11 November 2026 · VRAMfarm team
What happens to your files when you stop a GPU server, and how to keep them
You stop the server, start a new one the next day, and the LoRA you trained is gone. That happens because a GPU server’s disk is deleted the moment you stop it. The one exception on VRAMfarm is the /workspace folder: click Save and stop and it is kept for €0.03 per GB per month and put back on your next server, whichever card you choose. Everything else, on any path, is gone.
What is deleted when you stop
A server here is a container on a machine that belongs to a vetted host. When you stop, the machine goes back to the host and its disk is wiped. There is no “stopped but kept” state, so there is nothing left to restart and nothing to pay for.
What survives depends on which button you press and where the file was:
| What | Stop | Save and stop |
|---|---|---|
Your files in /workspace (outputs, LoRAs, datasets, notebooks, your own models) |
Deleted | Kept |
ComfyUI output folder, workflows and custom node folders (inside /workspace/ComfyUI) |
Deleted | Kept |
| Base models of the ready-made apps (Flux, Wan, Qwen-Image weights) | Deleted | Not saved, the app downloads them again |
Python packages installed with pip, system packages from apt |
Deleted | Deleted |
Hugging Face cache in ~/.cache |
Deleted | Deleted, unless you move it into /workspace |
Anything in /root, /tmp or other paths |
Deleted | Deleted |
| Servers started from your own Docker image | Deleted | Saving not available yet |
Two lines in that table cause most of the surprises.
Base models are skipped on purpose. The 17 GB of Flux or the 70 GB of Wan 2.2 14B are public files. The app downloads them itself on every launch, so storing them would only add a monthly charge for files you can always get again. Your own checkpoints and LoRAs, which nobody else has, are saved.
Packages live outside /workspace. Custom nodes come back as folders, but the Python libraries they installed do not. Keep a short setup script in /workspace, for example pip install -r /workspace/requirements.txt, and run it after each start. For Hugging Face downloads, set HF_HOME=/workspace/hf before you download so the cache lands inside the saved folder.
How the saved folder works
When you click Save and stop, the server uploads /workspace to storage that is not tied to the machine, checks that the upload is complete, and only then stops. The minutes it spends uploading are billed as part of the session, usually a few. If the upload fails, the server stays running so you don’t lose anything, and you get a notification; it is still billed until you stop it.
On your next launch the dashboard offers to restore the folder, with the size shown next to the checkbox, which is ticked by default. It comes back before the app starts, so ComfyUI sees your LoRAs and workflows when it opens. The new server can be a different card: save on an RTX 3090, continue on an RTX 4090 or an H100. That also means a busy machine can’t hold your files hostage. Your folder is not stuck on a host whose GPU someone else has since rented.
Storage costs €0.03 per GB per month and is charged daily, a thirtieth of the monthly amount each day. You can delete the folder in the dashboard at any time, and the charge stops.
Keep models or download them again?
This is the question with real money in it. Moving bytes onto a new machine is traffic, billed at that machine’s rate, which is shown before launch. Restoring a saved folder is traffic too. So storing a file does not save you the traffic. It saves you time, and it keeps files that you can’t download again.
Take 70 GB of models you downloaded yourself from Hugging Face, for example for a custom video workflow in the empty ComfyUI.
Option A: keep them in the saved folder. Storage is 70 × €0.03 per month, whether you launch once or twenty times. Each launch also pays the traffic to restore 70 GB, and you wait while it copies.
Option B: download them again each time. No storage charge. Each launch pays the traffic for 70 GB from Hugging Face, plus the GPU minutes you wait. On a fast machine, 70 GB takes about 10–15 minutes. On a 48 GB L40S that wait costs about €0.27 per launch.
The traffic is roughly a wash: at a typical rate of around a cent per GB or less, 70 GB costs well under a euro either way. What differs is the fixed monthly storage against the minutes spent re-downloading. Restoring from storage is not always faster than Hugging Face, so for public weights Option B usually wins, especially if you launch a few times a month.
| What it is | Save it? |
|---|---|
| LoRAs you trained, fine-tuned checkpoints | Yes, they exist nowhere else |
| Outputs, renders, generated datasets | Yes, or download them to your computer |
| Edited workflows, configs, scripts, notebooks | Yes, they are tiny |
| Models from links that may disappear (some Civitai files, gated repos) | Yes |
Public weights you can fetch with one wget or hf download command |
Usually no |
| Base models of the ready-made apps | Not needed, skipped automatically |
When your balance hits zero
Servers run from your prepaid balance. As it gets low, we warn you in the dashboard and by email, including when less than an hour of runtime is left. When the balance runs out, the server stops.
A stop because of balance is an ordinary stop: there is no time to upload the folder, so the server disk is deleted. A folder you saved earlier is not affected. It stays in storage and keeps being charged daily. If the balance stays at zero, we notify you and keep the folder for 30 days, then delete it. Top up and it is there for your next server.
Long job, unsure balance? Top up before you start, and keep anything important saved or copied as you go.
When a machine fails mid-job
Hosts are vetted, but hardware still fails now and then. If a machine goes offline, billing stops from the moment the machine went missing, not from the moment someone noticed, and you get a notification.
The machine’s disk is lost with it. Your saved folder is safe, but it only holds what was there at the last Save and stop, because saving happens on stop. For anything longer than an hour or two:
- Write checkpoints regularly, every 15–30 minutes of training or every few hundred steps.
- Copy them off the machine as well: push to a private Hugging Face repository,
rsyncorscpto your own computer (see the SSH key guide), or upload to your own S3 bucket. - For a run of several hours, stop and save at a natural break. Resuming from a checkpoint on a fresh server costs a couple of minutes.
Idle auto-stop
Forgotten servers are the most expensive kind. If a server has no SSH or Jupyter connection and no GPU load for about two hours, we send you an email. If it is still idle an hour later, the server saves its folder and stops itself, so your files are kept and the billing ends.
A training job counts as GPU load, so a run started with nohup or tmux is not stopped while it computes. If your job has long quiet phases, for example waiting for data, you can switch auto-stop off in the settings. Servers started from your own Docker image are not auto-stopped, because their folder can’t be saved yet. Stop those yourself or through the API.
Before your next stop, a quick check: is everything you need inside /workspace, do you want to keep it, and is anything big in there that you could simply download again? If so, delete it and pay storage only for the files that are truly yours. The billing guide shows how storage and traffic add up next to the GPU time.