Choose GPU memory disk space and a Docker template

Before you start

Know the model or application you plan to run. GPU memory is different from system RAM and disk storage.

Steps

1. Read the offering’s GPU type, per-GPU VRAM, GPU count, and provider location.
2. Read the compatible template’s minimum memory and recommended defaults. Multiple GPUs do not automatically combine VRAM for an application.
3. Allow disk capacity for the image, model files, caches, inputs, outputs, and temporary downloads. A 9 GB model needs more than 9 GB of free disk for a usable workflow.
4. For inference, begin with a smaller or quantized model and conservative context length. For image generation, start with modest resolution and batch size.
5. Review storage rates and persistent mount settings before ordering, then verify CUDA and the application after deployment.

Check the result

The model fits without out-of-memory errors, downloads have space to finish, and outputs are stored persistently.

Troubleshooting

Reduce batch size, context length, image resolution, or model size for CUDA out-of-memory errors. Template availability is specific to the GPU/provider combination; larger training workflows may need 24 GB or more.