TFLite(LiteRT) - Generative AI

Overview

The Generative AI section describes how to run large language model (LLM) and vision-language model (VLM) inference on Genio platforms running Yocto. MediaTek provides pre-compiled Deep Learning Archive (DLA) model packages and built-in command-line tools to enable immediate evaluation without requiring a host-side build environment.

For performance data and model downloads, refer to TFLite(LiteRT) - Generative AI.

Supported Platforms and DLA Grouping

Generative AI model packages are compiled for specific DLA hardware. Genio 360, 360P, 420, 520, and 720 share the same DLA binary, while MT8875, MT8883, and MT8893 each have their own dedicated DLA binary. Select the package that matches your platform group when downloading.

Platform Groups and DLA

Platform

DLA Package

Genio 360 / 360P / 420 / 520 / 720

Shared

MT8875

Dedicated

MT8883

Dedicated

MT8893

Dedicated

Built-in Inference Tools

The following inference tools are pre-installed in the Yocto image under /usr/share/tmp/llm/:

Tool

Description

llm_cmdline_tool

Runs LLM inference from a prompt file using a YAML configuration. See llm_cmdline_tool.

vlm_cmdline_tool

Runs VLM inference combining an image and text prompt. See vlm_cmdline_tool.