fix(freeu): run FFT in float32 for float16 inputs to avoid ComplexHalf

[Ricardo-M-L]

What does this PR do?

Closes #12504.

fourier_filter (the FFT helper used by enable_freeu) already upcasts
bfloat16 inputs to float32 before calling torch.fft.fftn, because
PyTorch's FFT does not support bf16. The same is true for float16:
depending on the PyTorch version, fftn either

• produces the experimental torch.complex32 (ComplexHalf) dtype and
  emits a UserWarning: ComplexHalf support is experimental… (the
  original symptom in freeU ComplexHalf warning #12504), or
• raises RuntimeError: Unsupported dtype Half outright (reproduced on
  CPU with torch==2.9.0).

Both paths were reachable from FreeU with half-precision models
(sd-turbo + torch_dtype=torch.float16 + enable_freeu(…)).

Fix

Extend the existing upcast branch to cover float16 as well as
bfloat16. The function already downcasts back to x_in.dtype at the
end, so the externally observable dtype is unchanged.

    # Non-power of 2 images must be float32
    if (W & (W - 1)) != 0 or (H & (H - 1)) != 0:
        x = x.to(dtype=torch.float32)
    # fftn does not support bfloat16, and produces the experimental ComplexHalf
    # dtype (torch.complex32) when given float16, which is numerically unstable
    # and triggers a UserWarning.
    elif x.dtype in (torch.bfloat16, torch.float16):
        x = x.to(dtype=torch.float32)

Running FFT in fp32 and downcasting is also more numerically accurate
than staying in fp16/complex32.

Tests

Added tests/others/test_utils.py::FourierFilterTester with three
cases:

• test_fourier_filter_float16_no_complexhalf_warning — fp16 input
  produces no ComplexHalf warning and returns an fp16 tensor.
• test_fourier_filter_bfloat16_no_complexhalf_warning — regression
  guard for the existing bf16 path.
• test_fourier_filter_preserves_dtype_and_shape — verifies dtype and
  shape round-trip for all three supported dtypes.

Locally on torch==2.9.0+cpu, these tests fail without the patch
(RuntimeError for fp16) and pass with it.

tests/others/test_utils.py::FourierFilterTester::test_fourier_filter_bfloat16_no_complexhalf_warning PASSED
tests/others/test_utils.py::FourierFilterTester::test_fourier_filter_float16_no_complexhalf_warning PASSED
tests/others/test_utils.py::FourierFilterTester::test_fourier_filter_preserves_dtype_and_shape PASSED

Why not PR #12511?

PR #12511 addresses the same warning by locally suppressing it with
warnings.catch_warnings() around fftn/ifftn. That hides the
warning but keeps the FFT running on ComplexHalf (or raising on
platforms where fp16 FFT is unsupported). This PR instead removes the
root cause: the FFT is run in fp32, matching the existing bf16 handling.

Before submitting

• Did you read the contributor guideline?
• Did you read our philosophy doc?
• Was this discussed/approved via a GitHub issue? (See freeU ComplexHalf warning #12504.)
• Did you make sure to update the documentation with your changes? (No docs touch API surface.)
• Did you write any new necessary tests?

Who can review?

@sayakpaul @yiyixuxu

AI assistance (Claude) was used to draft this patch. I reviewed every
line and ran the tests locally.

[sayakpaul]

Suggested change

	elif x.dtype in (torch.bfloat16, torch.float16):
	elif x.dtype != torch.float32:

[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.