Start a CUDA Python workload with Ubuntu or PyTorch

Verify GPU access and run Python compute inside a compatible GPU template.

Published on Oct 12, 2026 | Product: GPU

Before you start

Deploy an available certified Ubuntu/PyTorch-compatible template and connect with the rental’s SSH command.

Steps

Use nvidia-smi inside the container to confirm the assigned GPU and driver visibility.
Check the template’s Python and PyTorch installation before adding packages.
Create a project directory under the persistent workspace mount and keep datasets and outputs there.
Run a small CUDA tensor operation before starting a long training or inference job. Confirm that the device reports CUDA, not CPU.
Record package versions and your workload command. Watch container logs or your own output and save checkpoints frequently.

Example

nvidia-smi
python -c "import torch; print(torch.cuda.is_available()); print(torch.ones(4, device='cuda'))"

Check the result

The small test completes on CUDA and workspace files remain available after a supported stop/start.

Troubleshooting

A host GPU can be available even when a container’s CUDA libraries are incompatible. Check the image’s supported driver/runtime and avoid blindly upgrading all GPU packages.

Last updated October 12, 2026. Availability and prices are shown in the console before checkout.