Updated RQ signature

[nikita-malininn]

Changes

• Updated ReferenceQuantize.backward signature to align with the CUDA/CPU extensions signature.

Reason for changes

• Bugfix

Related tickets

• 166245

Tests

• Added tests/torch/quantization/test_functions.py

[ljaljushkin]

There's no test to catch the bug. Please add it.
E.g. you can mock call of Quantize_backward and check arguments type via call_args. This test would fail before this PR and pass with it.

[ljaljushkin]

There's no test to catch the bug. Please add it. E.g. you can mock call of Quantize_backward and check arguments type via call_args. This test would fail before this PR and pass with it.

call_args probably won't help with the test, but still there's no test checking that extension initialized by reference code is functional.
Before your PR the extension wasn't functional in that case and nobody has noticed that for several years until now, because there was no test.

[nikita-malininn]

There's no test to catch the bug. Please add it. E.g. you can mock call of Quantize_backward and check arguments type via call_args. This test would fail before this PR and pass with it.

call_args probably won't help with the test, but still there's no test checking that extension initialized by reference code is functional. Before your PR the extension wasn't functional in that case and nobody has noticed that for several years until now, because there was no test.

The test was added.

[ljaljushkin]

Suggested change

	self.input_low = torch.tensor([-0.5])
	self.input_range = torch.tensor([1.0])
	self.input_low = torch.tensor([[-0.5]])
	self.input_range = torch.tensor([[1.0]])

I guess there's a bug in sum_like when number of dimension doesn't match. In that way test passes for cuda

[nikita-malininn]

Tests passed with your fix. Thanks.

[ljaljushkin]

Do you have understanding why it didn't work without it? Was it valid behavior? Is some fix required in sum_like?

[nikita-malininn]

Not yet. I'll do my best to provide more details soon.

[nikita-malininn]

I've checked the test outputs regarding the issue.
I think that sum_like operates correctly, and the reason for failures was the incorrect input_low and input_range shapes.
For example, if the input_ shape [1, 2], then the low and range shapes should be [1, 1] each, but not [1].

Please, address your comment in the ticket.

[alexsu52]

LGTM