script

LiteRT model converter

model-converter.py wraps the Arm Ethos-U Vela compiler. It reads target settings and model information from a generated *.cbuild-mlops.yml file and converts one or more quantized LiteRT (TensorFlow Lite) input model files.

For every input model, the converter generates:

• a Vela-optimized .tflite model;
• a C source file containing the optimized model as a byte array; and
• a Markdown file containing the Vela report.

Prerequisites

• Python 3.13 or newer
• Arm Ethos-U Vela available as vela on PATH
• PyYAML

Install the Python dependency:

python -m pip install -r script/requirements.txt

Confirm that Vela is available:

vela --version

Basic usage

python script/model-converter.py <project.cbuild-mlops.yml>

The path to the *.cbuild-mlops.yml file may be absolute or relative. Example:

python script/model-converter.py Test-Ethos-U55.cbuild-mlops.yml

Using the converter from VS Code

The example contains a .vscode.d/tasks.json drop-in for the CMSIS Solution extension. After opening or updating the solution, run Tasks: Run Task from the Command Palette and select Convert LiteRT models. Choose the generated MLOps configuration for U55, U65, or U85 when prompted.

If the task is not visible yet, run Update Debug Tasks and Launch Configurations from the Command Palette. The extension merges the drop-in task into the generated workspace .vscode/tasks.json file.

Generate the selected *.cbuild-mlops.yml file by setting up or building the corresponding solution before running the conversion task.

MLOps input

The generated *.cbuild-mlops.yml supplies the Vela configuration and model selection:

cbuild-mlops:
  vela:
    ini: Model/vela.ini
    options: >-
      --accelerator-config ethos-u55-128
      --system-config Ethos_U55_High_End_Embedded
      --memory-mode Shared_Sram
    misc: --verbose-all
  model:
    dir: Model
    name: tiny_cnn/tiny_cnn_int8.tflite

Paths are resolved relative to the directory containing the MLOps file.

The converter uses from the *.cbuild-mlops.yml file the following nodes:

• vela.ini as Vela's configuration file when present;
• vela.options for generated Vela arguments;
• vela.misc for optional additional Vela arguments;
• model.dir as the model working directory; and
• model.name to select one or more input models.

Converting one model

Use a string-valued model.name for a single model:

model:
  dir: Model
  name: tiny_cnn/tiny_cnn_int8.tflite

The model path is relative to model.dir.

Converting multiple models

Use a sequence-valued model.name for multiple models:

model:
  dir: Model
  name:
    - hello_world/hello_world_int8.tflite
    - tiny_cnn/tiny_cnn_int8.tflite

Models are converted in declared order and must be located below model.dir.

Overwrite MLOps settings

Three optional arguments override settings from the *.cbuild-mlops.yml file:

--system <name>    Vela system configuration
--memory <name>    Vela memory mode
--misc <options>   Complete miscellaneous option string

Change the system or memory configuration:

python script/model-converter.py Test-Ethos-U55.cbuild-mlops.yml --system Ethos_U55_High_End_Embedded
python script/model-converter.py Test-Ethos-U55.cbuild-mlops.yml --memory Shared_Sram

Replace all miscellaneous Vela options:

python script/model-converter.py Test-Ethos-U55.cbuild-mlops.yml --misc="--verbose-all --timing"

An empty value removes them:

python script/model-converter.py Test-Ethos-U55.cbuild-mlops.yml --misc=""

--misc replaces vela.misc; it does not append to it.

Generated files

Given this input:

Model/tiny_cnn/tiny_cnn_int8.tflite

the converter produces:

Model/tiny_cnn/tiny_cnn_int8_vela.tflite
Model/tiny_cnn/tiny_cnn_model.c
Model/tiny_cnn/VELA_SUMMARY.md

The _int8 suffix is removed when naming the C source. For network.tflite, the outputs are network_vela.tflite and network_model.c.

The C source exports:

const uint8_t tiny_cnn_int8_vela_tflite[];
const uint32_t tiny_cnn_int8_vela_tflite_len;

The array is aligned to 16 bytes and placed in the ethos_model section.

VELA_SUMMARY.md records the effective accelerator, system, and memory configuration followed by Vela's report. Command-line overrides are reflected in this file.

Outputs are prepared in temporary storage and replace destination files only after Vela succeeds.

Exit status and errors

• 0: all models were converted.
• 1: Vela or output generation failed.
• 2: command-line or YAML configuration is invalid.

The converter validates configuration files, directories, and models before starting Vela. Common errors include:

• Vela is missing from PATH;
• model.dir, a model, or the Vela INI file does not exist;
• model.name is missing, empty, or malformed; and
• a model path escapes model.dir.

Running the tests

From the example directory:

python -m unittest discover -s script_test -p "test_*.py" -v

The suite includes a real Vela conversion when vela is installed.