gpumesh 0.8.0

Borrow your friends' GPUs: a terminal-based distributed compute mesh in pure Python

gpumesh

Borrow your friends' GPUs. A distributed compute mesh that lets you share GPU power across machines.

---

Features

• Transparent acceleration — Add @accelerate to any function, it uses all connected GPUs automatically
• One-line setup — gpumesh setup handles everything
• Auto-discovery — Workers find coordinators on the same network
• Smart scheduling — Fast GPUs get heavy tasks, slow ones get light tasks
• Fault tolerance — Dead workers are detected, tasks are re-queued
• Python API — Distribute functions from Jupyter notebooks

---

Installation

pip install gpumesh

Optional extras:

pip install gpumesh[gpu]       # GPU detection + CUDA benchmarks
pip install gpumesh[tunnel]    # ngrok for public URLs
pip install gpumesh[ui]        # Beautiful setup wizard (rich + questionary)
pip install gpumesh[all]       # everything

Requires Python 3.9+.

---

Quick Start

1. Start a coordinator (one machine)

gpumesh setup
# Choose option 1 (Coordinator)
# The wizard detects your hardware and shows a radar

2. Join a worker (another machine)

gpumesh setup
# Choose option 2 (Worker)
# Enter a token, start broadcasting
# The coordinator claims you from the radar

3. Use transparent acceleration

from gpumesh import GPUMesh, accelerate

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

@accelerate(mesh)
def train(lr, epochs):
    # Your code here — runs on all connected GPUs automatically
    return {"accuracy": 0.95}

# Single call → best local device
result = train(lr=0.01, epochs=100)

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

---

Transparent Acceleration

The @accelerate decorator makes your mesh resources transparent to your code:

from gpumesh import GPUMesh, accelerate

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

@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)}

# Single call → runs locally on best device
result = preprocess(chunk_id=0, data_path="data.parquet")

# Batch call → spreads across ALL mesh devices
results = preprocess.map([
    {"chunk_id": 0, "data_path": "part0.parquet"},
    {"chunk_id": 1, "data_path": "part1.parquet"},
])

How it works

Scenario	What happens
Single call func(x)	Runs on best local device (CPU/GPU)
Batch call func.map([...])	Spreads across all mesh devices
Mesh unreachable	Falls back to local execution silently
GPUMESH_LOCAL=1	Forces local-only (no mesh)
GPUMESH_VERBOSE=1	Prints which device handled each task

---

CLI Commands

Command	Description
gpumesh setup	Interactive setup wizard
gpumesh serve	Start coordinator
gpumesh join URL	Join as worker
gpumesh quickjoin	One-click: detect GPU and join mesh
gpumesh radar	Scan for nearby gpumesh devices
gpumesh worker	Start a worker that broadcasts and waits to be claimed
gpumesh submit SCRIPT --payloads FILE	Submit a job
gpumesh status JOB_ID	Check job progress
gpumesh cancel JOB_ID	Cancel a job
gpumesh workers	List connected workers
gpumesh devices	Show all GPUs as one pool
gpumesh show-connection	Show saved URL and token for sharing
gpumesh disconnect	Clear saved connection
gpumesh kill	Kill all gpumesh tasks (graceful or force)

---

Python API

from gpumesh import GPUMesh

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

# List workers
mesh.workers()

# Distribute a function across all workers
results = mesh.distribute(
    function=train_model,
    params=[
        {"lr": 0.01, "epochs": 100},
        {"lr": 0.05, "epochs": 200},
    ],
)

---

Network Options

Method	Setup	Best for
LAN	None	Same Wi-Fi, fastest
Tailscale	Install Tailscale	Remote teams
ngrok	pip install gpumesh[tunnel]	Public access

---

Limitations

• Python only — Tasks must be Python scripts
• No GPU memory sharing — Each task gets its own process
• No model sharding — Each task runs on one machine
• Single coordinator — Single point of failure

---

Troubleshooting

Problem	Fix
command not found	Use python -m gpumesh instead
401 bad token	Check coordinator and worker tokens match
coordinator unreachable	Check firewall and that coordinator is running
task timed out	Increase --timeout 600 or split into smaller tasks

---

Security

• Token authentication on all API requests
• Tokens stored with restricted file permissions (0o600 on Unix, icacls on Windows)
• Rate limiting after 5 failed attempts
• Process isolation for all tasks

Important: Workers execute code from the coordinator. Only share your URL and token with people you trust.

Security Model

gpumesh is designed for trusted networks (home labs, team clusters). Key security considerations:

• Code execution: Function tasks (@accelerate) execute in the worker's process. Only connect to machines you trust.
• Plaintext HTTP: All communication uses HTTP. Use Tailscale (--tailscale) for encrypted tunnels across untrusted networks.
• Token authentication: A shared token authenticates all communication. Keep it secret.
• No sandbox: Tasks have full access to the worker machine. Do not run untrusted code.

---

Contributing

git clone https://github.com/Samurai007AK/gpumesh.git
cd gpumesh
pip install -e ".[dev]"
pytest

---

License

MIT License. See LICENSE for details.