A template is a Docker image plus the right command, ports and settings. Choose one when you rent a GPU, or pass its id as template in the API.
- Min VRAM — total GPU memory the template needs (a chosen model may need more).
- Min CUDA — the host’s NVIDIA driver must support at least this CUDA version; other machines are skipped automatically.
- Multi-GPU —
tensor-parallel: the model is split across all GPUs automatically; distributed: all GPUs are visible, use torchrun/DeepSpeed/FSDP; —: the app uses one GPU. See Multi-GPU.
- Web interfaces are opened through SSH port forwarding; APIs can also be called through the proxy.
| id |
Name |
Image |
Ports |
Min VRAM |
Min CUDA |
Multi-GPU |
pytorch |
PyTorch 2.8 |
pytorch/pytorch:2.8.0-cuda12.8-cudnn9-devel |
— |
8 GB |
12.8 |
distributed |
jupyter |
Jupyter Lab + PyTorch |
pytorch/pytorch:2.8.0-cuda12.8-cudnn9-devel |
8888 Jupyter Lab |
8 GB |
12.8 |
distributed |
tensorflow |
TensorFlow + Jupyter |
tensorflow/tensorflow:latest-gpu-jupyter |
8888 Jupyter |
8 GB |
12.3 |
distributed |
cuda |
Ubuntu 24.04 + CUDA 12.8 |
nvidia/cuda:12.8.1-cudnn-devel-ubuntu24.04 |
— |
8 GB |
12.8 |
distributed |
custom |
Custom Docker image |
your image |
— |
8 GB |
— |
distributed |
- pytorch, cuda — idle containers for SSH work in
/workspace.
- jupyter, tensorflow — Jupyter opens in
/workspace; the login token is GYSGA_TOKEN.
- custom — any public image; its own entrypoint and command run as-is.
| id |
Name |
Image |
Ports |
Min VRAM |
Min CUDA |
Multi-GPU |
Model |
vllm |
vLLM · OpenAI API |
vllm/vllm-openai:v0.30.0-cu129 |
8000 OpenAI API |
16 GB |
12.9 |
tensor-parallel |
required (LLM) |
sglang |
SGLang · OpenAI API |
lmsysorg/sglang:latest |
30000 OpenAI API |
16 GB |
12.8 |
tensor-parallel |
required (LLM) |
open-webui |
Open WebUI + Ollama |
ghcr.io/open-webui/open-webui:ollama |
8080 Open WebUI, 11434 Ollama API |
8 GB |
12.0 |
— |
pulled in the UI |
ollama |
Ollama API |
ollama/ollama:0.34.4 |
11434 Ollama API |
8 GB |
12.0 |
— |
ollama pull |
text-generation-webui |
Text Generation WebUI |
atinoda/text-generation-webui:default-nvidia |
7860 Web UI, 5000 API |
8 GB |
12.1 |
— |
downloaded in the UI |
- vllm — high-throughput OpenAI-compatible server. The model name in requests is its Hugging Face id (e.g.
Qwen/Qwen3-8B). It uses 92% of GPU memory for weights and KV cache.
- sglang — fast serving with RadixAttention, OpenAI-compatible on port 30000.
- open-webui — a ChatGPT-like interface with Ollama inside; the first account you create in it becomes its admin. Ollama models are stored in
/workspace/ollama.
- ollama — Ollama server; models in
/workspace/ollama. Its OpenAI-compatible API is at /v1.
- text-generation-webui — oobabooga UI for GGUF, EXL2 and Transformers models, with an API on port 5000.
| id |
Name |
Image |
Ports |
Min VRAM |
Min CUDA |
Model |
comfyui |
ComfyUI |
yanwk/comfyui-boot:cu128-slim |
8188 ComfyUI |
12 GB |
12.8 |
optional (image / video) |
Node-based generation of images and video (SDXL, FLUX, HiDream, Qwen-Image, Wan, Mochi). ComfyUI-Manager is included for custom nodes and extra models. When you choose a model, its files are mounted at /models/<model-id> and ComfyUI is configured to find them automatically — they appear in the loaders right away.
| id |
Name |
Image |
Ports |
Min VRAM |
Min CUDA |
speaches |
Speech-to-text & TTS · OpenAI API |
ghcr.io/speaches-ai/speaches:latest-cuda |
8000 Audio API |
6 GB |
12.4 |
whisper |
Whisper large-v3 API |
onerahmet/openai-whisper-asr-webservice:latest-gpu |
9000 ASR API (Swagger at /docs) |
8 GB |
12.0 |
kokoro |
Kokoro TTS · OpenAI API |
ghcr.io/remsky/kokoro-fastapi-gpu:latest |
8880 TTS API (UI at /web) |
4 GB |
12.4 |
- speaches — faster-whisper transcription and TTS behind OpenAI-compatible
/v1/audio/transcriptions and /v1/audio/speech.
- whisper — speech recognition with faster-whisper large-v3.
- kokoro — fast natural TTS (82M) with
/v1/audio/speech.
| id |
Name |
Image |
Ports |
Min VRAM |
Min CUDA |
Multi-GPU |
llamafactory |
LLaMA-Factory |
hiyouga/llamafactory:latest |
7860 LLaMA Board |
16 GB |
12.1 |
distributed |
axolotl |
Axolotl |
axolotlai/axolotl:main-latest |
— |
16 GB |
12.4 |
distributed |
unsloth |
Unsloth |
unsloth/unsloth:latest |
— |
12 GB |
12.4 |
distributed |
- llamafactory — fine-tune 100+ LLMs with LoRA/QLoRA or full training from a web UI. Hugging Face cache in
/workspace/hf.
- axolotl — YAML-driven fine-tuning (LoRA, QLoRA, full, DPO). Connect over SSH and run
axolotl train config.yml.
- unsloth — faster, memory-efficient fine-tuning; run your Unsloth scripts over SSH.
| id |
Name |
Image |
Ports |
Min VRAM |
Min CUDA |
Model |
embeddings |
Embeddings · OpenAI API |
vllm/vllm-openai:v0.30.0-cu129 |
8000 OpenAI API |
8 GB |
12.9 |
required (embed) |
Serves embedding models through /v1/embeddings and rerankers through /v1/rerank — for search and RAG.
