Mac Studio M2 Ultra 64GB

[michristian]

I'm using a Mac Studio M2 Ultra with 64GB, which is one of the top Macs, but when I tried using stable-audio-tools to generate a 5-second sound effect, it took over 10 minutes for just 20 steps. And after all that, I couldn't even get the audio output.

Can any other Mac users share if they were able to get it to work?

--

• Running on local URL: http://127.0.0.1:7860

To create a public link, set share=True in launch().
Prompt: glass break
1357556058
/Users/michristian/stable-audio-tools/stable_audio_tools/models/conditioners.py:353: FutureWarning: torch.cuda.amp.autocast(args...) is deprecated. Please use torch.amp.autocast('cuda', args...) instead.
with torch.cuda.amp.autocast(dtype=torch.float16) and torch.set_grad_enabled(self.enable_grad):
/Users/michristian/miniconda3/envs/audio_tools/lib/python3.10/site-packages/torch/amp/autocast_mode.py:266: UserWarning: User provided device_type of 'cuda', but CUDA is not available. Disabling
warnings.warn(
huggingface/tokenizers: The current process just got forked, after parallelism has already been used. Disabling parallelism to avoid deadlocks...
To disable this warning, you can either:
- Avoid using tokenizers before the fork if possible
- Explicitly set the environment variable TOKENIZERS_PARALLELISM=(true | false)
90%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████ | 18/20 [09:40<01:04, 32.26s/it]/Users/michristian/miniconda3/envs/audio_tools/lib/python3.10/site-packages/torchsde/_brownian/brownian_interval.py:599: UserWarning: Should have ta>=t0 but got ta=0.009999999776482582 and t0=0.01.
warnings.warn(f"Should have ta>=t0 but got ta={ta} and t0={self._start}.")
100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 20/20 [10:45<00:00, 32.25s/it]

[fred-dev]

To make use of anything other than the cpu on a mac you need to use an MPS compatible version of PyTorch (and then lightning should take care of the rest). You can check if you have correctly installed the mps backend with something like this:

print(torch.version)
print(torch.backends.mps.is_available())

I have tried to run this code with MPS and I think there is an issue in that some of the operations are not yet supported by the PyTorch MPS backend. I only did some rudimentary tests and did not comprehensively debug this, but at first glance it seems this will not work with MPS, which is needed to use any kind of acceleration on a mac.

[sskalnik]

https://docs.pytorch.org/docs/stable/notes/mps.html

MPS should be supported since Pytorch 1.12.

I am using Pytorch 2.8.0, no issues.

[ymgenesis]

Really wished this was compatible with MPS. An hour to generate 40 seconds is unfortunate. I tried applying this PR, and it generates faster but the audio output is garbage.

[ymgenesis]

Been having much better luck with ACE-Step. Here's an example of what I can get from it using MPS. It uses the same Torch version, so yes it's definitely possible on MPS.

[Cortexelus]

Upgrade to Tahoe 26 which has upgraded MPS to Metal 4. The slow part is likely the convolution kernels in the decoder. The tensors are too big for earlier versions of MPS. It falls back to the CPU! You're not actually running it on your mac's GPU. Also use chunked decoding. Shrink the tensors, they go faster!!

…

On Mon, Nov 24, 2025 at 3:35 PM ymgenesis ***@***.***> wrote: *ymgenesis* left a comment (Stability-AI/stable-audio-tools#181) <#181 (comment)> Been having much better luck with ACE-Step <https://github.com/ace-step/ACE-Step>. Here's an example <https://github.com/user-attachments/files/23731693/output_20251121001159_0.mp3>, using MPS. — Reply to this email directly, view it on GitHub <#181 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AASXCZJXOSJ4A4OEEYPFSJL36NTYRAVCNFSM6AAAAABYUMXNT2VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTKNZSGYZDIOJZGE> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[Cortexelus]

Also there is a quantized tflite ARM version of stable audio open which runs in seconds on a CPU on a mac or iphone or android https://huggingface.co/innermost47/stable-audio-open-small-tflite https://huggingface.co/blog/Arm/ai-sound-gen-on-arm https://developer.arm.com/community/arm-community-blogs/b/ai-blog/posts/audio-generation-arm-cpus-stable-audio-open-small-kleidiai It may not be obvious how to run it I should just release an easy tutorial

…

On Mon, Nov 24, 2025 at 6:18 PM CJ Carr ***@***.***> wrote: Upgrade to Tahoe 26 which has upgraded MPS to Metal 4. The slow part is likely the convolution kernels in the decoder. The tensors are too big for earlier versions of MPS. It falls back to the CPU! You're not actually running it on your mac's GPU. Also use chunked decoding. Shrink the tensors, they go faster!! On Mon, Nov 24, 2025 at 3:35 PM ymgenesis ***@***.***> wrote: > *ymgenesis* left a comment (Stability-AI/stable-audio-tools#181) > <#181 (comment)> > > Been having much better luck with ACE-Step > <https://github.com/ace-step/ACE-Step>. Here's an example > <https://github.com/user-attachments/files/23731693/output_20251121001159_0.mp3>, > using MPS. > > — > Reply to this email directly, view it on GitHub > <#181 (comment)>, > or unsubscribe > <https://github.com/notifications/unsubscribe-auth/AASXCZJXOSJ4A4OEEYPFSJL36NTYRAVCNFSM6AAAAABYUMXNT2VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTKNZSGYZDIOJZGE> > . > You are receiving this because you are subscribed to this thread.Message > ID: ***@***.***> >