armdeveloperecosystem/sme-executorch-profiling/export-model

Export Model

Operating Rules

• Run from the repository root with .venv activated.
• Export to model_profiling/out_<model>/artifacts/ for the Learning Path examples unless the user requests another output root.
• Prefer --backend xnnpack for SME2 profiling because the runner comparison is intended to show delegated backend behavior.
• Do not assume .etrecord exists. Check for it and use it when present, but do not fail the export solely because it was not produced.
• If exporting a custom model, verify it is registered in model_profiling/models/__init__.py before running the exporter.

Required Inputs

• Completed setup-workspace.
• Model name, dtype, backend, and output directory.
• For custom models: an implementation of EagerModelBase with stable get_eager_model() and get_example_inputs().

Procedure

1. Activate the environment:

``bash source .venv/bin/activate ``

1. For the built-in smoke model, export with:

``bash python model_profiling/export/export_model.py \ --model toy_cnn \ --backend xnnpack \ --dtype fp16 \ --outdir model_profiling/out_toy_cnn/artifacts ``

1. For a named model, use the same shape:

```bash MODEL_NAME="<model_name>" DTYPE="fp16" OUTDIR="model_profiling/out_${MODEL_NAME}/artifacts"

python model_profiling/export/export_model.py \ --model "${MODEL_NAME}" \ --backend xnnpack \ --dtype "${DTYPE}" \ --outdir "${OUTDIR}" \ --save-graph ```

1. For quantized exports, add --quantize only after confirming the ExecuTorch checkout supports the model's quantization path:

``bash python model_profiling/export/export_model.py \ --model "${MODEL_NAME}" \ --backend xnnpack \ --dtype fp32 \ --quantize \ --outdir "${OUTDIR}" ``

1. Validate the output:

``bash find "${OUTDIR}" -maxdepth 1 -name '*.pte' -type f -size +0 -print find "${OUTDIR}" -maxdepth 1 -name '*.etrecord' -type f -print PTE_FILE="$(find "${OUTDIR}" -maxdepth 1 -name '*.pte' -type f -size +0 | head -1)" test -n "${PTE_FILE}" ``

Success Criteria

• A non-empty .pte exists in the selected artifact directory.
• Optional graph.json exists when --save-graph was requested.
• Optional .etrecord is recorded for later analysis if the export flow produced one.

Failure Triage

• Unknown model: register the model under model_profiling/models/ or use an ExecuTorch example model name that exists in the current checkout.
• ImportError from model code: install the model dependency into .venv or adjust the model wrapper.
• Torch export failure: inspect unsupported ops, dynamic control flow, input signatures, dtype casts, and shape constraints.
• XNNPACK partitioning failure: retry with --backend portable to isolate export validity from delegation, then return to XNNPACK after fixing unsupported delegated ops.
• Quantization failure: remove --quantize unless the model appears in the current ExecuTorch quantization options.

Handoff

After export succeeds, use run-profiling with the .pte path and the runner paths for the target platform.