2": 0.01, "val_accuracy": 0.952, "weights": [2.1907, -2.9448, 0.0191]}

Code normally

This is the entire point. Once a worker is connected, every machine sees the same pool, and the demo.py above is already the whole API — write normal Python, mark the heavy functions, call .map() when you have a batch. It works in VS Code, Jupyter, PyCharm, or a plain terminal. No job submission, no CLI commands, no ceremony.

Your function returns whatever it normally returns — a dict, an int, a list, a numpy array, a torch tensor — and you get that same object back:

@mesh
def evaluate(seed):
    import numpy as np
    return {"loss": np.float32(0.12), "preds": np.arange(10)}

out = evaluate(seed=1)      # {'loss': np.float32(0.12), 'preds': array([0, ..., 9])}

Jupyter notebooks

Load the extension once, in its own cell:

%load_ext gpumesh

Then %%mesh as the first line of any cell wraps every function defined in that cell with @mesh:

%%mesh
def preprocess(chunk_id, rows):
    return {"chunk": chunk_id, "rows": rows * rows}

results = preprocess.map([{"chunk_id": i, "rows": 100 + i} for i in range(6)])

Like %%time, the cell's own output displays normally. Loading the extension also injects a bare @mesh decorator into the notebook namespace, so you can decorate individual functions instead of a whole cell.

CLI reference

Every command — server, jobs, monitoring — plus the environment variables behind them, with --host vs --host-ip and colour handling explained: docs/cli.md.

Python API

from gpumesh import GPUMesh

mesh = GPUMesh("http://coordinator:8000", token=TOKEN)

Distribute a function

results = mesh.distribute(
    function=train_model,
    params=[{"lr": 0.01, "epochs": 100}, {"lr": 0.05, "epochs": 200}],
    timeout=600,
)

Inspect the pool

workers = mesh.workers()        # [{'id', 'device', 'device_name', 'hostname', 'score', 'alive'}]
devices = mesh.devices()        # unified pool view
count   = mesh.device_count()   # alive machines contributing compute (GPU or CPU)
gpus    = mesh.gpu_count()      # alive GPUs only
total   = mesh.total_score()    # combined compute score
best    = mesh.auto_device()    # most powerful alive device

Job management

job_id = mesh.submit(name="preprocess", script="process.py",
                     payloads=[{"file": "data.csv"}])
status = mesh.status(job_id)
df     = mesh.results_to_dataframe(results)   # requires pandas

From Python, non-blocking

GPUMesh.start_coordinator(port=8000, token=TOKEN)   # binds 127.0.0.1; pass host="0.0.0.0" to open it up
GPUMesh.add_worker("http://coordinator:8000", token=TOKEN)

@accelerate patterns

from gpumesh import GPUMesh, accelerate

mesh = GPUMesh("http://coordinator:8000", token=TOKEN)

# Basic
@accelerate(mesh)
def preprocess(chunk_id, data_path):
    import pandas as pd
    df = pd.read_parquet(data_path)
    return {"chunk": chunk_id, "rows": len(df)}

# Hardware selection — only run on an A100
@accelerate(mesh, gpu="A100")
def train(model):
    return model.cuda().forward(x)

# Resource specs
@accelerate(mesh, cores=8, memory="16GB", timeout=300)
def heavy_computation(data):
    return processed

# Batch: spread across every device
results = train.map([{"lr": 0.01}, {"lr": 0.05}])

# Bind to a specific device
gpu_predict = predict.to("cuda")
result = gpu_predict(x)

# Global install — @accelerate with no arguments
accelerate.install(mesh)

@accelerate
def train(lr, epochs):
    return {"accuracy": 0.95}

Smart routing

Either path returns the identical value, so switching between them never changes your results. Note that gpumesh serve joins your own machine to the pool by default, so a single call is dispatched through the mesh even when you are the only participant — pass --no-self-worker if you want it to stay purely local.

Docker

A prebuilt image is available on Docker Hub (k4-labs/gpumesh).

First, a token:

export GPUMESH_TOKEN=$(python -c "import secrets; print(secrets.token_urlsafe(32))")

There are two host decisions in a container, and they are not the same one:

1. What the coordinator binds to inside the container. It must be 0.0.0.0, because anything reaching it arrives from outside the container — a loopback bind inside a container answers nothing, including the port Docker published.
2. What Docker publishes on the machine running Docker. This is the one that decides who can reach you. -p 8732:8732 binds 0.0.0.0 on the host, which puts a coordinator on every network the machine is attached to. Prefix the host side with 127.0.0.1 unless you mean otherwise.

# Coordinator — reachable from this machine only
docker run -d --name gpumesh-coordinator \
  -p 127.0.0.1:8732:8732 -p 127.0.0.1:48900:48900/udp \
  -e GPUMESH_TOKEN \
  k4-labs/gpumesh:3.0.0 \
  serve --host 0.0.0.0 --port 8732

# Worker
docker run -d --name gpumesh-worker \
  -e GPUMESH_TOKEN \
  k4-labs/gpumesh:3.0.0 \
  join http://coordinator-ip:8732

Drop the 127.0.0.1: prefix — or better, name a specific interface, e.g. -p 192.0.2.10:8732:8732 — only when you deliberately want other machines to join. That is the same decision gpumesh serve --host 0.0.0.0 asks you to make on the host, and it deserves the same pause.

Both serve and join read GPUMESH_TOKEN from the environment, so -e GPUMESH_TOKEN is enough and the token never appears in docker ps output. join takes the coordinator URL as a positional argument, so there is no GPUMESH_URL to set here — that variable is only read by the job and monitoring commands.

Or use the included docker-compose.yaml for a coordinator + N workers, with a healthcheck the workers wait on before starting. It pulls the published image, so the file works on its own:

GPUMESH_TOKEN=$TOKEN docker compose up -d                      # localhost only
WORKER_REPLICAS=4 GPUMESH_TOKEN=$TOKEN docker compose up -d    # scale workers
GPUMESH_BIND=0.0.0.0 GPUMESH_TOKEN=$TOKEN docker compose up -d # reachable from the LAN

GPUMESH_TOKEN is required — compose stops with a clear message if it is unset, rather than starting a coordinator with a random token that no worker could authenticate against. GPUMESH_BIND defaults to 127.0.0.1 and is the host-side publish address from decision (2) above.

Ports: 8732 (TCP API) and 48900/udp (LAN discovery). Persistent state (the SQLite database and ~/.gpumesh/config.json) lives under /data; the compose file mounts a named volume there.

The container listens on 8732, while gpumesh serve on the host defaults to 8000. That is deliberate — the image pins an explicit port so published docker run and compose recipes stay stable. Both are just defaults: pass --port to use whatever you like, and make sure workers point at the same number the coordinator is listening on.

Network options

• LAN: gpumesh serve --host 0.0.0.0 + gpumesh join http://192.0.2.10:8000 --token $TOKEN. UDP broadcast helps workers find a coordinator, but joining still needs the token and a coordinator that is actually bound beyond loopback.
• Tailscale: gpumesh serve --host 0.0.0.0 --port 8000 --tailscale, then join via the Tailscale IP. This is the option to reach for if the mesh crosses any network you do not control — the tunnel, not the token, becomes the boundary.
• ngrok: gpumesh serve --host 0.0.0.0 --port 8000 --public prints a public https://... URL that workers anywhere can join. "Anywhere" includes people you did not invite; the token is the only thing keeping them out.

The default — no flag at all — is loopback only. Every row above is an opt-in to a wider audience.

Architecture

Jobs are stored in SQLite; workers pull tasks over HTTP with a lease (a crashed worker's task is automatically re-queued), run each task in an isolated subprocess, and post results back. Workers are scored by a benchmark and the scheduler assigns heavier tasks to stronger workers. Diagram and job flow: docs/architecture.md.

Security

Read that table as "what raises the cost of an attack", not as "what makes this safe". None of it is a sandbox.

The token is a licence to execute code, not a password guarding data. Anyone holding your URL and token can run arbitrary Python on every machine in your mesh, as the user who started each worker. And it runs both ways: a worker's result is deserialized on the submitting machine, so a hostile worker executes code on you. gpumesh is built for trusted networks — home labs, lab benches, your own machines, a team you know.

Use a real token: python -c "import secrets; print(secrets.token_urlsafe(32))". The coordinator hashes it with a single SHA-256 round and a deterministic salt, which is fine for 250 bits of entropy and useless for a token you chose by hand. Entropy is the whole defence.

Run the coordinator with --safe-mode to refuse function distribution and accept submitted scripts only. That closes the pickle path; it does not stop a submitted script from doing anything a script can do.

Rate limiting exempts loopback deliberately. Anyone who can open a socket from 127.0.0.1 can read the token out of the process and the config file rather than guess it, so a lockout there costs an attacker nothing and costs you your own mesh.

Traffic is not encrypted on a plain LAN. Use --tailscale or --public (ngrok) when the mesh crosses a network you do not control.

Full detail: SECURITY.md (what a token grants, how to report a vulnerability) and THREAT_MODEL.md (the same flow traced through the code, file and line).

Benchmark scoring

Each worker runs a benchmark on join and gets a score of gflops * 0.7 + bandwidth_gbps * 0.3. It is a relative, unbounded number — there is no normalisation and no ceiling, so it only means anything next to the other workers in your pool. The scheduler uses it to rank workers, nothing else.

Rough magnitudes, to read the numbers gpumesh workers prints:

Those are illustrative, not a scale — a laptop CPU commonly lands around 0.24, and a faster GPU than anything listed here would simply score higher.

Troubleshooting

The usual suspects — bind vs firewall, mismatched tokens, cross-version tasks, missing torch on a worker — in one table, plus how to read gpumesh doctor: docs/troubleshooting.md.

Development

git clone https://github.com/K4-LABS/gpumesh.git
cd gpumesh
pip install -e ".[dev]"
pytest                 # on Windows a few skip and one xpasses
python -m build        # build wheel + sdist

CI runs the suite on every push and pull request, plus a weekly scheduled run — that last one exists because the regressions this project actually hits come from dependencies changing under it, not from commits here. The matrix is Linux on every Python gpumesh claims to support, with Windows and macOS spot checks; the authoritative list is .github/workflows/tests.yml, and the tests badge at the top of this page is that matrix's live result, so neither can drift the way a hand-copied list does.

Contributing

Issues and pull requests are both welcome, and small ones are welcome too — a typo fix counts.

• CONTRIBUTING.md — dev setup, how the pieces fit together, and what makes a change easy to review
• Good first issues — scoped starting points that name the file to change
• Issue tracker — bugs, and questions about whether something is worth doing before you write the code

Documentation

Limitations

• Python only — tasks must be Python functions or scripts
• Arguments and return values must be picklable. Anything tied to a live process — open files, sockets, locks, database handles, CUDA handles — cannot cross machines. Return plain data (numbers, arrays, tensors, DataFrames) instead
• Every worker needs the imports your function uses already installed; gpumesh ships your code, not your environment
• Workers should run the same Python minor version as the submitter — cloudpickle falls back to source when they differ, which does not cover every function
• No GPU memory sharing — each task gets its own process
• No model sharding — each task runs on one machine at a time
• No task graphs — every task is independent; if task B needs task A's output, that plumbing is yours
• No distributed data structures — no shared arrays, dataframes or object store. If your data does not fit on one machine, gpumesh will not help
• Built and tested for meshes of roughly 2–20 machines. Nobody has run it at hundreds of workers
• Single coordinator — single point of failure (use Tailscale for reliability)
• No built-in encryption — use Tailscale for encrypted tunnels
• Trusted networks only — workers run whatever code the coordinator sends, and submitters deserialize whatever workers return

If several of those are dealbreakers, read Why not Ray or Dask? — it says plainly which tool to use instead.

Prior art and credits

gpumesh borrows shamelessly, at the level of API shape and scheduling strategy. No code was copied from any of these projects; what was taken is the idea of what a good interface looks like, and each one is credited in the source where its influence lands.

exo (Apache-2.0) is the closest relative in spirit — running AI workloads on the consumer hardware you already own. Its runner-supervisor pattern is why every gpumesh task runs in its own subprocess, and its crash diagnostics and topology-change events are reflected in worker.py and db.py. hivemind (MIT) supplied the TTL-based worker expiry that prunes a machine that has stopped heartbeating. Petals (MIT) supplied straggler deprioritisation — a worker slower than twice its peers' median gets lighter tasks rather than holding up the batch. cudf.pandas (Apache-2.0) is where accelerate.install(mesh) comes from: an import hook that makes existing code use the accelerator without editing it. Hugging Face Accelerate (Apache-2.0) is the source of the .to(device) placement idea. clustrix (MIT) is where @accelerate(cores=8, memory="16GB") gets its shape — declaring resource requirements on the decorator rather than in a separate config. burla contributed the remote_parallel_map batch pattern behind .map() and the func_gpu="A100" hardware-selection idea; note that burla is licensed FSL-1.1-Apache-2.0, which is source-available, not OSI open source, and reads back to Apache-2.0 after two years. Two smaller influences round it out: distry, for the plain single-function decorator, and ezpz, for setup_torch()-style automatic backend detection.

If you maintain one of these and think the credit is wrong, or the influence is closer to copying than we believe, open an issue and it will be corrected.

Maintainers

Samurai007AK

jinia-konar

License

GNU AGPL-3.0-or-later. See LICENSE for details.

gpumesh — a distributed compute mesh
Copyright (C) 2026 K4-LABS

This program is free software: you can redistribute it and/or modify
it under the terms of the GNU Affero General Public License as published
by the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.

This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
GNU Affero General Public License for more details.

You should have received a copy of the GNU Affero General Public License
along with this program.  If not, see <https://www.gnu.org/licenses/>.

That notice lives here rather than at the top of LICENSE on purpose. GitHub detects a project's license by matching the whole LICENSE file against the known texts, and a prepended notice pushes the match below its threshold — the repository then reports its license as "Other", which is exactly the wrong signal for a copyleft project. LICENSE is now the unmodified AGPL-3.0 text, and the "or (at your option) any later version" grant above is what the AGPL-3.0-or-later in pyproject.toml reports.

gpumesh was MIT-licensed through 2.0.0 and moves to the AGPL in the next release. That change is not retroactive: every release published up to and including 2.0.0 remains available under the MIT terms it was published under. If you run a modified gpumesh as a network service, AGPL section 13 requires you to offer your modified source to its users.

GitHub · Issues · PyPI