(Optional) Use the MLIA Python API

Run MLIA from Python and compare two models

The Arm ML Inference Advisor (MLIA) CLI is the best way to learn the tool, inspect output, and debug your environment.

Use the Python API when you want another product, dashboard, workflow runner, or CI system to integrate MLIA.

The mlia Python package exposes the same advisor functionality used by the CLI. Use run_advisor() as the main API entry point. Use helper functions such as list_targets(), list_target_profiles(), and list_backends() to discover what the installed environment supports.

The following example shows how you can use the API to analyze two LiteRT model variants, and print results and advice:

    

        
        
cat > compare_mlia_models.py <<'PY'
from pathlib import Path

from mlia import run_advisor


models = [
    Path("tflite/mv2_fp32.tflite"),
    Path("tflite/mv2_int8.tflite"),
]

for model in models:
    result = run_advisor(
        advice_category="compatibility",
        target_profile="ethos-u85-256",
        model=model,
        backends=["vela"],
    )

    print()
    print(model)

    for item in result["results"]:
        print(item["kind"], item["status"])

        for advice in item.get("advice", []):
            print("advice:", advice["severity"], advice["message"].splitlines()[0])
PY

    

Run the comparison script:

    

        
        
python compare_mlia_models.py

    

The expected result is that mv2_fp32.tflite reports compatibility incompatible, while mv2_int8.tflite reports compatibility ok.

You might still see warning advice for both models. For example, MLIA can report that SOFTMAX is a suboptimal activation even when the quantized model is otherwise compatible with the NPU.

Compatibility tells you whether the model can map to the target. Advice can still point out ways to improve it.

You can extend the example to do the following:

• Extract selected metrics into a dashboard
• Compare performance between model revisions
• Fail a CI job if a key metric regresses
• Surface advice messages in an internal model review tool

Discover capabilities from Python

MLIA also exposes helper functions for discovery. Depending on the installed MLIA version, useful helpers can include the following:

    

        
        
cat > discover_mlia.py <<'PY'
from mlia import list_backends, list_target_profiles, list_targets


print(list_targets())
print(list_target_profiles())
print(list_backends())
PY

    

Run the discovery script:

    

        
        
python discover_mlia.py

    

Use discovery in integrations so that your product can report what the current environment supports.

What you’ve accomplished

You’ve used the Python API to run the same kind of analysis you performed from the CLI. You’ve also learned how to compare model variants programmatically and why the API is useful for product integration or automation.

You can now incorporate these MLIA checks into an automation or integration workflow.