Skip to content

Conversation

@alvoron
Copy link
Contributor

@alvoron alvoron commented Nov 26, 2025

Details:

  • Do not fuse per-channel DQ scales to Convolution
  • Do not select int8 executor if output precision is quantized and FQ is not fused (i.e. requantize scale can not be applied)

Tickets:

  • ticket-id

@alvoron alvoron added the platform: arm OpenVINO on ARM / ARM64 label Nov 26, 2025
@github-actions github-actions bot added the category: CPU OpenVINO CPU plugin label Nov 26, 2025
@alvoron alvoron marked this pull request as ready for review November 26, 2025 14:02
@alvoron alvoron requested review from a team as code owners November 26, 2025 14:02
Copy link
Contributor

@v-Golubev v-Golubev left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Could you please extend the existing tests by adding a test case with per channel dequantization?

@alvoron
Copy link
Contributor Author

alvoron commented Nov 27, 2025

Could you please extend the existing tests by adding a test case with per channel dequantization?

Added per-channel test case.

Copy link
Contributor

@v-Golubev v-Golubev left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM

Comment on lines +130 to +136
ov::element::Type expectedPrecision = element::f32;
#if defined(OPENVINO_ARCH_ARM64)
const auto& [inputShape, inputPrecision, quantizeIntervals, fqConstShapes, targetName] = this->GetParam();
if (fqConstShapes.empty()) {
expectedPrecision = quantizeIntervals[0][0] < 0.f ? element::i8 : element::u8;
}
#endif
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Could you please add a short comment on why do we expect these precisions? The main thing I'd expect to see here is the mention that we don't support per channel dequantization for quantized convolution, so in this case we fallback on f32 implementation

@v-Golubev v-Golubev assigned EgorDuplensky and unassigned v-Golubev Nov 28, 2025
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

category: CPU OpenVINO CPU plugin platform: arm OpenVINO on ARM / ARM64

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants