ETRecord Adapter for Model Explorer

etrecord-adapter-model-explorer is the standalone Model Explorer adapter for ExecuTorch .etrecord files.

This repository contains the ETRecord-specific adapter code, tests, and fixtures. For a combined install that wires the PTE adapter, ETRecord adapter, and ETDump data provider together, use the umbrella package executorch-extension-model-explorer.

Installation

Install the standalone adapter package:

pip install etrecord-adapter-model-explorer

The adapter expects an environment with the ExecuTorch Python packages available.

Usage

Run Model Explorer with the standalone ETRecord adapter enabled:

model-explorer --extensions=etrecord_adapter_model_explorer

Then open an .etrecord file.

Delegate visibility

Delegated ExecuTorch programs often collapse a backend-supported subgraph into an opaque executorch_call_delegate node. If an ETRecord is generated from the post-delegation edge graph, this adapter can only display that delegate call for the delegated region because the original edge operators are no longer present in the recorded edge graph.

To keep the original operators visible, generate the ETRecord with the pre-delegation EdgeProgramManager as edge_dialect_program and the delegated ExecutorchProgramManager as executorch_program:

import copy
import torch
from executorch.backends.arm import EthosUCompileSpec, EthosUPartitioner
from executorch.devtools import generate_etrecord
from executorch.exir import to_edge, to_edge_transform_and_lower

exported_program = torch.export.export(model.eval(), example_inputs)

edge_program = to_edge(exported_program)
edge_program_for_record = copy.deepcopy(edge_program)

compile_spec = EthosUCompileSpec("ethos-u55-128")
delegated_edge_program = to_edge_transform_and_lower(
    programs=exported_program,
    partitioner=[EthosUPartitioner(compile_spec)],
)
executorch_program = delegated_edge_program.to_executorch()

generate_etrecord(
    "model.etrecord",
    edge_dialect_program=edge_program_for_record,
    executorch_program=executorch_program,
)

Backends that emit delegate debug mapping can also let the adapter group those operators under a delegate namespace such as delegates/CoreMLBackend/partition_0. Backends that do not emit that mapping can still show the pre-delegation operators, but delegate grouping may be limited.

Development

Local ETRecord test fixtures live under src/etrecord_adapter_model_explorer/tests/fixtures/.

pytest