Reject complex input in the pt2e min/max observers instead of silently dropping the imaginary part

[vijay-kapse]

Fixes #4910.

Problem

MinMaxObserver, MovingAverageMinMaxObserver, PerChannelMinMaxObserver and
MovingAveragePerChannelMinMaxObserver cast the incoming tensor to the dtype of their
min/max buffers:

x = x.to(self.min_val.dtype)

For a complex tensor that buffer is real, so the imaginary part is dropped and the observer
records the min/max of the real part alone. calculate_qparams() then hands back a scale and
zero point that look ordinary, with nothing to signal that most of the input was discarded:

x = torch.tensor([1+100j, 2+200j, 3+300j])
obs = MinMaxObserver(); obs(x)
obs.min_val, obs.max_val     # 1.0, 3.0   <- the 100..300 imaginary range is gone
obs.calculate_qparams()      # scale 0.0117647, zero_point 0

The quantize kernels reject the input a step later anyway — PerTensorAffineQuantizer::quantize
and PerChannelAffineFloatQParamsQuantizer::quantize in ATen's Quantizer.cpp both require a
float tensor. So the observer and the kernel disagreed about whether complex input is allowed,
and the observer was the one that failed quietly.

Measured before the change

observer	complex input
MinMaxObserver	accepted, scale 0.0117647
MovingAverageMinMaxObserver	accepted, scale 0.0117647
PerChannelMinMaxObserver	accepted, scale 0.00784314
MovingAveragePerChannelMinMaxObserver	accepted, scale 0.00784314
HistogramObserver	raises NotImplementedError: "aminmax_cpu" not implemented for 'ComplexFloat'

Change

The four now raise NotImplementedError naming the observer and the dtype, which matches how
this module already reports unsupported configurations (MinMaxObserver's qscheme only support ...) and gives a bug report enough to identify the input:

NotImplementedError: MinMaxObserver does not support complex input, got torch.complex64

HistogramObserver needed no change — it does not perform the cast and already fails through
torch.aminmax. I checked it rather than assuming; the four above now behave the way it
already does.

The numel() == 0 early return is deliberately left ahead of the check, so an empty tensor is
still returned unobserved whatever its dtype. That path is pinned by a test.

Tests

Adds test/quantization/pt2e/test_observer.py (there was no pt2e observer test module). It
covers all five observers over complex64 and complex128, asserts the message names both the
observer and the dtype, and pins that real input and the empty-tensor path are unaffected.

5 passed

Reverting only the observer change fails exactly the two complex tests, so the coverage
discriminates rather than passing either way:

FAILED test_observer.py::TestObserverComplexInput::test_per_channel_observers_reject_complex
FAILED test_observer.py::TestObserverComplexInput::test_per_tensor_observers_reject_complex
2 failed, 3 passed

No regressions in the neighbouring suites: test/quantization/test_observer.py 11 passed,
test/quantization/pt2e/test_duplicate_dq.py 4 passed. ruff format --check and ruff check
are clean on both files.

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/4970

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

This comment was automatically generated by Dr. CI and updates every 15 minutes.