Apple Silicon Support #115
Replies: 7 comments 10 replies
-
|
I've also looked into this and came back with the impression the MPS backend is not really ready yet. There's an issue here tracking supported operations pytorch/pytorch#77764 |
Beta Was this translation helpful? Give feedback.
-
|
Whoops wrong heading. :) |
Beta Was this translation helpful? Give feedback.
-
|
You could just do the ffts on cpu (or use precompiled conditioning latents). There is a section on the latter in the docs. I believe that is the only step that computes an fft. The rest of the model is traditional matrix math. It should probably work. The only other thing I'd be concerned about is memory consumption. Id love to hear how fast it is if you get it working. |
Beta Was this translation helpful? Give feedback.
-
|
Cool, thanks for the feedback! This project is amazing and I have a use, so I'll probably go to the cloud in the meantime, but I'll keep an eye on the MPS op coverage tracking issue you posted and keep tinkering with Apple Silicon support 🤓 |
Beta Was this translation helpful? Give feedback.
-
|
I'm going to accept your "MPS backend is not really ready yet" feedback and walk away from Apple Silicon support for now 😂
|
Beta Was this translation helpful? Give feedback.
-
|
LOL. |
Beta Was this translation helpful? Give feedback.
-
|
After a year it's still the same :'), why pytorch why! |
Beta Was this translation helpful? Give feedback.
-
Any news on "mps" support? |
Beta Was this translation helpful? Give feedback.
-
|
Try it, it might work with pytorch nightly now |
Beta Was this translation helpful? Give feedback.
-
I already tried it before posting the question. It didn't work, hence why I'm asking. Maybe someone found a way to make it work. |
Beta Was this translation helpful? Give feedback.
-
|
On nightly? What error are you getting? |
Beta Was this translation helpful? Give feedback.
-
|
Yes on nightly, I installed all the dependencies on my conda environment, everyhting works on cpu side, but it takes so much time, therefore I tried to add an output to see if the model is training on cpu or mps, unfortunately it was using the cpu, even though I tried forcing it to use mps. Here's the last block of code that I excuted to see if it's mps or cpu:
When I run verification code from another notebook it works fine!
This makes me guess that we need to change the code inside the tortoise api and swap it from cuda to mps too to make this work. |
Beta Was this translation helpful? Give feedback.
-
|
Do anyone have a version working on his mac arm machine? |
Beta Was this translation helpful? Give feedback.
-
|
No luck here either. |
Beta Was this translation helpful? Give feedback.
-
|
Is there any update on this? It has been over a year. |
Beta Was this translation helpful? Give feedback.
-
|
Nope tried it a week ago still got metal not supported for a attn:? function |
Beta Was this translation helpful? Give feedback.
-
|
Merged in #550! |
Beta Was this translation helpful? Give feedback.
-
Hi! Any plans to add Apple Silicon support? I'm just a basic C# programmer professionally, but I gave it a good try. I had to compile torch audio from source to get tortoise-tts running natively on my M1 Max with GPU support. Unfortunately, it seems that the M1 Max GPU (or specifically the "mps" device equivalent to "cuda" in torch) doesn't support ... something over my head. I got errors about fft ops not being supported.
I don't know how involved this would be, but if it could be made to work it seems like the unified memory architecture in Apple Silicon could be an advantage. Thanks!
Beta Was this translation helpful? Give feedback.
All reactions