Your first hour on a rented GPU: check it, load it, save it, stop it

The bill starts when the machine starts. Even on today's cheapest on-demand H100, at $1.80/hr, a machine left running over a weekend bills 48 hours that did no work: $86.40 for each GPU. So the first hour is worth doing in order, through to step 5. The steps below work on any Linux GPU machine you reach over SSH or a web terminal; the commands are standard tools, not one provider's.

1. Check that the card is the one you are paying for

Run nvidia-smi first. It names each GPU, its memory, the driver and the newest CUDA version that driver supports.

nvidia-smi --query-gpu=name,memory.total,driver_version --format=csv

Compare the name with the offer you chose: an H100 SXM reports itself as "NVIDIA H100 80GB HBM3", the PCIe card as "NVIDIA H100 PCIe". The two are priced differently, and our tables keep them apart for that reason. On a machine with several GPUs, nvidia-smi topo -m shows how they connect: NV followed by a number means NVLink, the fast link that multi-GPU training gains most from; PIX, PXB, PHB, NODE or SYS mean the cards talk over PCIe.

Then check that your framework sees the card:

python -c "import torch; print(torch.cuda.is_available(), torch.cuda.get_device_name(0))"

If it prints False, the usual cause is a PyTorch build for a newer CUDA version than the driver supports, or a CPU-only build: pick an image whose CUDA version is at or below the one nvidia-smi shows.

2. Put the weights on the disk that survives

Most machines have two kinds of disk: a system disk that disappears with the machine, and a volume that stays until you delete it. Run df -h to see them. Download model weights to the volume, so a restart does not mean a second download:

hf download Qwen/Qwen2.5-7B-Instruct --local-dir /workspace/models/qwen2.5-7b

Gated models need hf auth login first; older versions of the Hugging Face tools call the same commands huggingface-cli download and huggingface-cli login. Check the model's memory need before downloading a large one: how much VRAM a model needs.

3. Keep a long run alive when your connection drops

A closed laptop ends an SSH session and anything running in it. Start long jobs inside tmux:

tmux new -s train        # start a session, then run the job inside it
# detach with Ctrl-b then d; the job keeps running
tmux attach -t train     # come back to it later

Save checkpoints to the volume at a steady interval, not only at the end. On spot or community capacity, where the machine can be taken back, that interval is what you lose when it is: spot, on-demand or serverless.

4. Copy results out before you stop

Copy what you need to your own machine or storage before stopping anything:

rsync -avP user@MACHINE_IP:/workspace/outputs/ ./outputs/

Moving data out can be billed per GB, and large checkpoints add up: check the provider's data-out rate on the fees page and copy only what you will use.

5. Stop every charge, not only the GPU

Stopping a machine does not always stop its bill. Some providers keep charging the full rate until the machine is deleted; most still charge for its disk. When the results are safe, delete the machine and any volume you no longer need, then check the provider's billing page the next day. Our list of what each provider charges while stopped is below.