Usage

Installation

Install the plugin in the same Python environment as MLIA:

python -m pip install mlia-model-explorer

The base package includes Model Explorer's built-in PyTorch ExportedProgram and TFLite conversion support. Install an extra when the model format needs a separate adapter, for example:

python -m pip install "mlia-model-explorer[onnx]"
python -m pip install "mlia-model-explorer[spirv]"
python -m pip install "mlia-model-explorer[vgf]"
python -m pip install "mlia-model-explorer[executorch]"

See Model Adapters before choosing an extra; installing an adapter does not by itself guarantee that this plugin has a verified source-identity mapping for the format.

The ETRecord profiling guide adds measured ETDump commands, method selection and an optional PTE companion to the workflows below.

Generate a standalone visualization

Use files when another process will archive, publish, or open the result:

mlia check MODEL --target-profile TARGET --compatibility --model-explorer files

Use open for an interactive local run:

mlia check MODEL --target-profile TARGET --compatibility --model-explorer open

With the default MLIA output directory, artifacts are written under mlia-output/model-explorer/ and the page is mlia-output/model-explorer/bundle/index.html. The HTML is self-contained and can be moved or opened directly with a file:// URL.

Interpreting the view

For ordinary inputs, when both are available, the initial layout shows the source model beside the generated MLIA hierarchy with the richest overlay coverage. Select placement, status, metric, check, or advice overlays from the node-data controls. Synchronized navigation uses explicit mappings so selecting related nodes across graph panes does not depend on coincidentally matching node IDs.

Generated hierarchies reflect standardized output exactly. Configure entity or code-stack collapse through MLIA core's [[filtering.collapse]] settings; legacy Model Explorer plugin settings are rejected.

Collapsed entity frames

Entity collapse is owned by MLIA core and is applied before this plugin receives standardized output. Model Explorer does not parse filtering settings or perform a second transformation; its standardized-output artifact, generated hierarchy, mappings, and node data all reflect the processed result unchanged.

Configure structured [[filtering.collapse]] rules in MLIA's user config.toml. MLIA loads this file from the platform-standard configuration directory, such as %LOCALAPPDATA%\arm\mlia\config.toml on Windows or $XDG_CONFIG_HOME/mlia/config.toml on Linux. See the MLIA configuration documentation for defaults, raw-value glob semantics, and how to disable collapse.

PyTorch code stacks

For .pt2 models, prefer torch.export.export(..., strict=True). Strict export uses PyTorch Dynamo's logical source stack and generally preserves ordinary helper-function frames and their individual call-site lines.

In recent PyTorch versions, non-strict export may fall back to live FX tracing that filters stack information mainly to forward methods. Operations on different lines inside an ordinary helper function can then receive the same recorded call site. Later PyTorch or ExecuTorch transformations may preserve, replace, or append frames, so displayed stacks remain provenance information rather than a guaranteed complete runtime Python call stack.

Troubleshooting

Symptom	Action
Adapter import error	Install the extra or adapter package named in the error, in the same environment as MLIA.
Unsupported model format or mapping	Check the adapter matrix; the plugin fails rather than guessing node identity.
Materialized model hash or size mismatch	Regenerate the backend output and ensure the model file is directly inside the MLIA output root.
Plugin settings are no longer supported	Move collapse and filtering rules to MLIA core configuration.
Vendored visualizer is missing	Restore it from the package or follow the contributor update procedure in Development.
Page does not open automatically	Use files, then open bundle/index.html manually in a modern browser.