Model Adapters
The plugin selects a Model Explorer adapter from the effective model's filename suffix. It then requires a verified mapping from MLIA source-operator entity IDs to the adapter's graph nodes. Adapter installation and end-to-end mapping support are therefore separate requirements.
Support matrix
| Format | Dependency | Source-identity behavior | End-to-end status |
|---|---|---|---|
.pt2 | Base ai-edge-model-explorer dependency | Matches top-level exported-program FX node names to one adapter graph. | Supported |
.tflite | Base ai-edge-model-explorer dependency | Uses operator indices within each flatbuffer subgraph; identical indices remain graph-qualified. | Supported |
.onnx | mlia-model-explorer[onnx] | Correlates top-level ONNX nodes through name, output tensors, and operation labels, requiring a complete one-to-one match. | Supported with restrictions |
.spv | mlia-model-explorer[spirv] | Uses native SPIR-V result-ID metadata exposed by the adapter. | Supported |
.vgf | mlia-model-explorer[vgf] | Combines VGF segment identity with native SPIR-V result-ID metadata. | Supported |
.etrecord, .bin | mlia-model-explorer[executorch] | Uses method-qualified ETRecord source identities and measured-output compiler associations. | Supported for ETRecord profiling |
Companion .pte | mlia-model-explorer[executorch] | Consumes analyzed method/instruction/segment/SPIR-V identities; requires matching structural IDs from the adapter. | VGF companion only; see ETRecord profiling |
.tosa | mlia-model-explorer[tosa] | No verified native source-operator identity contract is implemented. | Not yet supported end to end |
PTE visualization is available through --companion-model with an ETRecord input.
Standalone .pte inputs are not supported; they are rejected before adapter
execution because no primary-PTE source-identity mapping is implemented.
Common adapter requirements
For non-ETRecord inputs, adapter output must normalize to exactly one graph collection. Multiple graph collections are rejected because flattening them could collide graph IDs or break subgraph references. ETRecord adapters may also return recorded deployable and delegate collections; the plugin explicitly selects one identifiable ETRecord source collection. Within the accepted collection, graph and node IDs used for mapping must be valid and unambiguous.
The source graph collection is always relabeled Source Model; this stable label
is used to route node-data overlays and qualify mapping references.
ONNX restrictions
ONNX mapping currently supports the top-level ModelProto.graph only. Models
with ONNX functions or graph-valued attributes are rejected because their node
identity requires a future graph-qualified native ID format. Initializer, input,
and output helper nodes emitted by the adapter are not treated as source
operations.
Names alone are insufficient when they are duplicated. The mapper combines all available authoritative evidence and rejects missing, ambiguous, conflicting, or incomplete matches rather than attaching MLIA data to a plausible-looking node.
Adapter and visualizer versions
The Python adapter package converts models; it does not provide the runtime used
by the standalone page. Adapter packages may themselves depend on
ai-edge-model-explorer because they use its adapter API and graph-building
helpers; this remains separate from serving the UI.
Both files and open use the vendored standalone Model Explorer runtime.
Loading these artifacts in a separately installed Model Explorer server is
unsupported until that project builds its web_app, bumps the package version,
and publishes a release containing the graph-mapping runtime. Do not infer or
substitute an upstream version that has not been published.