MPS device not properly tested in EulerAncestralDiscreteScheduler

[florian-ivadolabs]

Describe the bug

When the device set to mps, str(device) can return "mps:0".
The proper way to test the device would be device.type == "mps".
The wrong test is used in EulerAncestralDiscreteScheduler preventing to run the right code:

device = model_output.device if torch.is_tensor(model_output) else "cpu"
if str(device) == "mps":
    # randn does not work reproducibly on mps
    noise = torch.randn(model_output.shape, dtype=model_output.dtype, device="cpu", generator=generator).to(
        device
    )
else:
    noise = torch.randn(model_output.shape, dtype=model_output.dtype, device=device, generator=generator).to(
        device
    )

if str(device) == "mps" should be replaced by if device.type == "mps":

Reproduction

Run twice an inference with the same seed and the EulerAncestral scheduler doest generate the same image

Logs

No log as no error

System Info

diffusers 0.7.2, Mac OS ventura, M1 Max env with Python 3.9.2

[pcuenca]

Will be fixed by #1169.

[pcuenca]

@florian-ivadolabs Should be fixed now. Thanks a lot!