How qualcomm/ai-hub-models works
Qualcomm AI Hub Models is a catalog and Python package for deploying machine-learning models on Qualcomm devices, with a companion catalog/download CLI. Users browse or fetch published assets, or select a model and run its export workflow to compile, optionally quantize, profile, infer, and download; they can also evaluate accuracy or run local/on-device demos. The graph captures documented capabilities and indexed module relationships; several internal utility paths are named in the source index but were not excerpted, so their detailed behavior is not inferred.
Main entry points: export/pipeline.py Export orchestration · models Model recipes · app.py Demo applications · evaluate/pipeline.py Accuracy evaluation
Components 24
Catalog CLI
- Catalog commands
cli/qai_hub_models_cli/cli.py - CLI common
cli/qai_hub_models_cli/common.py - Asset fetch
cli/qai_hub_models_cli/fetch.py
Models and Data
- Model recipes
src/qai_hub_models/models - Demo applications
src/qai_hub_models/models/templates/yolo/app.py - Dataset adapters
src/qai_hub_models/datasets
Export and Evaluation
- Export orchestration
src/qai_hub_models/utils/export/pipeline.py - Collection export
src/qai_hub_models/utils/export/collection_pipeline.py - Multi-graph export
src/qai_hub_models/utils/export/multi_graph_collection_pipeline.py - Accuracy evaluation
src/qai_hub_models/utils/evaluate/pipeline.py - Collection evaluation
src/qai_hub_models/utils/evaluate/collection_pipeline.py - Workbench helpers
src/qai_hub_models/utils/qai_hub_helpers.py
Runtime Foundations
- Inference executors
src/qai_hub_models/utils/inference.py - Asset loading
src/qai_hub_models/utils/asset_loaders.py - Tensor specifications
src/qai_hub_models/utils/input_spec.py - Image processing
src/qai_hub_models/utils/image_processing.py - Model interfaces
src/qai_hub_models/utils/base_model.py - App interfaces
src/qai_hub_models/utils/base_app.py - Dataset interfaces
src/qai_hub_models/utils/base_dataset.py
Model Metadata
- Model configuration
src/qai_hub_models/configs/model_metadata.py - Recipe manifests
src/qai_hub_models/configs/manifest_yaml.py
Other components
- Model user
- AI Hub Workbench
- Qualcomm device
Connections 25
- Model user → Catalog commands: browse or fetch
README.md - Catalog commands → Asset fetch: fetch assets
cli/qai_hub_models_cli/cli.py - Catalog commands → CLI common: use shared CLI
cli/qai_hub_models_cli/cli.py - Model user → Model recipes: choose model
README.md - Model recipes → Demo applications: provide app example
src/qai_hub_models/models/templates/yolo/app.py - Demo applications → Image processing: preprocess inputs
src/qai_hub_models/models/templates/yolo/app.py - Demo applications → Inference executors: run inference
README.md - Model recipes → Model interfaces: implement model interface
- Dataset adapters → Dataset interfaces: extend dataset interface
src/qai_hub_models/datasets/__init__.py - Model user → Export orchestration: request export
README.md - Export orchestration → Collection export: dispatch collection
src/qai_hub_models/utils/export/pipeline.py - Export orchestration → Multi-graph export: dispatch multi-graph
src/qai_hub_models/utils/export/pipeline.py - Export orchestration → Recipe manifests: load recipe metadata
src/qai_hub_models/utils/export/pipeline.py - Export orchestration → Asset loading: load model assets
src/qai_hub_models/utils/export/pipeline.py - Collection export → Workbench helpers: use Workbench helpers
src/qai_hub_models/utils/export/collection_pipeline.py - Collection export → AI Hub Workbench: compile and profile
README.md - AI Hub Workbench → Qualcomm device: run on device
README.md - Model user → Accuracy evaluation: evaluate accuracy
README.md - Accuracy evaluation → Collection evaluation: dispatch collection
src/qai_hub_models/utils/evaluate/pipeline.py - Accuracy evaluation → Dataset adapters: evaluate with data
- Accuracy evaluation → Inference executors: run model executor
src/qai_hub_models/utils/evaluate/pipeline.py - Inference executors → AI Hub Workbench: submit inference
README.md - Model interfaces → Tensor specifications: describe inputs
- Demo applications → App interfaces: wrap model inference
README.md - Model recipes → Model configuration: publish model details