Skip to main content

Copying files

Move data to and from your machine with scp and rsync, and download models directly on the machine.

On this page

Everything on the machine is wiped when your reservation ends, so plan how you get data on and — more importantly — off. The examples use the short name gtl from Using ~/.ssh/config; with no config, use the full user@…ssh.gputolease.ai address from your dashboard instead.

Download on the machine, not through your laptop

Models and datasets usually come from the internet — Hugging Face, GitHub, a cloud bucket. Download them on the machine, not to your computer first: it's faster and doesn't use your own connection. For example, with the Hugging Face CLI in a Python virtual environment:

bash
ssh gtl
python3 -m venv ~/.venvs/hf
~/.venvs/hf/bin/pip install --upgrade "huggingface_hub[cli]"
~/.venvs/hf/bin/hf download mlx-community/Qwen3-8B-4bit

A startup script can start the download before you even log in.

scp: copy a few files

Copy a file to the machine:

bash
scp ./prompts.jsonl gtl:~/

Copy a file back:

bash
scp gtl:~/results/summary.csv ./

Add -r to copy a folder:

bash
scp -r gtl:~/results ./results

rsync: large or repeated transfers

rsync copies only what changed, can resume, and shows progress — the better choice for anything large:

bash
rsync -avh --progress ./dataset/ gtl:~/dataset/
rsync -avh --progress gtl:~/checkpoints/ ./checkpoints/

The trailing / on the source means "the contents of this folder". Run the same command again to pick up where an interrupted copy stopped.

On Windows, rsync isn't built in: use scp, or run rsync inside WSL.

Before your reservation ends

  • The Ending soon email arrives an hour before the end (you can turn it off in Settings, but we recommend keeping it).
  • Start copying results well before 23:59 UTC on your last day. A large copy that's still running at the end is cut off.
  • Don't rely on the machine for storage between reservations: every reservation starts from a clean machine, and there are no backups.
  • For big outputs, push them from the machine to your own storage (for example an S3 bucket or the Hugging Face Hub) instead of pulling them through your connection.