Title here
Summary here
Formats
Capability matrix
| Format | Inline weights ( weights) | External weights | Subgraph inlining | Multi-subgraph |
|---|---|---|---|---|
| ONNX | yes - parseOnnx populates graph.weights from initializer payloads | not loaded | yes - If / Loop / Scan bodies inlined with prefixed names, depth ≤ 4 | n/a |
| TFLite | yes - parseTflite exposes the buffer table via graph.weights | not loaded | no | first subgraph only - secondary subgraphs (If / While bodies) are not yet inlined |
| Keras | no - .weights.h5 lives outside the .keras archive and is not read | not loaded | n/a | n/a |
| TorchScript | no - bytecode does not carry tensor payloads | not loaded | n/a | n/a |
| ExecuTorch | no - constant buffers are not surfaced | not loaded | n/a | only first execution plan |
SavedModel (saved_model.pb) | no inline payload; Const shape/dtype folded into initializers | yes - TF2 checkpoint pair via loadSavedModelWeights + attachCheckpointToGraph | yes - StatefulPartitionedCall / PartitionedCall function bodies inlined | n/a |
SavedModel (keras_metadata.pb) | no | not loaded | n/a | n/a |
All parsers set ModelGraph.fileSizeBytes. The renderer’s weight panel uses it to gate decoding for files larger than 20 MB.