How We Implemented MLX-Demo (MacOS)

This document explains how DeepExtension internally implemented the built-in training method: MLX-Demo.

The MLX-Demo serves as a read-only reference example that demonstrates how to integrate third-party training code into the DeepExtension workflow.

---

Demo Source

• MLX-Demo is based on the LoRA training implementation from the mlx-lm project

After following the Installation Guide, the necessary files are located at:

{deepextension_base_dir}/deep-e-python/mlx_lm

---

Prerequisites: Make It Runnable Locally

By default, the MLX example reads datasets from directories, while DeepExtension requires file-based inputs.
Some minor changes are necessary and are automatically applied when running prepare_mlx_changes.sh during installation.

Base Model (Local)

Ensure a compatible base model (e.g., Qwen2.5-1.5B-Instruct) is downloaded to:

{deepextension_base_dir}/models/Qwen2.5-1.5B-Instruct

You may use any other MLX-compatible base model available locally.

Dataset (Local)

For the MLX-Demo, use the following dataset:

examples/mlx-demo-train-dataset.jsonl

---

Test the Localized Script

Before integrating, verify that the MLX demo code runs correctly on your machine.

cd {deepextension_base_dir}/deep-e-python

python3 -m mlx_lm.lora \
    --model ../models/Qwen2.5-1.5B-Instruct \
    --train \
    --data examples/mlx-demo-train-dataset.jsonl \
    --iters 2

If everything is set up properly, the script should run and terminate without errors.

Only proceed to the next integration steps if this test is successful.

---

Integrating mlx-demo.py

After validating the training script, we created mlx-demo.py.

Unlike the CUDA demo described in How You Implement Own Training (CUDA), which embeds training logic directly inside a script,
mlx-demo.py instead invokes training using a subprocess call:

cmd = ['python3', '-m', 'mlx_lm','lora']

This makes the integration more modular and easier to maintain when using well-structured external modules like mlx_lm.

---

Summary

• MLX-Demo is a working example that shows how to integrate MLX-based training pipelines into DeepExtension.
• It demonstrates that even non-CUDA environments like MacOS can support training workflows via DeepExtension.
• The script structure makes it easy to replicate for other MLX-compatible fine-tuning tools.

---

DeepExtension — Bring real training examples into your enterprise AI stack.
No hacks. No surprises. Just clean integration.