Use case · Fine-tuning
Train on your data. Keep the weights.
LoRA and QLoRA fine-tuning of open-weight LLMs with Hugging Face TRL on a datacenter GPU. Train, copy the adapter, end the session.
$ curl -LsSf https://astral.sh/uv/install.sh | sh && source $HOME/.local/bin/env$ uv venv --python 3.12 --seed && source .venv/bin/activate$ uv pip install torch --torch-backend=auto && uv pip install "trl[peft]"$ trl sft --model_name_or_path Qwen/Qwen2.5-0.5B \ --dataset_name trl-lib/Capybara --use_peft --output_dir qwen-loraAdapt an open-weight model to your data with LoRA. The adapter is small: train for a few hours, copy it out, end the session.
For whom
- Teams specializing a base model on their domain, format or tone.
- Researchers running short training experiments.
- Anyone who needs a large GPU for hours, not months.
Which GPU
| Base model | LoRA (16-bit base) | QLoRA (4-bit base) |
|---|---|---|
| Up to 3B | 24 GB | 16 GB |
| 7B to 8B | 48 GB | 24 GB |
| 13B to 14B | 80 GB | 48 GB |
| 30B to 34B | Several GPUs | 80 GB |
Longer sequences and larger batches need more memory. Full fine-tuning needs several times the LoRA figures.
Steps
- Book a machine. Choose a duration that covers training plus a margin.
- Upload your dataset with
rsync, or load it from the Hugging Face Hub. - Follow the LoRA recipe. Run training inside
tmux. - Copy the adapter to your computer or your storage.
- End the session.
Tips
- Pick a non-interruptible offer for runs you cannot resume. On interruptible offers, save checkpoints often (
--save_steps). - Run a few steps first to check memory and throughput, then launch the full run.
- The session ends when the booked time is up. Budget the run with a margin.
- Push the adapter to a private Hub repository from the machine to skip the download step.
Request access.
Invitations go out by email.