Fine-Tune With LLaMA-Factory on a Rented GPU, Billed by the Minute (RTX 4090 / 4090 48G / A100, 2026)

LLaMA-Factory is the quickest way to fine-tune an open model without writing a training loop: a YAML file or a web UI, LoRA and QLoRA out of the box, and support for most of the Llama, Qwen, DeepSeek and Mistral families. What it does not come with is a GPU. This post covers renting one by the minute on cloudgpu.app, which card fits which job, and what a run costs.

What each card fits, and what it costs (27 September 2026)

Rough VRAM needs for a LoRA fine-tune; real usage moves with sequence length, batch size and gradient checkpointing, so leave headroom. Prices are per hour, billed per minute.

Job	Rough VRAM	Card	Price
QLoRA (4-bit), 7B to 8B	6 to 10 GB	RTX 3090 24G / RTX 4090 24G	$0.21/h / $0.33/h
LoRA (16-bit), 7B to 8B	18 to 22 GB	RTX 4090 48G (24 GB is tight)	$0.65/h
QLoRA, 14B	12 to 16 GB	RTX 4090 24G	$0.33/h
QLoRA, 32B	24 to 30 GB	RTX 4090 48G	$0.65/h
QLoRA, 70B	45 to 55 GB	A800 80G / A100 80G	$1.09/h / $1.49/h

For the VRAM side in more detail, see how much VRAM each open model needs. Stock moves during the day; the pricing page shows what is available right now.

What a run costs. Three hours of QLoRA on an 8B model on an RTX 4090 is about $0.99. The same three hours on an RTX 4090 48G is about $1.95, and a 70B QLoRA for three hours on an A800 80G is about $3.27. One hour is pre-authorised when the machine starts; when you stop it, unused minutes go back to your balance.

Step by step

1. Sign up and top up. Email or Google account, no phone number. Pay by Alipay, Google Pay, Apple Pay, Visa or Mastercard from $5, or USDT (TRC-20). A first top-up of $10 or more gets a $5 bonus.
2. Rent the card. On the deploy page, pick the card from the table above and the PyTorch template. If you need more than one GPU, the card-count picker only shows counts that can be placed right now.
3. Open a terminal. The machine comes up in a couple of minutes with JupyterLab and SSH; the instance page shows both. Use JupyterLab’s terminal or SSH in.
4. Install LLaMA-Factory.

git clone --depth 1 https://github.com/hiyouga/LLaMA-Factory.git
cd LLaMA-Factory
pip install -e ".[torch,metrics]"

5. Point model downloads somewhere fast. Most of our supplier data centres are in mainland China, where Hugging Face is slow or unreachable. Either use the Hugging Face mirror or pull from ModelScope:

export HF_ENDPOINT=https://hf-mirror.com   # Hugging Face mirror
# or
export USE_MODELSCOPE_HUB=1                # download from ModelScope instead

6. Train. From a config file:

llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml

Or with the web UI (LLaMA Board), which listens on port 7860 on the machine. Start it, then forward the port from your laptop over SSH and open http://localhost:7860:

# on the GPU machine
llamafactory-cli webui

# on your laptop (user, host and port are on the instance page)
ssh -L 7860:localhost:7860 -p <ssh-port> <ssh-user>@<ssh-host>

7. Save your adapter before you stop the machine. Download the output directory from JupyterLab, scp it off, or push it to Hugging Face or ModelScope. Destroying the instance deletes its disk, and an instance is also destroyed when your balance runs out, so keep a few hours of balance on hand for a long run and save checkpoints as you go.

Things worth knowing

• Latency does not matter much for training, but downloads do. Base models of 15 to 140 GB are the slow part of the first run; the mirror settings above are the fix.
• Multi-GPU is not always on offer. A multi-GPU instance needs that many free cards on one machine. We only show counts we can place, so a count you saw in the morning may be gone by the evening, and the other way round.
• Where the machines are. Mainly mainland China, some Singapore and the US. If your dataset must not leave a particular jurisdiction, this is not the right place to train it.
• After training. Serving the fine-tuned model is a separate job: an Ollama or vLLM template on the same kind of card gives you a public HTTPS endpoint. See running Ollama with a public API.

Start at the deploy page or check current prices.