m-gpux

The GPU workbench for Modal.

m-gpux is a CLI for getting real work done on Modal GPUs. It turns a folder on your laptop into a GPU dev box, a Jupyter session, a hosted web app, an OpenAI-compatible LLM endpoint, or a whole Docker Compose stack — across every Modal account you own, with the price per hour in front of you before anything starts.

Quickstart

pip install m-gpux         # also installs the `modal` CLI that m-gpux drives
m-gpux account add         # paste a `modal token set ...` command

cd my-project
m-gpux dev up              # pick an account and a GPU; this folder lands in /workspace
m-gpux dev code            # open it in VS Code (Remote-SSH)
m-gpux dev pause           # snapshot everything, stop paying for compute

If you'd rather click through a menu, m-gpux hub launches Jupyter, a script run, a browser terminal, or vLLM on any GPU from one guided wizard.

For the same workflows inside your editor, install the m-gpux VS Code extension.

m-gpux ecosystem

The CLI works on its own, and each piece below plugs into the same accounts, presets, and budgets.

• CLI — dev boxes, Hub, hosting, serving, Compose, budgets, and billing from one command
• VS Code extension — accounts, sessions, presets, and the Hub / host / serve wizards in a sidebar
• Documentation — guides for every workflow, plus the full command reference
• Examples — ready-to-deploy FastAPI, Flask, and static-site projects for m-gpux host

Why use m-gpux?

m-gpux wraps Modal's primitives (Sandboxes, Functions, Servers, Volumes, named Images) into workflows that fit how you already develop.

• Dev boxes that survive the night — Sandboxes running sshd, reachable with plain ssh or VS Code Remote-SSH. pause snapshots the whole filesystem and resume boots it back in seconds; sync push/pull moves files both ways
• Every GPU, priced up front — T4 through H100, H200, B200, and B300, including multi-GPU containers (H100:8), with $/hour in every picker and m-gpux billing rates for the live price table
• Many accounts, one budget — per-account monthly limits, budget-aware AUTO selection, a warning before you launch on an account that's nearly empty, and budget check --stop to shut down apps on accounts that go over
• From localhost to a URL — m-gpux host deploys FastAPI, Flask, Django, or a static folder with scale-to-zero; --server uses Modal's low-latency @app.server() primitive
• Your own OpenAI-compatible endpoint — m-gpux serve deploy puts any Hugging Face model behind vLLM with API keys, streaming, warmup, logs, and a live metrics dashboard
• Kaggle & Colab too — m-gpux kaggle and m-gpux colab drive Kaggle notebooks and Colab runtimes across many accounts: weekly GPU hours and compute units at a glance, AUTO picks the account with the most left, kernels pushed as any account (Colab CLI works on Windows)
• A local control plane — m-gpux ui opens a web dashboard on localhost: every account's credit and workspace, live apps, containers and sandboxes across accounts, sessions, endpoints and GPU prices, with stop / logs / budget actions
• Scripts as an async API — m-gpux gateway deploy exposes the tasks you declare in mgpux.gateway.toml as an authenticated job API: submit JSON, poll for results, each task autoscaling on its own CPU or GPU
• Docker Compose, lifted as-is — run a docker-compose.yml as one container, a full VM-style image (Triton, gRPC), or one Sandbox per service with readiness probes and working service-to-service addresses
• Starts in seconds, not minutes — m-gpux image build bakes your dependencies once and publishes them to every account; Hub and dev boxes offer them as the base image
• Nothing hidden — every workflow generates a plain Modal script you can read and edit before it runs

Resources

• Documentation — getting started, guides, and FAQ
• Command reference — every command and option
• Dev boxes — SSH, VS Code, pause/resume, and sync in depth
• Costs, budgets & images — GPU catalog, prices, budgets, and published images
• Docker Compose guide — subprocess, VM, and sandbox modes
• Architecture — plugins, the generated-script pattern, and how third-party plugins add commands
• Issues — bug reports and feature requests

Contributing

git clone https://github.com/PuxHocDL/m-gpux.git
cd m-gpux
pip install -e ".[dev,docs]"
m-gpux --help

Each command group is a plugin in m_gpux/plugins/<name>/plugin.py, registered in both m_gpux/plugins/__init__.py and the m_gpux.plugins entry-point group in pyproject.toml.

Run the same checks used by CI before opening a pull request:

ruff check m_gpux tests
python -m unittest discover -s tests -v
mkdocs build --strict

cd m-gpux-vscode
npm ci
npm run typecheck
npm run compile

cd ../m-gpux-ui
npm ci
npm run build

License

MIT