Software templates

Pick the tools. Then pick the GPU.

Q: What exactly is a template?

A container image plus the configuration needed to make it useful on a GPU node: the driver, the CUDA or ROCm runtime, an SSH server, and the tool itself. Images are pre-pulled on every node, so the instance is reachable in under 60 seconds instead of after a long download.

Q: Can I use my own image?

Yes. Give the console any OCI image reference from a public or private registry, with credentials if it is private. The only requirements are a Linux base with the GPU runtime your framework needs, and an entrypoint that stays in the foreground.

Q: Do templates cost anything?

No. Templates, updates and the images themselves are included in the hourly price of the instance. You only pay for the GPU-minutes and, if you attach one, the persistent disk.

Q: Are they kept up to date?

Images are refreshed as the upstream projects publish releases. The exact image reference is printed on every card below and in the console, so you always know which build you are about to run — and can pin an older one by using its reference directly.

Q: Which templates work on AMD GPUs?

The ROCm template, and any image you bring that ships ROCm builds of your framework. NVIDIA-specific images do not run on AMD Instinct instances, and the console will not offer them for those models.

Q: Where does my code live?

Wherever you put it: clone it into the instance at boot, bake it into your own image, or keep it on a persistent volume so it survives reclaims. Templates are environments, not storage.

9 environments that are already on the node when your instance starts — the framework, the CUDA or ROCm runtime, JupyterLab and SSH, configured and tested. Or bring any container image you like; nothing here is mandatory.

Deploy with a template Bring your own image

Templates: 9Training & fine-tuning · Development · Inference · Notebooks · Image generation
Boot time: < 60sImages are pre-pulled on every node
Extra cost: NoneIncluded in the GPU-hour price
Your own image: Any OCIPublic or private registry, with credentials

Training & fine-tuning

Train and fine-tune

PyTorch 2.5 · CUDA 12.4

Ubuntu 22.04, PyTorch 2.5, CUDA 12.4, cuDNN, JupyterLab, SSH.

Image: spotgpus/pytorch:2.5-cu124
Runs on: All NVIDIA GPUs in the catalogue

Deploy with this template

TensorFlow 2.17

Ubuntu 22.04, TensorFlow 2.17 GPU, CUDA 12.3, JupyterLab, SSH.

Image: spotgpus/tensorflow:2.17
Runs on: All NVIDIA GPUs in the catalogue

Deploy with this template

Inference

Serve a model

vLLM inference

vLLM with OpenAI-compatible server on port 8000, Hugging Face cache on the persistent disk.

Image: spotgpus/vllm:latest
Runs on: All NVIDIA GPUs in the catalogue

Deploy with this template

Ollama

Ollama server with the API on port 11434, models cached on the persistent disk.

Image: spotgpus/ollama:latest
Runs on: All NVIDIA GPUs in the catalogue

Deploy with this template

Notebooks

Work in a notebook

JupyterLab

JupyterLab with PyTorch and common data-science packages, token shown in the console.

Image: spotgpus/jupyter:latest
Runs on: All NVIDIA GPUs in the catalogue

Deploy with this template

Image generation

Generate images

ComfyUI

ComfyUI with the manager, models folder on the persistent disk.

Image: spotgpus/comfyui:latest
Runs on: All NVIDIA GPUs in the catalogue

Deploy with this template

Development

Start from bare metal

CUDA 12.4 base

Ubuntu 22.04, NVIDIA driver, CUDA 12.4 toolkit, Docker, SSH. Bring your own stack.

Image: spotgpus/cuda:12.4
Runs on: All NVIDIA GPUs in the catalogue

Deploy with this template

ROCm 6.2 (AMD)

Ubuntu 22.04, ROCm 6.2, PyTorch for ROCm, SSH. AMD Instinct instances only.

Image: spotgpus/rocm:6.2
Runs on: AMD Instinct instances

Deploy with this template

Ubuntu 22.04 (driver only)

Plain Ubuntu with the GPU driver and SSH. Nothing else.

Image: spotgpus/ubuntu:22.04
Runs on: Any instance in the catalogue

Deploy with this template

Your own image

A template is a convenience, not a lock.

Every instance can boot any OCI image you can pull. If your team already ships a training image, use it — the platform only adds the driver, the network and the disks.

Any registryDocker Hub, GHCR, Quay, NVIDIA NGC or your own — with credentials for private ones, stored encrypted in your account.
The GPU is already thereThe driver and the device are exposed to the container; you provide CUDA or ROCm at the version your framework needs.
Your volume is mountedThe persistent disk is mounted before the entrypoint runs, so checkpointing works with no changes.
First pull is slowerAn image we have never seen is downloaded once per node; after that it is cached like the built-in templates.

Dockerfilea minimal training image

FROM nvidia/cuda:12.4.1-cudnn-runtime-ubuntu22.04

RUN apt-get update && apt-get install -y --no-install-recommends \
      python3 python3-pip openssh-server && rm -rf /var/lib/apt/lists/*
RUN pip3 install --no-cache-dir torch transformers accelerate

# Checkpoints go to the persistent volume, datasets to local scratch.
ENV HF_HOME=/mnt/scratch/hf OUTPUT_DIR=/mnt/vol

COPY train.py /opt/train.py
CMD ["python3", "/opt/train.py"]

Keep the entrypoint in the foreground.

The instance is considered running while the entrypoint runs. A process that daemonises and exits will look like a finished job, and the instance will stop.

Pairing

Which GPU for which template.

A rough guide from the memory each tool tends to need. Every model page has the exact arithmetic for the model sizes you care about.

What you are doing	Template	A GPU that fits	From
Serving a 7B model with an OpenAI-compatible API	vLLM	RTX 4090 · 24 GB	$0.06/h spot
Fine-tuning 7B–13B with LoRA	PyTorch	A100 PCIe 80GB · 80 GB	$0.09/h spot
Serving a 70B model at FP8	vLLM	H200 SXM · 141 GB	$0.75/h spot
Full fine-tune of a 30B model	PyTorch	H100 SXM · 80 GB	$0.55/h spot
Generating images with diffusion models	ComfyUI	RTX 5090 · 32 GB	$0.09/h spot
Exploring data in a notebook	JupyterLab	L4 · 24 GB	$0.05/h spot
Running a model on AMD Instinct	ROCm 6.2 (AMD)	MI300X · 192 GB	$0.99/h spot

These are starting points, not limits: any template runs on any compatible GPU in the catalogue. Compare all 47 models.

FAQ

Templates and images

What exactly is a template?

A container image plus the configuration needed to make it useful on a GPU node: the driver, the CUDA or ROCm runtime, an SSH server, and the tool itself. Images are pre-pulled on every node, so the instance is reachable in under 60 seconds instead of after a long download.

Can I use my own image?

Yes. Give the console any OCI image reference from a public or private registry, with credentials if it is private. The only requirements are a Linux base with the GPU runtime your framework needs, and an entrypoint that stays in the foreground.

Do templates cost anything?

No. Templates, updates and the images themselves are included in the hourly price of the instance. You only pay for the GPU-minutes and, if you attach one, the persistent disk.

Are they kept up to date?

Images are refreshed as the upstream projects publish releases. The exact image reference is printed on every card below and in the console, so you always know which build you are about to run — and can pin an older one by using its reference directly.

Which templates work on AMD GPUs?

The ROCm template, and any image you bring that ships ROCm builds of your framework. NVIDIA-specific images do not run on AMD Instinct instances, and the console will not offer them for those models.

Where does my code live?

Wherever you put it: clone it into the instance at boot, bake it into your own image, or keep it on a persistent volume so it survives reclaims. Templates are environments, not storage.

Get started

Choose a template, choose a GPU, be running in under 60 seconds.

Pay as you go — no contracts, no minimum commitment. Add credit, launch, stop whenever you want.

Open the deployment screen See GPU prices

Billed per minute from the moment the instance is reachable. Minimum credit $40, no subscription.