Your AI says you need a GPU? Run your Python script on a GPU in minutes

Why your AI says you need a GPU

Your assistant (Claude, Claude Code, ChatGPT, Codex, Cursor or Copilot) usually sees one of these signs: your code calls .to("cuda") but torch.cuda.is_available() returns False, training ends with CUDA out of memory, or it takes hours on the CPU. In all three cases you need an NVIDIA card with enough memory (VRAM).

• You are on a Mac or a PC without an NVIDIA card: read torch.cuda.is_available() returns False.
• Your card runs out of memory: read CUDA out of memory.
• Your code runs but too slowly: a GPU rented by the hour runs the same script without changing your code.

Which GPU to rent, and what it costs

For a PyTorch script, start with the RTX A6000 with 48 GB. If your model does not fit in memory, move to a card with 80 GB or 96 GB. The pricing page also lists the servers with 2 to 8 GPUs.

The steps, with the code to copy

1. 1.Check what your code sees

   On your computer, check whether PyTorch finds a GPU. False confirms there is no usable CUDA GPU.

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

2. 2.Launch a server with the PyTorch (CUDA) template

   In the console, pick the PyTorch (CUDA) template, an RTX A6000 and your SSH key, then confirm the price. PyTorch and CUDA install into /opt/pytorch in a few minutes; the console shows the progress and the IP address.

3. 3.Copy your project to the server

   Send your folder with rsync or scp. The SSH user is ubuntu.

   # Copy the project without the local virtual environment
   rsync -avz --exclude .venv --exclude __pycache__ ./my-project/ ubuntu@IP_ADDRESS:~/my-project/
   
   # Without rsync (on Windows for example), scp works too
   scp -r ./my-project ubuntu@IP_ADDRESS:~/

4. 4.Log in and install your dependencies

   The PyTorch environment is activated at login: install only what your project adds.

   ssh ubuntu@IP_ADDRESS
   # The /opt/pytorch environment is activated at login
   cd ~/my-project
   pip install -r requirements.txt

5. 5.Check the GPU

   nvidia-smi shows the card, and PyTorch should now answer True.

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

6. 6.Run your script

   With nohup, the job keeps running even if you close your terminal.

   # The script keeps running even if your SSH connection drops
   nohup python train.py > train.log 2>&1 &
   tail -f train.log

7. 7.Fetch your results

   Copy your output files (models, logs, figures) back before you stop the server.

   rsync -avz ubuntu@IP_ADDRESS:~/my-project/outputs/ ./outputs/

8. 8.Stop the charges

   An on-demand server can be hibernated ($0.02 per hour (about $14.60 per month)): the disk is kept and you resume later. A Spot machine cannot be hibernated: delete it. Stopping the server is not enough, it is still billed at the full rate. Do it in the console or with the API.

   export GPUCLOUD_API_KEY="<your API key>"
   
   # On-demand server: hibernate it (the disk is kept)
   curl -X POST "https://gpucloud.ca/api/servers/SERVER_ID/actions" \
     -H "Authorization: Bearer $GPUCLOUD_API_KEY" \
     -H "Origin: https://gpucloud.ca" \
     -H "Content-Type: application/json" \
     -d '{"action": "hibernate"}'
   
   # Spot machine, or work finished: delete the server
   curl -X DELETE "https://gpucloud.ca/api/servers/SERVER_ID" \
     -H "Authorization: Bearer $GPUCLOUD_API_KEY" \
     -H "Origin: https://gpucloud.ca"

Keep your code portable between your computer and the GPU

Pick the device at startup instead of hard-coding cuda: the same script runs on your computer (CPU) and on the server (GPU).

import torch

device = "cuda" if torch.cuda.is_available() else "cpu"
print("device:", device)

model = model.to(device)
for inputs, targets in loader:
    inputs, targets = inputs.to(device), targets.to(device)
    ...

Your assistant can rent the GPU for you

If your assistant speaks MCP (Claude Code, Cursor, VS Code or Claude Desktop), connect it to the GPU Cloud MCP server, https://gpucloud.ca/api/mcp, with a spending-capped MCP key. It reads the catalog, asks for a quote, shows you the price in Canadian dollars and creates the server only after you confirm. The configuration to paste is on the Developers page.

Good to know before you launch

• Billing is hourly, at the server’s hourly rate, from prepaid credit. There is no per-minute billing, and the first hour is charged when you deploy.
• A stopped server is still billed at the full hourly rate. To pay less, hibernate it ($0.02 per hour (about $14.60 per month), disk kept) or delete it. A Spot machine cannot be hibernated: delete it when you are done.
• The minimum top-up is CA$25.00, paid by credit card.
• Servers and their disks stay in Canada (region canada-montreal, in Montreal). Account information (identity, billing, e-mails) may be processed outside Quebec, as the privacy policy explains.

Frequently asked questions