Skip to content
tensorlessv0.1.0

See exactly what will run

Before you download 13 GB or start a server, you can see exactly what the launcher would do: which file it downloads from where, which image it pulls by digest, and the full podman run or docker run command. tensorless info shows one recipe; --dry-run shows the plan of a command.

You need the launcher installed. See Install the launcher. Neither needs a GPU, so you can do this on a laptop.

Show a recipe

Section titled “Show a recipe”
Terminal window
tensorless info iamMrDuncan/Qwen3.8-27B-PN3-GGUF
iamMrDuncan/Qwen3.8-27B-PN3-GGUF on p100: Qwen3.8-27B-PN3 (P100-native 3-bit, mix P0k+19)
  short name      pn3
  status          released
  recipe          recipes/iamMrDuncan/Qwen3.8-27B-PN3-GGUF/p100.toml
  fits this host  unknown (nvidia-smi not found)


Model
  repository  iamMrDuncan/Qwen3.8-27B-PN3-GGUF
  revision    c701b6690ec360b16d40f85df1cb76062e30cd38
  file        Qwen3.8-27B-PN3-P0k19.gguf
  sha256      15ce2d4469673f1f1753c55de3780cfe04311dd6f05cb41633d73df128e91476
  size        12.95 GB (12954114496 bytes)
  license     Apache-2.0 (from Qwen/Qwen3.8-27B)
  url         https://huggingface.co/iamMrDuncan/Qwen3.8-27B-PN3-GGUF/resolve/c701b6690ec360b16d40f85df1cb76062e30cd38/Qwen3.8-27B-PN3-P0k19.gguf
  cache       /home/you/.cache/huggingface/hub/models--iamMrDuncan--Qwen3.8-27B-PN3-GGUF/blobs/15ce2d4469673f1f1753c55de3780cfe04311dd6f05cb41633d73df128e91476
  cached      no


Card
  title    NVIDIA Tesla P100 16 GB
  compute  6.0
  memory   at least 16000 MiB
  driver   525.60.13 or newer


Engine
  release  engine-v0.1.0
  arch     sm60
  image    ghcr.io/iammrduncan/tensorless-engine@sha256:35409ff97c2f518e468790c28bd19206ae777a945fcffe5775585e28fd9939a7

The output goes on with Run (the context, the arguments, the engine switches and the full container command) and Receipt (what was measured). Each part:

  • The header says which recipe file this is, whether it is released, and whether a GPU on this host fits it.
  • Model is the exact file: its repository and pinned revision, its sha256 and size, the download address, and where it goes in your cache.
  • Card is what the GPU needs: compute capability, memory and driver.
  • Engine is the image, by digest, and the engine release it was measured on.
  • Run is how the server is started, with the command your host would run.
  • Receipt is what was measured with exactly these settings. The Receipts page has them all.

info takes most of the flags run takes. Add --context, --port, --host, --gpu or arguments after --, and it shows the command with them, with an UNTUNED line in the header when they change the measured configuration. Add --runtime docker to see the docker command.

Get it as JSON

Section titled “Get it as JSON”
Terminal window
tensorless info pn3 --json

The JSON holds the whole recipe under recipe, and what the launcher worked out for this host: published, fits, download (the URL, the blob and snapshot paths, the size and whether it is cached), runtime, gpu, threads, command (as a list of arguments) and endpoint.

tensorless models --json and tensorless capabilities --json list every recipe in short form: model, card, status and engine release.

Show the plan of a command

Section titled “Show the plan of a command”

Add --dry-run to run, pull, serve or stop. Nothing is pulled, downloaded or started:

Terminal window
tensorless run iamMrDuncan/Qwen3.8-27B-PN3-GGUF --card p100 --dry-run
dry run of 'tensorless run iamMrDuncan/Qwen3.8-27B-PN3-GGUF': nothing is pulled, downloaded or started.
recipe:   iamMrDuncan/Qwen3.8-27B-PN3-GGUF on p100 (released), from recipes/iamMrDuncan/Qwen3.8-27B-PN3-GGUF/p100.toml
gpu:      chosen with --card; GPU detection skipped; the container gets GPU all
runtime:  podman


# 1. pull the engine image, unless 'podman image inspect' finds it
podman pull \
    ghcr.io/iammrduncan/tensorless-engine@sha256:35409ff97c2f518e468790c28bd19206ae777a945fcffe5775585e28fd9939a7


# 2. download the model (12.95 GB), resuming a partial download, and check its sha256 15ce2d4469673f1f1753c55de3780cfe04311dd6f05cb41633d73df128e91476
GET https://huggingface.co/iamMrDuncan/Qwen3.8-27B-PN3-GGUF/resolve/c701b6690ec360b16d40f85df1cb76062e30cd38/Qwen3.8-27B-PN3-P0k19.gguf
    into /home/you/.cache/huggingface/hub/models--iamMrDuncan--Qwen3.8-27B-PN3-GGUF/blobs/15ce2d4469673f1f1753c55de3780cfe04311dd6f05cb41633d73df128e91476

It goes on with the snapshot link, the full podman run command, and how it waits for the server. This example ran on a machine without the card, so --card p100 chose the recipe. On a P100 host, leave --card out: the gpu: line then names the GPU that matched.

A dry run still asks nvidia-smi for the GPU list, so it shows what would happen on this host. Without a GPU, --dry-run and info fall back to the model’s only recipe; pass --card p100 to choose it yourself.

Check that it worked

Section titled “Check that it worked”

Compare the command with the recipe file named on the recipe line, in recipes/: the image digest, the model’s revision and sha256, the arguments and the switches are the same. When you then run the command without --dry-run, the launcher downloads and starts exactly that.