fix(randn_tensor): compare device.type, not torch.device, when suppressing MPS info log

[Ricardo-M-L]

What does this PR do?

Fixes a latent bug in randn_tensor where the MPS-specific info-log suppression has been broken since PR #1902 (Jan 2023).

The bug

# src/diffusers/utils/torch_utils.py
if isinstance(device, str):
    device = torch.device(device)             # ← device is now a torch.device object
...
if generator is not None:
    gen_device_type = ...
    if gen_device_type != device.type and gen_device_type == "cpu":
        rand_device = "cpu"
        if device != "mps":                   # ← always True
            logger.info("The passed generator was created on 'cpu' ...")

The if device != "mps": guard was meant to silence the info log on MPS, because MPS doesn't support device-side generators — suggesting the user "create a generator on the mps device" would be misleading.

But by this point device is a torch.device object (coerced a few lines earlier). torch.device("mps") == "mps" is False — torch.device.__eq__ returns NotImplemented when compared with a string, Python falls back to identity, and the two are different types. So the guard is effectively always true, and MPS users get the spurious log on every call where a CPU generator is passed — the opposite of the documented intent.

Sanity check:

>>> import torch
>>> torch.device("mps") == "mps"
False
>>> torch.device("mps") != "mps"
True
>>> torch.device("mps").type == "mps"
True

The fix

One-character change: compare device.type (a str) to "mps".

-            if device != "mps":
+            if device.type != "mps":

Test

Added RandnTensorTester.test_mps_suppresses_cpu_generator_info_log in tests/others/test_utils.py, which:

• Asserts no CPU-fallback info log is emitted when a CPU generator targets MPS.
• Asserts the log is emitted when a CPU generator targets a non-MPS accelerator.

Verified that the test fails on main and passes with this fix. All existing tests in tests/others/test_utils.py still pass (14/14).

Before submitting

• This PR fixes a bug that affects MPS users.
• Did you read the contributor guideline?
• Did you write any new necessary tests? — Yes, regression test added.

Who can review?

@sayakpaul @DN6

[yiyixuxu]

thanks
I left a comment

[yiyixuxu]

we can use this? https://github.com/huggingface/diffusers/blob/main/tests/testing_utils.py#L1320

[Ricardo-M-L]

Good call! Refactored the test to use CaptureLogger — much cleaner. Also rebased on latest main to resolve the conflict.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.