Skip to main content

Images on Superheat CN

Superheat CN keeps its template images on Alibaba Cloud Container Registry (ACR), and GPU hosts pull them from there. This page covers where the catalog's images live, how an image reference you type yourself is pulled, and how to get your own image onto a host.

The registry​

superheat-cn-registry.cn-shanghai.cr.aliyuncs.com

The images templates use sit in two namespaces of that registry:

NamespaceContentsExample
superheat/Images Superheat publishes itselfsuperheat-cn-registry.cn-shanghai.cr.aliyuncs.com/superheat/base
mirror/Third-party images the catalog usessuperheat-cn-registry.cn-shanghai.cr.aliyuncs.com/mirror/open-webui-open-webui

Every repository listed below is public and allows anonymous pulls. Hosts need no credentials to pull them, and a template that uses them needs no private registry credentials.

superheat/ repositories​

RepositoryTags on ACR
superheat/basecuda-12.1, cuda-12.4, cuda-12.6, cuda-12.8, latest
superheat/pytorchcuda-12.1, cuda-12.4, cuda-12.6, cuda-12.8
superheat/axolotlcuda-12.4
superheat/comfyuicuda-12.4, cuda-12.8
superheat/a1111cuda-12.1
superheat/fooocuscuda-12.1
superheat/faster-whispercuda-12.6
superheat/kohyacuda-12.8

With a template tag of [Automatic], the launch picks the newest of that repository's published tags whose CUDA the chosen machine can run. The base image itself is covered in Base image.

mirror/ repositories​

A third-party image's ACR reference is derived like this: drop the source registry's host, replace each / in the remaining path with -, and put the result under mirror/. For example, ghcr.io/open-webui/open-webui becomes mirror/open-webui-open-webui; a Docker Hub official image gets library/ prefixed to its path first.

The image sync copies only the images of catalog templates you can launch, and only the one tag each template pins — mirror/ is not a full copy of the upstream repositories. The third-party images synced today:

TemplateReference on ACR (registry host omitted)Tag
TensorFlow (Jupyter)mirror/tensorflow-tensorflowlatest-gpu-jupyter
Data Science (Jupyter)mirror/jupyter-datascience-notebooklatest
Open WebUImirror/open-webui-open-webuimain
Text Embeddings Inferencemirror/huggingface-text-embeddings-inference1.5

vLLM (OpenAI API), Ollama, InvokeAI and LLaMA-Factory are no longer synced. Their servers cannot require the instance access token on every request, so they are marked coming soon until they can, and the image sync skips them. Instances already running are unaffected.

Catalog references and the ones you type​

Curated templates have their image references rewritten to ACR references when they are written into the template catalog. So a curated template's card, its detail view and any duplicate of it all show the ACR reference, and hosts pull from ACR.

An image reference you enter when you create or edit a template is pulled verbatim, with no rewriting. Write pytorch/pytorch and the host goes to Docker Hub for it; Docker Hub is unreachable from mainland hosts, so the deploy fails.

So either duplicate a curated template and edit the copy, or write a full reference that mainland hosts can pull: one of the ACR references above, or your own mainland registry.

Registries mainland hosts cannot pull from​

RegistryFrom mainland hosts
Docker Hub, including short names without a registry host such as pytorch/pytorchUnreachable
GHCR (ghcr.io)Very slow
quay.io, nvcr.ioSlow or blocked

An instance whose image cannot be pulled does not reach running, and the instance logs carry the pull error. See Logs.

nvcr.io templates show "Coming soon"​

Images from NVIDIA NGC (nvcr.io) are not copied to ACR: the NGC licence does not clearly permit redistribution. So the catalog's nvcr.io templates — Triton Inference Server and RAPIDS, for example — are flagged coming soon on Superheat CN and cannot launch on any machine. The card shows a Coming soon badge and a disabled Launch button with the tooltip "Execution ships in a later release."; in the detail dialog the same button reads Coming soon (即将推出).

You can still open these templates to inspect them. See What "coming soon" means.

ssh and jupyter modes need an image on ACR​

In ssh and jupyter modes Superheat takes over the container with the base image's entrypoint: it starts sshd (plus JupyterLab in jupyter mode) and your on-start script, and opens the SSH port. Superheat decides whether to take over from the image reference. On Superheat CN:

Image on the templateWhat happens in ssh / jupyter mode
superheat-cn-registry.cn-shanghai.cr.aliyuncs.com/superheat/<repository>Superheat takes over the entrypoint and the instance starts normally
superheat/base and other spellings without the ACR hostThe deploy fails: Superheat CN only accepts the base image on ACR
Any other image, including one you built on the base image and pushed to your own registryRuns its own entrypoint. Superheat does not take it over and does not open an SSH port for it

For that last kind of image, use args mode and declare the ports it serves on the template. See Launch modes and Ports.

Using your own image​

  1. Push the image to a registry in mainland China, such as your own Alibaba Cloud ACR instance.
  2. In the template's Image field, enter the full reference including the registry host, for example acme-registry.cn-shanghai.cr.aliyuncs.com/acme/trainer. The registry host is derived from this reference.
  3. Pin a tag that exists in the Tag field. [Automatic] resolves to a cuda-<minor> tag only for images under superheat/; for any other image it means latest.
  4. If the repository is private, fill in a username and an access token in the template's Private registry section — for example an access credential for your ACR instance. See Private registries.
  5. Set Mode to Args (参数) and declare the ports the image serves. The previous section explains why.

If you want the tools in the base image, build FROM the base image on ACR, which allows anonymous pulls. See Base image.