Wondering if there will be support for the mac M2 chip

[xinmans]

+1

[Debuggernaut]

came here wondering the same thing, incredulous that nobody has gotten it working already even though it's been out for so long

I guess I've been too online, since it's only been out for one day

[StevenCyb]

I did some changes as workaround to get the Step 1 and Step 2 running on my Mac with M1 with a pretty good performance. What I did: 1) Cloned main 2) On gradio_app.py changed rng = torch.Generator(device=memory_management.gpu).manual_seed(int(seed)) to use device=memory_management.mps 3) On memory_management.py I... 3.1) replaced cpu = torch.device('cpu') with mps = torch.device('mps') 3.2) changed torch.zeros((1, 1)).to(gpu, torch.float32) to use mps instead of gpu 3.3) modified load_models_to_gpu function to:

def load_models_to_gpu(models):
    global models_in_gpu

    if not isinstance(models, (tuple, list)):
        models = [models]

    models_to_remain = [m for m in set(models) if m in models_in_gpu]
    models_to_load = [m for m in set(models) if m not in models_in_gpu]
    models_to_unload = [m for m in set(models_in_gpu) if m not in models_to_remain]

    if not high_vram:
        for m in models_to_unload:
            with movable_bnb_model(m):
                m.to(mps)
            print('Unload to MPS:', m.__class__.__name__)
        models_in_gpu = models_to_remain

    for m in models:
        m.to(mps)  #
        print('Load to MPS:', m.__class__.__name__)

    models_in_gpu = list(set(models_in_gpu + models))
    torch.cuda.empty_cache()
    return

I know step 3.3 is quite dirty, but it does the trick for me.