https://en.bioerrorlog.work/entry/ace-step-15-local-m2-macbook

These are my notes from trying out ACE-Step 1.5 locally.

Introduction

I decided to try running ACE-Step 1.5 on my local M2 MacBook — a music generation AI that's been getting a lot of attention for delivering strong performance despite being lightweight.

# Setup
MacBook Air
Chip: Apple M2
Memory: 16GB

Note: This article was translated from my original post.

Running the Music Generation AI "ACE-Step 1.5" Locally

Installing ACE-Step 1.5

I followed the official guide for installation.

First, if you don't already have the Python package manager uv installed, go ahead and install it.

# macOS / Linux
curl -LsSf https://astral.sh/uv/install.sh | sh

# Windows (PowerShell)
powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"

I wrote an article about the basics of uv before, so feel free to check that out too.

en.bioerrorlog.work

Once uv is installed, clone the ACE-Step 1.5 repository and install the dependencies.

git clone https://github.com/ACE-Step/ACE-Step-1.5.git
cd ACE-Step-1.5
uv sync

That's it for installation.

There are five ways to use it:

Gradio Web UI
Studio UI
Python API
REST API
CLI

This time, I'll go with the Gradio Web UI.

Running ACE-Step 1.5 with the Gradio Web UI

Launch the Gradio Web UI with the following command:

uv run acestep

Once it's up and running, open http://localhost:7860/ in your browser to access the Gradio Web UI.

Next, click the "Initialize Service" button to download the model.

I left all the settings at their defaults and clicked the button.

This kicks off the model download.

In my case, the download took about 40 minutes.

Once the download is complete, it's time to generate some music.

I clicked the "Click Me" button to randomly generate a prompt and lyrics, then hit "Generate Music" to create a track.

Randomly generating a prompt with "Click Me", then generating music with "Generate Music"

Generation took roughly 5–10 minutes.

Troubleshoot - RuntimeError: MPS backend out of memory

On my first attempt, I ran into the following error:

RuntimeError: MPS backend out of memory (MPS allocated: 11.74 GiB, other allocations: 9.95 GiB, max allowed: 18.13 GiB). Tried to allocate 10.99 MiB on private pool. Use PYTORCH_MPS_HIGH_WATERMARK_RATIO=0.0 to disable upper limit for memory allocations (may cause system failure).

This means there isn't enough GPU memory.

As the error message suggests, you can disable the memory cap by setting the environment variable PYTORCH_MPS_HIGH_WATERMARK_RATIO=0.0.

export PYTORCH_MPS_HIGH_WATERMARK_RATIO=0.0

After setting this and launching the Gradio Web UI, music generation completed without any errors.

Example Output

Here's an example of a generated track.

As mentioned, the prompt was auto-generated by the "Click Me" button:

A melancholic German singer-songwriter track with fingerpicked acoustic guitar, subtle piano, and heartfelt female vocals. The intimate production captures the pain of heartbreak and the journey toward healing.

Here's the result: