Run model zoo pipelines

Discover models across the YAML zoo, standalone recipe catalog, and AxScript registry while keeping listed, downloadable, buildable, and validated states separate.

Start with the surface that exists in the SDK release you installed. The established YAML model zoo provides pipeline definitions and downloadable prebuilt payloads. The standalone recipe catalog exposes editable Python build recipes. AxScript exposes authored-graph model cards with per-card lifecycle status. A model name appearing in any list proves discovery only; it does not by itself prove that an artifact is downloaded, compiled, or validated on your hardware.

Find and run a zoo model

Discovery and runtime routes

Discover, obtain, and run a model

1. List what this release contains

   List the selected release’s model surfaces instead of relying on a copied catalog. Ask Agent Dev for a task in plain language, or use each local front door directly.

   bash

   ./deploy.py --help
   axzoo list
   python -m model_axscript list

2. Fetch or build through the owning surface

   For a YAML zoo model, fetch the prebuilt payload when one is published. For a recipe or AxScript card, build locally through that surface’s documented command. Do not describe local recipe or AxScript output as a model download.

   YAML payload

   axdownloadmodel <yaml-model>

   Recipe or AxScript

   axzoo build <config>
   python -m model_axscript build <tag>

   Published LLM

   axdownloadmodel llama-3-2-1b-1024-4core-static
   axllm llama-3-2-1b-1024-4core-static --prompt "Give me a joke"

3. Run with the documented input contract

   Use only the input forms documented by the selected pipeline or card. Common YAML pipeline sources include a video, image, USB camera, RTSP stream, or the model’s evaluation dataset; a specific model may support fewer forms. Recipe and AxScript run verbs are defined by their own help.

   YAML pipeline

   ./inference.py <yaml-model> media/traffic1_1080p.mp4
   ./inference.py <yaml-model> image.jpg
   ./inference.py <yaml-model> usb:0
   ./inference.py <yaml-model> rtsp://camera.example/stream
   ./inference.py <yaml-model> dataset

   AxScript card

   python -m model_axscript run <tag> <input>
   python -m model_axscript benchmark <tag> <input>

4. Report what was actually proved

   Record lifecycle and evidence separately: listed, payload downloaded or artifact built, runtime-loaded on Axelera hardware, and validated against the model card’s task-specific gate. A successful build cannot skip the runtime and accuracy steps.

Established zoo starting points

Keep AxScript lifecycle evidence model-specific

Read each AxScript card’s lifecycle from the selected release and keep it separate from artifact evidence. Pending, validated-source, artifact-built, payload-present, downloadable, hardware-run, and task-quality-validated are different states. Do not publish a source-snapshot inventory or turn one card’s status into a capability claim for another graph.

Separate published LLM inference from custom compilation

When no listed model is close

If none of the three surfaces contains a close model, move to Deploy a custom model. It explains safe artifact inspection, the AxScript authored-graph boundary, calibration, compilation, and the evidence required before calling a port validated.