---
sidebar_position: 30
---

> ## Documentation Index
> Fetch the complete documentation index at: https://docs.lium.io/llms.txt
> Use this file to discover all available pages before exploring further.

# `lium templates`

List available Docker templates and images used to create pods.

```bash
lium templates [SEARCH] [--arch hopper|blackwell] [--format json]  # --arch: lium#132, since 0.0.46
```

## Arguments

- `SEARCH` — Optional filter string. Templates and images whose name matches are shown; omit it to list everything.

## Options

:::note Since lium 0.0.46
`--arch`, the *Runs on* column and the `cuda_version`, `torch_version` and `arch` JSON fields shipped in lium 0.0.46 (lium#132, released 14 Sep 2026). `--format json` and `--json` exist since 0.0.39. On an older CLI, `pip install -U lium.io` first.
:::

| Flag | Effect |
|------|--------|
| `--arch hopper\|blackwell` | Only templates whose CUDA build runs on that GPU generation (Blackwell needs CUDA 12.8 or newer) *(lium#132, since 0.0.46)* |
| `--format table\|json`, `--json` | JSON rows carry `id`, `huid`, `name`, `docker_image`, `docker_image_tag`, `category`, `status` (since 0.0.39), plus `cuda_version`, `torch_version` and `arch` *(lium#132, since 0.0.46)* |

## "Runs on"

Since lium 0.0.46 (lium#132) the table has a **Runs on** column: the CUDA build the image was made with, then the GPU generations it runs on — `13.0 Hopper+Blackwell` (CUDA 12.8+), `12.4 Hopper only` (11.8 – 12.7) or `11.7 pre-Hopper`. In JSON the same reading is `arch`: `hopper+blackwell`, `hopper`, `pre-hopper`, or `null` when neither the tag nor the image name carries a CUDA version. Nothing in a template record says "Blackwell" — the CLI reads the CUDA version from the image tag, then the image name (`cu126`, `cuda12.6`, `cuda13.0.2` …). A tag it cannot read (`latest`, `12.6.0-cudnn-…`) shows `?` and `--arch` leaves that template out (lium#132, since 0.0.46). A template built on CUDA 12.6 or older has no kernels for B200, B300, RTX PRO 6000 or RTX 50x0 nodes: the first CUDA call fails with *no kernel image is available*. Pick with `--arch blackwell` before renting one of those.

## Examples

```bash
lium templates            # List all templates and images
lium templates pytorch    # Show entries matching "pytorch"
lium templates --arch blackwell           # Only images that run on B200 / B300 / RTX PRO 6000 / RTX 50x0 — lium#132, since 0.0.46
lium templates pytorch --format json | jq '.[] | {id, name, cuda_version, arch}'  # cuda_version, arch: lium#132, since 0.0.46
lium templates --format json              # Full objects: id, huid, name, docker_image, docker_image_tag, … — since 0.0.39
```

To create a pod from a template, pass its ID to [`lium up --template_id`](./up). Since 0.0.39 the table shows the id and `--format json` (or `--json`) gives the full object; earlier versions print neither.

## See also

- [Templates](/pod-users/templates) — concept guide
- [`lium up`](./up) — pick a template at pod creation