macOS Support Question

[jasonwang0228]

Hi! I'm trying to use SUPPORT on macOS but running into CUDA compatibility issues since CUDA isn't available on Mac.

A few questions:

1. Are there known workarounds for macOS users (Docker, Colab, MPS backend)?
2. Any plans for macOS/GPU-agnostic support?

I'm happy to test solutions or help document workarounds for other Mac users. Thanks!

[SteveJayH]

Hi,
Thank you for your interest in SUPPORT.
I don’t use macOS myself, so I can’t confirm the best workaround, but here are a few notes:

• Colab is probably the easiest option for quick testing, though model training can take more than a day, so the free tier may disconnect before completion. Running inference (testing) should work relatively easily, but without training the model on your own data, performance may be limited.
• Docker or MPS backend could work in theory, but I haven’t personally tested either for SUPPORT.

Currently, there’s no dedicated support for macOS, and there are no short-term plans to add it.
That said, contributions are welcome—if you’d like to open a pull request to help improve compatibility, that would be great. I might be able to take a look myself later this winter when things are less busy.

[jasonwang0228]

MPS works on Apple Silicon.
You just need to add a simple device picker and replace hard-coded .cuda() calls with .to(device).

import torch

if torch.cuda.is_available():
    device = torch.device("cuda")
elif torch.backends.mps.is_available():
    device = torch.device("mps")
else:
    device = torch.device("cpu")

If you want to reuse it across multiple files:

def get_device():
    if torch.cuda.is_available():
        return torch.device("cuda")
    elif torch.backends.mps.is_available():
        return torch.device("mps")
    else:
        return torch.device("cpu")

device = get_device()

Then replace:

model.cuda()
tensor.cuda()

with:

model.to(device)
tensor.to(device)

That’s sufficient to run training and testing on Mac MPS. Tested on Apple Silicon.