Use Metal Performance Shaders when training SDV synthesizers on Apple Silicon machines

[liuup]

Problem Description

I have saw the CUDA support in SDV, but it only works on Nvidia GPU. Mac with Apple Silicon now have MPS backend to accelerate the computing process. Could SDV R&D team consider to add the support of MPS?

Some offical docs may help:
Apple: https://developer.apple.com/metal/pytorch/
PyTorch: https://pytorch.org/docs/master/notes/mps.html

[srinify]

Hi there @liuup 👋

This is a neat idea and I'll leave this open as a feature request! We are a small team so we appreciate your patience as we triage the different requests. We will typically prioritize based on demand and importance to your use case.

To help us prioritize this one, it would be great it if you could describe more about your project. Which synthesizer you're using, what you plan to do with the synthetic data, etc. 🙏

[TiStef]

Hi there @liuup 👋

This is a neat idea and I'll leave this open as a feature request! We are a small team so we appreciate your patience as we triage the different requests. We will typically prioritize based on demand and importance to your use case.

To help us prioritize this one, it would be great it if you could describe more about your project. Which synthesizer you're using, what you plan to do with the synthetic data, etc. 🙏

Good morning srinify,
same issue here. Particularly when using CTGAN, having torch.backend.mps added as GPU option would be REALLY helpful. Should be an easy fix which would enlarge SDV library to whoever code with Apple Silicon terminals.
Thank you so much for your great work and support

[yawetse]

@TiStef I added a pull request for CTGAN with Apple Metal support here: sdv-dev/CTGAN#425

[liuup]

@TiStef I added a pull request for CTGAN with Apple Metal support here: sdv-dev/CTGAN#425

Hi, I think this is helpful for this repo, I hope the code could be checked by maintainers and merged into main branch! Thanks for your contribution!