USE CASE

Fine-tune an LLM with LoRA on a cloud GPU

Fine-tune Llama, Qwen or Mistral with LoRA or QLoRA on an A100 or RTX 5090 and pay only for the training time.

Billed per second from your USD balance. Stop or terminate the machine when you are done.

Recommended GPU

A100 SXM4 · RTX 5090

QLoRA on a 7B–14B model fits in 24–32 GB. Full LoRA on larger models needs an A100 80 GB.

PyTorch

NVIDIA CUDA + PyTorch development environment, ready with SSH and Jupyter.

1
Rent the machine
Open the catalog with A100 and the PyTorch template already selected and press Rent.
See A100 SXM4 machines
2
Connect and upload your data
Connect with the SSH command from your machine's page and copy your dataset with scp.
```
$ scp -P <PORT> dataset.jsonl root@<HOST>:/root/
```
3
Install the libraries
Install the Hugging Face training stack, then run your training script.
```
$ pip install -U transformers peft trl datasets accelerate bitsandbytes
python train.py
```
4
Download the adapter
Copy the trained adapter back to your computer, then terminate the machine.
```
$ scp -P <PORT> -r root@<HOST>:/root/output ./adapter
```

How long does a fine-tune take?

A QLoRA run on a 7B model with a few thousand examples usually takes under an hour on an A100.

What if my balance runs low during training?

You get a warning first. If the balance runs out the machine is stopped and the disk is kept, so save checkpoints often.