ExecuTorch Extension Model Explorer

ExecuTorch Extension Model Explorer is a combined extension package for Model Explorer. It registers adapters and data providers for inspecting ExecuTorch model artifacts and profiling data in one package:

• PTEAdapter visualizes .pte model files.
• ETRecordAdapter visualizes .etrecord graph files.
• ETDumpProfilingOverlayProvider overlays runtime profiling data from .etdp or .etdump files onto ETRecord graphs.

Supported File Types

Requirements

• Python >=3.10, <3.13
• Model Explorer, provided by ai-edge-model-explorer
• ExecuTorch Python packages for ETRecord and ETDump workflows
• PTE delegate support packages installed through the PTE adapter: vgf-adapter-model-explorer, tosa-adapter-model-explorer, and flatbuffers

The umbrella depends directly on PTE, ETRecord and ETDump only. Those packages declare and install their own Model Explorer, ExecuTorch and delegate requirements; the umbrella does not duplicate those version constraints.

Installation

For local development from this repository:

scripts/install-editable.sh

Install the published package from PyPI:

pip install executorch-extension-model-explorer

Usage

Launch Model Explorer with this combined extension package enabled:

model-explorer --extensions=executorch_extension_model_explorer

To load individual adapters or providers, pass the extension names directly:

model-explorer --extensions=pte_adapter_model_explorer
model-explorer --extensions=etrecord_adapter_model_explorer
model-explorer --extensions=etrecord_adapter_model_explorer,etdump_data_provider_model_explorer

PTE Adapter

Open a .pte file in Model Explorer to inspect the ExecuTorch program. The PTE adapter includes delegate support for Arm Ethos-U, Arm VGF, and XNNPACK. See the PTE adapter documentation for details and schema regeneration notes.

ETRecord Adapter

Open an .etrecord file in Model Explorer to inspect exported ExecuTorch graph structure, metadata, connections, inputs, and outputs. See the ETRecord adapter documentation for details.

ETDump Data Provider

After opening an .etrecord graph, click Add per-node data, select ETDump Profiling Overlay, upload an .etdp or .etdump file, and run the provider. The overlay maps runtime latency metrics onto matching ETRecord graph nodes. See the ETDump data provider documentation for details and mock artifact tooling.

Development

Install the package in editable mode, then run the test suite with pytest:

scripts/install-editable.sh
scripts/test-all.sh

For coordinated development across all leaf repos, keep the adapter repositories as sibling checkouts and install them in editable mode before the umbrella package:

scripts/install-local-adapters.sh

The umbrella wheel uses the package versions pinned in pyproject.toml. Local adapter checkouts are for multi-repo development and adapter-specific testing; they are not the release pin.

Run adapter-specific tests from each adapter repository root so their local fixtures and pytest configuration are resolved correctly.

Development dependency groups are declared in pyproject.toml for test, coverage, lint, type-checking, and build tools.

CI

Pull requests and pushes to main run the aggregate CI workflow. It installs this package in editable mode, then runs Ruff, Pyright, and the aggregate pytest suite. The workflow also builds a wheel after tests pass.

Internal pushes to main publish development builds to Artifactory. Internal release tags publish release builds there for testing. Tags are not automatically copied to the public repository.

Use the Trigger Release workflow to cut a release from the private repository. It calculates the next version with git-cliff, updates pyproject.toml, pushes the version bump commit, and pushes the internal release tag. After testing, run Sync release tag to public repo with that tag. The public tag triggers PyPI publication and a GitHub release. Main commits still sync automatically.

Dependency update policy

This repository declares adapter/provider minimum versions in pyproject.toml. PTE, ETRecord and ETDump notify it after successful internal tagged releases to Artifactory, not after main development builds or public releases.

After installing the development dependencies, use the helper script for local dependency lower-bound updates:

python scripts/update-dependency-pin.py etrecord-adapter-model-explorer v0.2.0

The same update can be automated through the Update dependency lower bound workflow. Leaf release workflows can trigger it with a repository_dispatch event named update-dependency and a payload containing dependency or name, plus version or tag. The workflow also accepts the old update-submodule dispatch name for compatibility, but it now updates pyproject.toml. Stable versions and package prereleases such as v0.5.1rc1 are accepted. Updates preserve other constraints and never lower an existing minimum. Each update opens a PR for normal CI and review; it does not merge or publish the umbrella automatically.

Trademark Notice

This project uses some of the Arm product, service, or technology trademarks, as listed in the Trademark List, in accordance with the Arm Trademark Use Guidelines.

Subsequent uses of these trademarks throughout this repository do not need to be prefixed with the Arm word trademark.

Contributions

We are not accepting direct contributions at this time. If you have feedback or feature requests, please use the repository issues section.