--- title: Quickstart by task sidebar_label: Quickstart by task description: "The shortest path for each common job on Lium: give an AI agent a GPU, run a quick experiment, train, serve an LLM with vLLM, rent by spec under a price ceiling, multi-node training, control spend." --- > ## Documentation Index > Fetch the complete documentation index at: https://docs.lium.io/llms.txt > Use this file to discover all available pages before exploring further. > > Lium (lium.io) is a GPU rental marketplace: rent NVIDIA GPUs by the hour, billed per second. # Quickstart by task Pick the job you have; each path is a handful of commands, with a link to the full page. Every command works with the current CLI and SDK (`pip install lium.io`, 0.16.1). Authenticate once with `lium signup` (new account) or `lium init` (existing key), or set `LIUM_API_KEY`. Billing is per second with no minimum, so a short run costs only the seconds it uses. ## Give my AI agent a GPU An agent with a shell can sign up, rent, run and release a GPU on its own. - Claude Code: [Give Claude Code a GPU](/developers/claude-code-gpu) - Any agent: [AI Agents](/developers/agents) for the no-account-to-running-pod path, then [Scripting for agents](/developers/agent-scripting) for JSON output, exit codes and teardown - Give each agent its own key with a budget: [API keys → Budgets](/pod-users/api-keys#budgets) ## Run a quick experiment ```bash lium up --gpu RTX4090 --name exp --ttl 1h -y --json # rent, wait, remove itself after 1 hour lium exec exp "nvidia-smi" # exits with the remote command's exit code lium scp exp ./experiment.py # upload to ~/experiment.py lium exec exp "python experiment.py" lium rm exp -y # done early? release it now ``` More: [CLI quickstart](/developers/cli/quickstart) · [`lium up`](/developers/cli/reference/up) · [`lium exec`](/developers/cli/reference/exec) ## Train a model Keep data on a volume (mounted at `/mnt`, survives the pod), back up checkpoints on a schedule, and run the job detached so your shell is free. ```bash lium up --gpu H100 --name train --volume new:name=train-data --ttl 2d -y --json lium bk set train --path /root/checkpoints --every 1h --keep 7d -y lium exec train -d --log /root/train.log "python train.py --data /mnt --out /root/checkpoints" lium exec train "tail -n 50 /root/train.log" # check progress any time lium scp train /root/checkpoints/final.pt ./ -d # pull the result back ``` More: [Volumes](/pod-users/volumes) · [Backups](/pod-users/backups) · [SDK training workflow](/developers/sdk/examples/training-workflow) ## Serve an LLM with vLLM Tested on lium.io (15 Oct 2026, one RTX 4090): Qwen2.5-7B-Instruct served in about 4 minutes from `lium up`, at about 61 tokens/s for one stream. vLLM goes into its own `uv` venv because the default template pins its own torch. ```bash lium up --gpu RTX4090 --name vllm --ttl 2h -y --no-ssh lium exec vllm 'uv venv -q /root/vllm-env && uv pip install -q --python /root/vllm-env vllm' export VLLM_API_KEY=$(openssl rand -hex 16) lium exec vllm -d --log /root/vllm.log -e VLLM_API_KEY=$VLLM_API_KEY \ '/root/vllm-env/bin/vllm serve Qwen/Qwen2.5-7B-Instruct --host 127.0.0.1 --port 8000 --max-model-len 8192' lium exec vllm 'tail -n 5 /root/vllm.log' # wait for "Application startup complete" $(lium describe vllm --json | python3 -c 'import json,sys; print(json.load(sys.stdin)["access"]["ssh_cmd"])') \ -N -L 8000:127.0.0.1:8000 & # tunnel: laptop 127.0.0.1:8000 -> the pod curl http://127.0.0.1:8000/v1/models -H "Authorization: Bearer $VLLM_API_KEY" ``` More: [vLLM server (SDK)](/developers/sdk/examples/vllm) · [Templates](/pod-users/templates) ## Rent the cheapest GPU that fits a spec under a price ceiling One API call picks the cheapest node that meets every constraint and rents it; `dry_run: true` returns the pick and price without renting. ```bash curl -X POST https://lium.io/api/executors/rent-by-spec \ -H "X-API-Key: $LIUM_API_KEY" -H "Content-Type: application/json" \ -d '{"gpu_type": "H100", "gpu_count": 1, "min_vram_gb": 79, "max_price_per_gpu_hour": 2.5, "termination_hours": 4, "pod_name": "spec-pod", "dry_run": true, "user_public_key": "'"$(cat ~/.ssh/id_ed25519.pub)"'"}' ``` The same from Python: `Lium().rent(gpu_type="H100", max_price_per_gpu_hour=2.5, dry_run=True)`. From the shell, `lium ls --gpu H100 --max-price 2.50` shows what fits and `lium up --gpu H100` rents the cheapest match. More: [Rent by spec](/developers/quickstart#step-4-the-short-way--rent-by-spec) · [Error codes](/developers/error-codes) ## Multi-node training Rent whole nodes on one InfiniBand or RoCE fabric as a single all-or-nothing cluster, with launcher config printed for you. ```bash lium clusters # fabrics with free nodes lium clusters up 1 --nodes 2 --name job --ttl 12h -y lium clusters show job --hostfile # mpirun / DeepSpeed hostfile lium clusters show job --torchrun 0 # torchrun rendezvous flags for rank 0 lium clusters rm job -y # remove every member ``` More: [Multi-node clusters](/pod-users/clusters) · [`lium clusters`](/developers/cli/reference/clusters) ## Control spend ```bash lium up --gpu H100 --ttl 6h --budget 25 -y # removed after 6 h or at $25, whichever comes first lium spend # hourly burn, spend per pod, runway lium schedules list # every pod with a scheduled removal ``` Give each agent or job its own API key with server-enforced USD budgets (daily, monthly, lifetime); a rent past a budget is refused with `402 API_KEY_BUDGET_EXCEEDED`: ```bash curl -X POST https://lium.io/api/keys -H "Authorization: Bearer $JWT" -H "Content-Type: application/json" \ -d '{"name": "agent-1", "scopes": ["read", "rent"], "daily_budget_usd": 20, "monthly_budget_usd": 300}' ``` More: [API keys](/pod-users/api-keys) · [Billing](/pod-users/billing) · [Scheduled termination](/pod-users/scheduled-termination) ## Related - [What's live on lium.io](/developers/availability): every capability by surface - [Developer quickstart](/developers/quickstart): the REST API step by step