voyager-list-models

name	voyager-list-models
description	List and search models in the Voyager SDK model zoo for Axelera AI hardware. Use when the user wants to find available models by task, framework, or dataset.
argument-hint	[search criteria]
allowed-tools	Read, Bash, Glob, Grep, mcp__voyager__*

List Available Models

List and search models in the Voyager SDK model zoo for Axelera AI hardware

Important Context

Do not call a model deployable or validated until SDK evidence and runtime artifacts show that status.

Instructions

List models based on the specified criteria: $ARGUMENTS

Step 0: Data Source & Environment Selection

{{INCLUDE common/voyager-sdk-setup.md}}

Step 0.5: Axelera Voyager Project & Task Integration

{{INCLUDE common/voyager-task-integration.md}}

Step 1: List All Models

# Show all available models
./deploy.py --help

# Or use make (help is the default target; prints per-model help)
make help

Classify each result as:

• listed: YAML or deploy help entry exists
• payload-present: compiled payload exists under build/<model>/
• downloadable: axdownloadmodel <model> succeeds and payload exists
• metis-validated: inference.py ran on visible Metis hardware and produced output

Step 2: Search by Task Type

Search models by their primary task:

Object Detection:

# YOLO models
ls ax_models/zoo/yolo/object_detection/

# Oriented bounding boxes (OBB)
ls ax_models/zoo/yolo/obb_detection/

# SSD models
ls ax_models/zoo/tensorflow/object_detection/

# RetinaFace (face detection)
ls ax_models/zoo/torch/ | grep -i retinaface

Classification:

# TorchVision classifiers
ls ax_models/zoo/torchvision/classification/

# Timm models
ls ax_models/zoo/timm/

Segmentation:

# Instance segmentation (YOLO-based)
ls ax_models/zoo/yolo/instance_segmentation/

# Semantic segmentation
ls ax_models/zoo/mmlab/mmseg/
ls ax_models/zoo/yolo/semantic_segmentation/

Pose Estimation:

# Pose models (YOLO-based keypoint detection)
ls ax_models/zoo/yolo/keypoint_detection/

Depth Estimation:

ls ax_models/zoo/torch/ | grep -i depth

LLM/Language Models:

ls ax_models/zoo/llm/
ls ax_models/llm/

Step 3: Search by Framework

The zoo YAMLs declare their framework via the class: key (there is no source: key). Search by model class:

# ONNX models
find ax_models/zoo -name "*-onnx.yaml"

# Timm models
find ax_models/zoo -name "*.yaml" | xargs grep -l "class: AxTimmModel"

# Ultralytics models
find ax_models/zoo -name "*.yaml" | xargs grep -l "class: AxUltralyticsYOLO"

# Darknet YOLO models (v1.7.0 adds yolov4 variants)
find ax_models/zoo -name "*.yaml" | xargs grep -l "class: AxYoloDarknet"

# TorchVision models
ls ax_models/zoo/torchvision/

Step 4: Search by Dataset

# COCO trained models
find ax_models -name "*coco*.yaml"

# ImageNet trained models
find ax_models -name "*imagenet*.yaml"

# Custom/other datasets
find ax_models -name "*widerface*.yaml"
find ax_models -name "*lfw*.yaml"

Step 5: Reference Pipelines

# Cascade pipelines (multi-model)
ls ax_models/reference/cascade/

# Parallel pipelines
ls ax_models/reference/parallel/

# With tracker
ls ax_models/reference/cascade/with_tracker/

# Image preprocessing pipelines
ls ax_models/reference/image_preprocess/

Step 6: Model Information

Get details about a specific model:

# View the model YAML directly
cat ax_models/zoo/yolo/object_detection/yolov8n-coco.yaml

# Print per-model help for the whole zoo (default make target)
make help

Step 7: List by Model Size/Speed

YOLO model variants (speed vs accuracy), using the YOLO11 family as an example; zoo YOLOv5/YOLOv8 builds stop at the l size:

| Model    | Size  | Speed | Accuracy |
|----------|-------|-------|----------|
| yolo11n  | nano  | fast  | lower    |
| yolo11s  | small | fast  | good     |
| yolo11m  | medium| med   | better   |
| yolo11l  | large | slow  | high     |
| yolo11x  | xlarge| slow  | highest  |

Step 8: Tutorial Models

# General tutorials
ls ax_models/tutorials/general/

# YOLO tutorials
ls ax_models/tutorials/yolo/

# Custom training tutorials
ls ax_models/tutorials/torch/
ls ax_models/tutorials/onnx/

Step 9: Custom Search

# Find all detection models
find ax_models -name "*.yaml" -exec grep -l "decodeyolo\\|ssd\\|retinaface" {} \;

# Find models with specific input size
find ax_models -name "*.yaml" -exec grep -l "640, 640" {} \;

# Find models with tracker support
find ax_models -name "*tracker*.yaml"

Step 10: Model Categories Summary

Detection (COCO unless noted):
- yolov5n/s/m/l-v7-coco          YOLOv5 v7 (also yolov5s-v5, yolov5s-relu)
- yolov7-coco, yolov7-tiny-coco, yolov7-640x480-coco
- yolov8n/s/m/l-coco             YOLOv8 variants
- yolov9t/s/m/c-coco-onnx        YOLOv9 (plus gelan-s/m/c-coco-onnx)
- yolov10n/s/b-coco-onnx         YOLOv10 variants
- yolo11n/s/m/l/x-coco-onnx      YOLO11 family
- yolo26n/s/m/l/x-coco-onnx      YOLO26 family
- yolov3-coco-onnx               YOLOv3
- yolov4-416-coco, yolov4-csp-leaky-coco   Darknet YOLOv4 (new in 1.7)
- yolox-s/m-coco-onnx, yolox-x-crowdhuman-onnx
- yolonas-s/m/l-coco-onnx        YOLO-NAS
- ssd-mobilenetv1/v2-coco-poc-onnx
- retinaface-*                   Face detection (WiderFace)

Oriented bounding boxes (DOTAv1):
- yolov8n/l-obb-dotav1-onnx
- yolo11n/l-obb-dotav1-onnx
- yolo26n/s/m/l/x-obb-dotav1-onnx

Classification (ImageNet):
- resnet18/34/50/101/152-imagenet (plus resnet10t via timm)
- efficientnet_b0..b4-imagenet
- mobilenetv2-imagenet, mobilenetv3_small/large-imagenet
- mobilenetv4_small/medium/large/aa_large-imagenet (timm)
- densenet121, squeezenet1.0/1.1, inception_v3, regnet_x/y,
  resnext50_32x4d, wide_resnet50

Pose (COCO keypoints):
- yolov8n/s/m/lpose-coco
- yolo11n/lpose-coco-onnx
- yolo26n/s/m/l/xpose-coco-onnx

Instance segmentation (COCO):
- yolov8n/s/m/lseg-coco
- yolo11n/lseg-coco-onnx
- yolo26n/s/m/l/xseg-coco-onnx

Semantic segmentation (Cityscapes):
- unet_fcn_256/512-cityscapes
- yolo26n/s/m/l/xsem-cityscapes-onnx

Depth:
- fastdepth-nyudepthv2-onnx      Monocular depth

Face:
- facenet-lfw                    Face recognition
- retinaface-*                   Face detection

Re-identification and other:
- osnet-x1-0-market1501-onnx, sbs-s50-market1501-onnx   Re-ID
- lprnet                         License plate recognition
- real-esrgan-x4plus-onnx        Super resolution

LLM:
- llama-3-1-8b / llama-3-2-1b / llama-3-2-3b (1024-4core-static)
- phi3-mini-512-static, phi3-mini-1024/2048-4core-static
- velvet-2b-1024-4core-static

Treat this table as illustrative only. Verify exact model names against ./deploy.py --help, ax_models/**/*.yaml, or Voyager SDK RAG before recommending them.

Step 11: Check Model Availability

# Verify model is listed
./deploy.py --help | grep -F "<model-name>"

# Verify YAML exists
find ax_models -name "<model-name>.yaml" -o -name "*<model-name>*.yaml"

# Verify compiled or downloaded payload
find build/<model-name> -type f \( -name "*.axnet" -o -name "model.json" \) -print

Availability is not hardware validation. Only label a model validated after a bounded inference.py run on a probed Metis host succeeds and produces an output artifact.