- Python 99.9%
- Makefile 0.1%
| Filename | Latest commit message | Latest commit date |
|---|---|---|
453c5bfd6b
Production registers the public device-flow client with the fixed id airigs-cli, so zero-config login needs no override. The old comment claimed AIRigs generates a deployment-specific aic_... id, which contradicts the verified flow; AIRIGS_CLIENT_ID is documented as an escape hatch for non-standard deployments instead. |
||
| airigs | ||
| docs | ||
| tests | ||
| .gitignore | ||
| AGENTS.md | ||
| LICENSE | ||
| Makefile | ||
| pyproject.toml | ||
| README.md | ||
airigs
The AIRigs CLI: sign in, benchmark a machine, submit a rig, and publish reproducible recipes. It talks to AIRigs (identity, rigs, benchmarks) and the Recipes platform (recipes, benchmarks, scripts), and uploads only when you submit or publish.
Install (Python 3.11+, Linux/WSL2 or macOS)
git clone https://git.fedesito.me/fedes1to/airigs-cli.git
cd airigs-cli && pipx install .
pipx install '.[gpu]' adds PyTorch for TFLOPS, GPU memory bandwidth and P2P;
without it those fields are null and a rig upload is refused with the exact
missing measurements.
Log in
Run airigs login. The OAuth device flow (RFC 8628) prints a URL and a short
code, opens your browser, and stores the token once approved. On a headless host
use --no-browser, or pipe a token on stdin (argv leaks through ps):
printf '%s' "$AIRIGS_TOKEN" | airigs login --token -
An install from this repository talks to the public deployment out of the box:
https://airigs.net (AIRigs) and https://recipes.airigs.net (Recipes). No
environment variables are needed. Only for a local instance set both URLs (env
or --airigs-url/--recipes-url) and log in again — tokens are bound to their
issuer origin, so a new host always needs a fresh login:
export AIRIGS_URL=http://localhost:4567
export RECIPES_URL=http://localhost:4568
airigs login
Benchmark the machine
airigs bench --output bench.json
airigs bench --quick # fewer iterations for a smoke test
CPU, RAM, GPUs, PCIe, interconnect and power are measured; a missing sensor or
optional library yields null plus a note, never a made-up number. --json
prints the Benchmark v1 payload.
Submit a rig
airigs rig submit --name "workstation"
This requests a single-use challenge, measures, uploads, and fails unless the
server returns verified: true. Reuse a saved run with --benchmark bench.json.
Also: rig list, rig show <id|name>, rig delete <id> --yes.
Work with recipes
airigs recipe init --name "70B on 2x4090" # then edit recipe.json
airigs recipe validate
airigs recipe submit --publish
Benchmark an already-running OpenAI-compatible endpoint, then publish:
airigs recipe bench my-recipe \
--endpoint http://localhost:8000/v1 \
--engine vllm --model meta-llama/Llama-3.1-8B \
--context 8192 --concurrency 4 --requests 16 --publish
The launch script is always shown first; run needs --yes (or
AIRIGS_YES=1) when stdin is not a TTY.
airigs recipe script my-recipe --out run.sh
airigs recipe run my-recipe --yes -- --port 8000
Troubleshooting
airigs doctor is a read-only checklist: Python, libraries, NVIDIA/AMD tools,
config permissions, token state, and service reachability. airigs --self-check
reports whether this host can produce an uploadable rig.
Docs
docs/installing.md— install, login, config, env vars, logout.docs/benchmark-v1.md— the frozen Benchmark v1 contract.- Recipes deployments publish this walkthrough at
/docs/cli. AGENTS.md— layout and conventions for coding agents.
Every command accepts --json; exit codes: 0 ok, 1 general, 3 auth,
4 not found/conflict, 5 validation, 6 rate limit.