Templates · 37 maintained stacks
37 GPU templates: from "deploy" to working stack in 30 seconds
Every template is a pinned, GPU-ready image with ports mapped and storage mounted — official builds of the tools you already use. Launch PyTorch, serve with vLLM, generate with ComfyUI, fine-tune with Axolotl — or point us at your own Docker image.
Base & frameworks 6
-
The bare CUDA base image — drivers, toolkit, SSH and Jupyter. Build your own stack on top.
powergpu/base:cuda13
GPUs & details Deploy -
The training and research default: PyTorch 2.6, cuDNN, JupyterLab and SSH out of the box.
powergpu/pytorch:2.6-cuda12.8
GPUs & details Deploy -
TensorFlow with GPU support, Keras and TensorBoard on a mapped port.
powergpu/tensorflow:cuda12.1
GPUs & details Deploy -
NVIDIA's NGC-optimised PyTorch build — tuned kernels for the datacenter cards.
nvcr.io/nvidia/pytorch
GPUs & details Deploy -
GPU-accelerated data science — cuDF, cuML, cuGraph in a notebook.
rapidsai/notebooks
GPUs & details Deploy -
A launcher bundling the most-used AI apps behind one desktop — pick and run.
powergpu/aio-studio
GPUs & details Deploy
LLM serving & chat 9
-
Production LLM serving with PagedAttention — an OpenAI-compatible endpoint from any HF model.
powergpu/vllm
GPUs & details Deploy -
vLLM extended for multimodal models — vision and audio inputs on the same fast server.
powergpu/vllm-omni
GPUs & details Deploy -
High-throughput serving with RadixAttention — excels at structured and agentic workloads.
powergpu/sglang
GPUs & details Deploy -
GGUF inference with a built-in server — the lightest way to run quantized models.
powergpu/llama-cpp
GPUs & details Deploy -
Pull and run quantized models in one command, with a clean REST API.
powergpu/ollama
GPUs & details Deploy -
A full chat UI over Ollama — conversations, RAG and model management in the browser.
powergpu/openwebui
GPUs & details Deploy -
The text-generation WebUI & API — load GPTQ, EXL2 or GGUF with extensions.
powergpu/oobabooga
GPUs & details Deploy -
Hugging Face Text Generation Inference — battle-tested production serving with an API.
ghcr.io/huggingface/text-generation-inference
GPUs & details Deploy -
Visually build LLM pipelines and agents on top of a local Ollama backend.
powergpu/langflow
GPUs & details Deploy
Training & fine-tuning 5
-
Fine-tune Llama, Qwen or Mistral from one YAML — LoRA, QLoRA and multi-GPU FSDP.
axolotlai/axolotl-cloud
GPUs & details Deploy -
2× faster, lower-VRAM fine-tuning — Unsloth kernels with a notebook workflow.
powergpu/unsloth-studio
GPUs & details Deploy -
Train SDXL and Flux image LoRAs in a browser GUI — dataset in, .safetensors out.
powergpu/kohyas-gui
GPUs & details Deploy -
A dead-simple UI for training Flux LoRAs, built on the Kohya scripts.
powergpu/fluxgym
GPUs & details Deploy -
The AI-Toolkit trainer for Flux and diffusion models, with a web UI.
powergpu/ostris-ai-toolkit
GPUs & details Deploy
Image generation 6
-
Node-based image and video workflows with ComfyUI-Manager pre-installed.
powergpu/comfyui
GPUs & details Deploy -
A polished Stable Diffusion studio — canvas, layers and workflow nodes.
powergpu/invokeai
GPUs & details Deploy -
The Forge fork of the SD WebUI — faster attention, lower VRAM, same extensions.
powergpu/sd-forge
GPUs & details Deploy -
The classic AUTOMATIC1111 Stable Diffusion WebUI with the full extension ecosystem.
powergpu/a1111
GPUs & details Deploy -
Prompt-and-go image generation — Midjourney-style simplicity on SDXL.
powergpu/fooocus
GPUs & details Deploy -
A ComfyUI-backed UI that scales generation across multiple GPUs.
powergpu/swarmui
GPUs & details Deploy
Video generation 2
-
Run Wan and other video-generation models on modest VRAM, with a simple UI.
powergpu/wan2gp
GPUs & details Deploy -
The open text-to-video model, ready to generate and fine-tune.
powergpu/open-sora
GPUs & details Deploy
Audio & speech 3
-
Batch-transcribe audio with faster-whisper behind a UI and a REST endpoint.
powergpu/whisper
GPUs & details Deploy -
Text-to-speech and voice cloning with a web interface and API.
powergpu/voicebox
GPUs & details Deploy -
The ACE-Step music generation model with a UI and API.
powergpu/acestep
GPUs & details Deploy
Desktops & VMs 4
-
A full GPU Linux desktop over the browser — XFCE, VNC/RDP, run GUI apps.
powergpu/linux-desktop
GPUs & details Deploy -
The Pinokio 1-click app browser on a desktop — install AI apps with a click.
powergpu/pinokio
GPUs & details Deploy -
A full KVM virtual machine — your own kernel, root, systemd, any stack.
powergpu/kvm:ubuntu-22.04
GPUs & details Deploy -
The Ubuntu VM with a GNOME desktop over VNC/RDP — a full graphical workstation.
powergpu/kvm:ubuntu-desktop
GPUs & details Deploy
Specialised 2
-
Stream an Unreal Engine app to the browser — pixel streaming on a cloud GPU.
powergpu/unreal-pixel-streaming
GPUs & details Deploy -
GPU password recovery for authorised security testing — CUDA-accelerated Hashcat.
dizcza/docker-hashcat
GPUs & details Deploy
Bring your own image
Nothing here fits? The deploy form takes any OCI reference — we inject the NVIDIA runtime and run your entrypoint untouched.
- Public or private registries (credentials stored encrypted).
- Your ports, env vars and volume mounts, set at deploy.
- Pinned by tag — a redeploy is byte-identical next month.
$ powergpu launch --gpu l40s \
--image ghcr.io/acme/trainer:v14 \
--ports 8080 --disk 100
✓ instance i-4fd02b11 running (31.2s)GPU templates FAQ
Environment variables, ports and custom images in the template docs.
What exactly is a template?
A maintained container image plus sane launch defaults: exposed ports, volume mount points, health checks and environment variables. Pick one at deploy, override anything, and the instance boots straight into a working stack — usually in about 30 seconds.
Can I use my own Docker image instead?
Yes — the deploy form accepts any public or private OCI image reference (registry credentials are stored encrypted per-account). We inject the NVIDIA runtime; your entrypoint is untouched.
What do the ARM, SSH, Jupyter and VM tags mean?
They tell you how a template runs: ARM = also builds for ARM hosts, SSH = direct shell access, Jupyter = a notebook server on a mapped port, VM = a full virtual machine rather than a container.
Do templates cost extra?
No. You pay the GPU price — a ComfyUI session on an RTX 4090 is $0.262/hr, template included, billed per second.