Rent a cloud GPU for vLLM, billed by the minute

vLLM is an open-source inference and serving engine for large language models. It runs an OpenAI-compatible server, so code that talks to the OpenAI API can talk to your own model instead. On CloudGPU you can run it on a rented GPU billed by the minute: start a raw GPU, install vLLM over SSH, serve your model, and destroy the instance when you are done.

There is no one-click vLLM template today. You use Raw GPU: a dedicated container with JupyterLab and CUDA pre-installed and root SSH, where you install what you need.

Pick a card on the pricing page · Deploy a raw GPU

Which card for which model

vLLM needs a Linux machine and an NVIDIA GPU with compute capability 7.5 or higher. These cards on our pricing page meet it:

Card VRAM Compute capability
RTX 5090 32G 32 GB 12.0
RTX 4090 48G 48 GB 8.9 (RTX 4090 chip)
A100 80G 80 GB 8.0
RTX 4090 24G 24 GB 8.9
RTX 3090 24G 24 GB 8.6

Two things decide the card:

  • Weights plus KV cache. The model weights must fit, and vLLM uses the rest of the memory it is given for the KV cache, which grows with context length and the number of concurrent requests. By default vLLM uses 0.92 of the GPU’s memory (--gpu-memory-utilization). If the model does not fit at its full context, cap it with --max-model-len. See how much VRAM an LLM needs.
  • One GPU per instance. Instances on CloudGPU are single-GPU, so pick a card where the model fits on one GPU; vLLM’s --tensor-parallel-size stays at its default of 1.

Availability changes during the day.

Start vLLM in five steps

  1. Deploy a raw GPU. On Deploy, choose Raw GPU and a card. About 2 minutes later the instance row shows root SSH and JupyterLab.
  2. Install vLLM. In a JupyterLab terminal or over SSH, the vLLM quickstart recommends uv:
uv venv --python 3.12 --seed
source .venv/bin/activate
uv pip install vllm --torch-backend=auto
  1. Serve a model. vllm serve starts an OpenAI-compatible server on http://localhost:8000 by default. Pass --api-key (or set VLLM_API_KEY) so the server checks for a key:
vllm serve Qwen/Qwen2.5-1.5B-Instruct --api-key "$VLLM_API_KEY"
  1. Reach it from your machine. The public HTTPS URL is provided for the API templates (Ollama and ComfyUI); for vLLM on a raw GPU, forward the port over SSH. Take the SSH command from the instance row and add -L 8000:localhost:8000, then call it locally:
curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $VLLM_API_KEY" \
  -d '{"model": "Qwen/Qwen2.5-1.5B-Instruct", "messages": [{"role": "user", "content": "Hello"}]}'

Any OpenAI client works too: set its base URL to http://localhost:8000/v1.

  1. Destroy when done. Persistent storage across restarts is still on the roadmap, so model downloads happen again on a new instance. Copy off anything you want to keep, then destroy: billing stops and unused minutes go back to your balance.

Billing in plain words

  • Billed per minute of GPU time, with a 1-minute minimum. Model downloads and idle time count as GPU time, so destroy the instance when you stop serving.
  • While an instance runs, each started hour is pre-charged from your balance; unused minutes go back to your balance when you destroy the instance.
  • Pay with Visa or Mastercard, Alipay (in most countries) or Google Pay / Apple Pay in one checkout, USDT on TRC-20, or bank transfer for business customers (payments).

vLLM or Ollama?

If you want a public HTTPS endpoint without any setup, the Ollama template gives you an OpenAI-compatible endpoint after deploy. Choose vLLM when you need its serving features, Hugging Face model formats, or control over memory and context settings. If you would rather not run a GPU at all, see self-hosting vs an API.

FAQ

Is there a vLLM template? Not today. Use Raw GPU, which comes with JupyterLab, CUDA and root SSH, and install vLLM yourself in a few commands.

Which GPUs can run vLLM? vLLM’s documentation requires an NVIDIA GPU with compute capability 7.5 or higher. The RTX 5090, RTX 4090 48G, A100 80G, RTX 4090 and RTX 3090 on our pricing page meet that.

Does vLLM give me an OpenAI-compatible API? Yes. vllm serve starts an OpenAI-compatible server, by default on localhost port 8000, with endpoints such as /v1/chat/completions.

How do I call it from my laptop? Forward port 8000 over SSH using the SSH command from the instance row, then point your OpenAI client at http://localhost:8000/v1. Start the server with --api-key so requests need a key.

How is it billed? Per minute of GPU time, with a 1-minute minimum. Each started hour is pre-charged, and unused minutes go back to your balance when you destroy the instance. Rates are on the pricing page.

Ready to start?

Deploy a raw GPU · See the current rates

Sources

All read on 24 October 2026 (Beijing time, UTC+8).