GrandaddyShmax/audiocraft_plus

MacOS Build and Performance

KannManMachen opened this issue · 3 comments

KannManMachen commented

Hey @GrandaddyShmax,

after you already helped me to solve a compatibility installation problem for MacOS (by pointing me to the additional branch for MacOS called "mac-os-fix"), I have a general question regarding the performance.

First, I noticed that the MacOS version ("MusicGen+ V1.2.8c Mac OS Version") is still different from the current release ("AudioCraft Plus - v2.0.0a") - but this is probably irrelevant for the way the (audio) data is generated I guess :). I had tried to generate a 4 seconds long file (mono, 32 kHz, small model. But after 10 minutes of waiting I had the feeling that maybe something is not working properly or that my hardware is not optimal.

My question: the data generation is done locally using your own processing power, isn't it? If yes: is use made of Apple's M1/2 chip with Neural Engine, as it happens for example with the tool "TextToSample") (which is also based on METAs Audiocraft)?

Thanks (again)

My Specs:

  • Apple M2 Max (12-core CPU, 30-core GPU, 16-core Neural Engine)
  • 64 GB RAM
  • MacOS Ventura 13.5 (current release)
GrandaddyShmax commented

im not sure i am capable of answering you this question as i do not own mac os or the apple's M1/2 chip 😅
in terms of version v2.0.0a for mac os, hopefully I will do it soon enough as it does require some changes.

Hopefully someone who does own mac os would be able to answer your question

  • 👍1
KannManMachen commented

Update:

the long calculation time surely had something to do with the download of the different models (in the background). At least I had seen corresponding processes afterwards in my terminal app. Probably the generation had not yet started at my first attempt :).

Here are my current export times:

  1. 3 sec, mono, 32 khz, small model = 28 s
  2. 3 sec, mono, 32 khz, medium model = 90 s
  3. 3 sec, mono, 32 khz, large model = 160 s
  • 👍1
tob-har commented

is use made of Apple's M1/2 chip with Neural Engine, as it happens for example with the tool "TextToSample") (which is also based on METAs Audiocraft)?

Unfortunately no. It is running on the CPU only.
And it loads the models on first time using them. This can cause quite some delay depending on your download bandwidth.

  • 👍1