[Tests] Skip layerwise casting tests on devices without float8_e4m3fn support

[GiGiKoneti]

What does this PR do?

torch.float8_e4m3fn is a storage-only dtype that is not supported on all hardware backends. On Apple Silicon (MPS), attempting to cast a tensor to this dtype raises a TypeError:

TypeError: Cannot convert a MPS Tensor to float8_e4m3fn dtype as the MPS framework doesn't support float8 types.

This causes five layerwise-casting tests to crash instead of being skipped:

File	Test
tests/pipelines/test_pipelines_common.py	test_layerwise_casting_inference
tests/models/testing_utils/memory.py	test_layerwise_casting_memory
tests/models/testing_utils/memory.py	test_layerwise_casting_training
tests/lora/utils.py	test_layerwise_casting_inference_denoiser
tests/lora/utils.py	test_layerwise_casting_peft_input_autocast_denoiser

Fix

Each affected test now includes a runtime check at the very top:

try:
    torch.zeros(1, device=torch_device).to(dtype=torch.float8_e4m3fn)
except TypeError:
    self.skipTest(f"Device {torch_device} does not support float8 storage dtype.")

If the device cannot handle float8_e4m3fn, the test is skipped with a clear message rather than crashing the entire test suite.

Fixes #14072

Before submitting

• Did you use an AI agent (Claude Code, Codex, Cursor, etc.) to help with this PR? If so:
  • Did you read the Coding with AI agents guide?
  • Did you self-review the diff against .ai/review-rules.md?
• Did you read the contributor guideline?
• Did you read our philosophy doc? (important for complex PRs)
• Was this discussed/approved via a GitHub issue or the forum? Please add a link to it if that's the case.
  • Issue: [Bug] Layerwise casting tests fail on MPS backend due to float8 unsupported dtype #14072
• Did you make sure to update the documentation with your changes? (N/A — no docs changes needed)
• Did you write any new necessary tests? (N/A — this fixes existing tests)
• Are you the author (or part of the team) of the model/pipeline? (N/A)

Who can review?

@pcuenca (MPS), @sayakpaul @DN6 (General functionalities)

[GiGiKoneti]

Alternative approaches considered

This PR uses an inline try/except guard at the top of each affected test. It works, but I wanted to flag two cleaner alternatives for your consideration:

1. @require_float8_support decorator (preferred)

Instead of repeating the same 4-line block across 5 tests, a reusable decorator in the test utilities would be more consistent with the existing @require_* pattern in diffusers:

def require_float8_support(test_func):
    @functools.wraps(test_func)
    def wrapper(*args, **kwargs):
        try:
            torch.zeros(1, device=torch_device).to(dtype=torch.float8_e4m3fn)
        except TypeError:
            raise unittest.SkipTest(f"Device {torch_device} does not support float8_e4m3fn")
        return test_func(*args, **kwargs)
    return wrapper

Each test then becomes:

@require_float8_support
def test_layerwise_casting_inference(self):
    ...

2. Guard in source code (enable_layerwise_casting)

A more complete fix would be to also add a check inside enable_layerwise_casting() itself, raising a clear ValueError when the device does not support the requested storage_dtype, rather than letting it crash deep inside the pipeline forward pass. The test-side skip would still be needed, but users calling the API directly would get a much better error message.

---

let me know which you prefer and I will update the PR. cc @pcuenca @DN6

[sayakpaul]

Would it not be better to use a skip_mps decorator directly to skip those tests?

diffusers/src/diffusers/utils/testing_utils.py

Line 468 in bd2c919

def skip_mps(test_case):

[GiGiKoneti]

Using the @skip_mps decorator is much cleaner and fits the existing test codebase conventions perfectly.
The only reason I initially went with the dynamic try/except check was to make it future-proof and device-agnostic (in case other accelerators/configurations also lack float8 casting support and trigger a similar TypeError). But since MPS is the primary backend experiencing this crash and @skip_mps is idiomatic for the suite, using it is definitely better.

u want me to update PR ?

[sayakpaul]

Future-proof would xfailing those tests when torch_device == "mps". That way, if support is added in the future, xfail won't work anymore.

[GiGiKoneti]

@sayakpaul
Since float8 support is a device capability (similar to bfloat16), would you prefer we xfail on MPS specifically to track when they add support, or use a generic @require_float8 decorator to skip on any unsupported device?

[GiGiKoneti]

@sayakpaul
I have updated the PR to use @pytest.mark.xfail(condition=torch_device == "mps", reason="MPS does not support float8 casting.", strict=True) for all five affected layerwise casting tests instead of skipping them.

This way, if MPS adds float8 casting support in the future, the tests will unexpectedly pass (XPASS) and fail the suite, alerting us to remove the decorator.