Skip to main content

Conversion Flow

Supported input​

This repo is focused on LiteRT / TensorFlow Lite .tflite models.

What the converter does​

At a high level, the converter:

  • Reads the LiteRT / TensorFlow Lite .tflite model representation.
  • Maps supported operations into the TOSA-oriented path used by MLIA.
  • Prepares .tosamlir text or .tosa.mlirbc bytecode artifacts suitable for downstream MLIA backends.

Operational model​

In most workflows this converter is a dependency backend. You usually see the results of the later analysis backend rather than interacting with this package as a destination in its own right.

Output format​

The default route writes MLIR bytecode with a .tosa.mlirbc suffix. Workflows that request mlir-text output receive a .tosamlir artifact instead.

Maintenance considerations​

When changing the conversion path, keep registration, conversion behaviour, and any downstream assumptions aligned so the rest of the MLIA pipeline continues to consume the produced artifacts correctly.