veloxml

veloxml init [name] [-g | --gpu ACCEL]veloxml checkveloxml deploy [--dry-run] [--cloud aws|gcp] [--region REGION] [-g | --gpu ACCEL] [-v | --verbose]veloxml statusveloxml logs [name] [-t | --tail N] [-f | --follow]veloxml down [-n | --name name] [-a | --all]veloxml version

veloxml is the command-line interface for VeloxML, a thin Python wrapper around SkyPilot Serve. It deploys a FastAPI (or similar) model service into your AWS or GCP account rather than onto a third-party inference host.The usual loop is init (or hand-write `app.py` + `veloxml.yaml`), check, deploy, then down when finished. deploy generates a SkyPilot service spec under `.veloxml/`, installs `requirements.txt` plus an optional `runtime.setup` command on the VM, starts `uvicorn app:app --host 0.0.0.0 --port 8000` by default, and waits for `/health`. Re-running deploy against an already-running name performs a rolling `sky serve update`.The PyPI package is veloxml-deploy; the console script it installs is veloxml. It depends on skypilot (the `sky` binary), truss, and typer. Python 3.10+ is required. The project is Apache-2.0 licensed.

-g, --gpu ACCEL

GPU accelerator string understood by SkyPilot, e.g. T4:1, A10G:1, L4:1. On init it is stored in `veloxml.yaml`; on deploy it overrides the file for that run.

deploy only. Write the SkyPilot spec and exit without provisioning. Prints a dummy `http://127.0.0.1:8000` endpoint.

deploy only. Target cloud. Default aws.

deploy only. Cloud region. Default us-east-1.

deploy only. Stream SkyPilot's full output and re-raise exceptions instead of collapsing them into a short error box.

logs only. Number of trailing lines to print. Default 100. Ignored with --follow.

logs only. Stream live logs (passes through to `sky serve logs`).

down only. Service to tear down instead of the name in `veloxml.yaml`.

down only. After stopping the service, purge every SkyPilot cluster (`sky down -a -y`) so leftover controllers stop billing.

Typer help for the CLI or a subcommand. Running veloxml with no arguments also prints help.

veloxml.yaml

Per-project spec in the working directory. Loaded by deploy, logs, and down; written by init. Missing keys take the defaults below.

Service name passed to SkyPilot (default my-model-service).

Cloud and region (defaults aws, us-east-1). Overridden by --cloud / --region on deploy.

Optional GPU, e.g. `T4:1`. Set by --gpu.

Size floors with a trailing `+` (defaults 2+ / 4+). Values without `+` get it appended when the spec is compiled.

Request spot instances. Default false in the schema; README examples set true.

Published port (default 8000).

HTTP paths (defaults /health and /predict). The readiness probe is given a 240 s initial delay.

Replica policy (both default 1). Scale-to-zero is not the compiled default; set min_replicas to 0 if SkyPilot should idle the service.

ASGI target and the remote start command (defaults app:app and `uvicorn app:app --host 0.0.0.0 --port 8000`).

Extra shell run on the VM after `pip install -r requirements.txt` (and a best-effort CPU PyTorch install).

Generated SkyPilot Serve spec. Do not treat it as the source of truth; it is rewritten on every deploy.

init [name]

Write a starter FastAPI `app.py` (health + echo `/predict`), `requirements.txt`, and `veloxml.yaml` in the current directory. name defaults to my-model-service. Existing `app.py` / `requirements.txt` are left untouched.

Run SkyPilot's cloud probe (`sky check`) and report whether AWS credentials and EC2 quota look usable. Fails with a hint to run `aws configure` when nothing is found.

Compile `.veloxml/<name>.sky.yaml`, then `sky serve up` (or `sky serve update` if the service already exists). Polls until the replica `/health` probe returns HTTP 200 (up to 10 minutes) and prints a sample `curl` against `/predict`. Requires `veloxml.yaml` or `app.py` in the current directory.

Parse `sky serve status` and print a table of VeloxML services (name, version, uptime, status, replicas, endpoint).

Print application logs via `sky serve logs`. name defaults to the `name` field in `veloxml.yaml`.

Stop the named service with `sky serve down`. With --all, also run `sky down -a` to destroy leftover controller nodes and clusters.

Print `VeloxML CLI version <x.y.z>`.

The repository marks itself alpha and warns of breaking changes. Version 0.1.1 is the package version at documentation time.Deployments run in your cloud account and cost money. Spot instances can be reclaimed. Always veloxml down --all when finished; leftover SkyPilot controllers keep billing after a single-service down.check looks for AWS in `sky check` output even though --cloud gcp is accepted. GCP still needs application-default credentials for deploy to work.deploy must be run from the project directory. Without `veloxml.yaml` or `app.py` it exits 1.The marketing site talks about scale-to-zero; the compiled replica policy defaults to min_replicas: 1. Setting `use_spot: true` is independent of replica count.logs --tail is ignored when --follow is set, because follow mode execs `sky serve logs` without capturing output.The `sky` binary must be on PATH (or next to the same Python that installed veloxml-deploy). A missing or unconfigured SkyPilot install makes every cloud subcommand fail.

VeloxML is an Apache-2.0 Python CLI by Pedro Águas Marques (paguasmar). The public GitHub repository paguasmar/veloxml-deploy was created on 29 August 2026. It wraps SkyPilot (UC Berkeley) for provisioning and borrows packaging conventions from Truss. The advertised goal is one-command LLM endpoints inside the caller's own VPC, without Dockerfiles, Kubernetes YAML, or a hosted inference vendor.

aws(1), gcloud(1), uvicorn(1), pip(1), curl(1), docker(1), flyctl(1)

Source code · Homepage · Documentation