Skip to main content

CLI Guide

Use the core mlia CLI for Ethos-U runs. This guide focuses on the command patterns that are most useful in practice.

If you are starting from scratch​

A sensible first sequence is:

  1. List what is installed with mlia target list and mlia backend list.
  2. Run one compatibility command.
  3. Run one performance command with vela.
  4. Only then add a target-compatible Corstone backend or --json.

Compatibility​

Use compatibility when the first question is whether the model can map cleanly to the selected Ethos-U target profile.

mlia check my_model.tflite --target-profile ethos-u55-256 --compatibility

If you want to make the backend path explicit during investigation, add Vela:

mlia check my_model.tflite \
  --target-profile ethos-u65-512 \
  --compatibility \
  --backend vela

Fast performance estimate​

For most day-to-day work, start with vela:

mlia check my_model.tflite \
  --target-profile ethos-u55-256 \
  --performance \
  --backend vela

Use this when you want the quickest answer about whether the model looks promising and where the main cost seems to be.

Deeper performance investigation​

Use a target-compatible Corstone backend when you want more detail than the quick estimate gives you:

mlia check my_model.tflite \
  --target-profile ethos-u65-512 \
  --performance \
  --backend corstone-310

If you are trying to decide whether the deeper path changes your view of the model, run both backends in one invocation:

mlia check my_model.tflite \
  --target-profile ethos-u65-512 \
  --performance \
  --backend vela \
  --backend corstone-310

ExecuTorch AOT through Corstone​

The Corstone path can also run supported ExecuTorch workloads. If mlia-converters-pytorch is installed, MLIA can accept a .pt2 input and convert it to .pte before running the Corstone backend:

mlia check my_model.pt2 \
  --target-profile ethos-u55-256 \
  --performance \
  --backend corstone-300

If you already have a prepared .pte artifact, you can run it directly:

mlia check my_model.pte \
  --target-profile ethos-u85-256 \
  --performance \
  --backend corstone-320

ExecuTorch performance is supported only on Ethos-U55 with corstone-300 and Ethos-U85 with corstone-320. For Ethos-U65 .pt2 inputs, use --performance with --backend vela; Ethos-U65 .pte performance is unsupported. See the support matrix.

JSON output​

Use --json when you want to save results, compare runs, or inspect output more carefully later.

mlia check my_model.tflite \
  --target-profile ethos-u55-256 \
  --performance \
  --backend vela \
  --json

When a run is confusing​

If the result is not making sense, try this sequence:

  1. Rerun with one explicit backend at a time.
  2. Start with compatibility if unsupported mapping looks likely.
  3. Compare the top-level cycle counts before reading deeper detail.
  4. Use a Corstone backend only after you know what the quick estimate is telling you.

Quick rules of thumb​

  • Use --compatibility first if you suspect unsupported operators.
  • Use --backend vela first if you want the fastest useful answer.
  • Use a target-compatible Corstone backend such as --backend corstone-300, --backend corstone-310, or --backend corstone-320 when you need deeper evidence before deciding what to change.
  • Use .pt2 or .pte with Corstone only when you are explicitly exercising the ExecuTorch AOT path.
  • Use --json when the result is something you expect to compare or archive.