Usage and Integration

Installation

Start with a target plugin for the hardware you want to analyze. It installs MLIA as a dependency. Add this converter in the same Python environment only if your workflow needs LiteRT conversion. A converter alone cannot provide target-specific analysis.

pip install mlia-converters-litert

Overview

This package registers the transformer name tflite_to_tosa through the mlia.plugin.transformer entry point.

Naming convention

This repo uses two names that matter in different places:

• tflite_to_tosa: the transformer name exposed to MLIA workflows.
• tosa_converter_for_tflite: the Python implementation package.

Integration model

This package adds conversion to an MLIA workflow with a target plugin. It does not provide a standalone analysis command. Once installed, MLIA can discover the converter and route LiteRT / TensorFlow Lite .tflite conversion requests through the transformer registry.

Typical workflow

The converter usually participates in a larger MLIA flow:

1. MLIA receives a .tflite model.
2. The converter plugin transforms it into a TOSA-oriented form.
3. A downstream backend consumes the converted artifact.
4. MLIA returns target-specific analysis results.

Output controls

The transformer supports two output formats:

• mlir-bytecode, which writes a .tosa.mlirbc artifact and is the default.
• mlir-text, which writes a .tosamlir artifact.

When emit_debug_info is enabled, the plugin passes --emit-debug-info to the underlying converter.

Source layout

• conversion.py: conversion logic.
• converter_plugin.py: MLIA plugin registration.
• install.py: installation metadata for MLIA-managed backend tooling.

Cross-links

• See conversion_flow.md for a more pipeline-oriented view.
• See conversion_outputs.md for what success looks like from this converter's perspective.
• See troubleshooting.md for conversion-specific failures.