[OpenVINO] calibration_device parameter

[daniil-lyakhov]

Changes

Adds a calibration_device parameter to AdvancedCompressionParameters and AdvancedQuantizationParameters that allows users to specify which OpenVINO device to use for calibration inference (e.g. "GPU", "AUTO:GPU,CPU").

Source changes:

• src/nncf/openvino/engine.py — Added calibration_device_context() context manager using contextvars.ContextVar. OVNativeEngine reads the context variable to determine the compile device instead of hardcoding "CPU". FP32 inference precision config is only applied for CPU device.
• src/nncf/quantization/advanced_parameters.py — Added calibration_device: str | None = None field to both AdvancedQuantizationParameters and AdvancedCompressionParameters.
• src/nncf/openvino/quantization/quantize_model.py — Wraps calibration calls in calibration_device_context() for quantize_impl, quantize_with_accuracy_control_impl, and compress_weights_impl.
• src/nncf/quantization/quantize_model.py — Added ParameterNotSupportedError guards for calibration_device in non-OV backends (Torch, TorchFX, ONNX) across compress_weights, quantize, and quantize_with_accuracy_control.

Reason for changes

Up to 120x speed up for PI0.5 quantization calibration on Intel(R) Arc(TM) B580 Graphics

Users need the ability to run calibration inference on a device other than CPU (e.g. GPU) for faster quantization/compression. This parameter is only meaningful for the OpenVINO backend; other backends now raise an explicit ParameterNotSupportedError instead of silently ignoring it.

Related tickets

184686

Tests

• tests/openvino/native/test_engine.py — Unit test for calibration_device_context and OVNativeEngine device propagation.

• tests/cross_fw/test_templates/template_test_weights_compression.py — Template test test_compress_weights_calibration_device verifying ParameterNotSupportedError is raised on non-OV backends.

• tests/openvino/native/quantization/test_weights_compression.py — OV override verifying the device is correctly passed through to ov.Core.compile_model via monkeypatch.

• tests/cross_fw/test_templates/template_test_quantize_api.py — New template TemplateTestQuantizeApi with test_quantize_calibration_device for non-OV backends.

• tests/openvino/native/quantization/test_quantize_api.py — OV overrides for quantize and quantize_with_accuracy_control verifying device propagation.

[Copilot]

Pull request overview

Adds an OpenVINO-specific calibration_device advanced parameter to route calibration-time inference to a user-selected OpenVINO device (e.g., GPU) and makes non-OpenVINO backends explicitly reject this option.

Changes:

• Added calibration_device: str | None to advanced quantization and compression parameter dataclasses (documented as OpenVINO-only).
• Implemented a calibration_device_context() (contextvars-based) and made OVNativeEngine compile models for the context-selected device.
• Propagated the option through OpenVINO quantization/weight-compression flows and added non-OV guards + cross-framework tests.

Reviewed changes

Copilot reviewed 12 out of 12 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
src/nncf/openvino/engine.py	Introduces calibration_device_context() and makes OVNativeEngine compile on the context-selected device (CPU default).
src/nncf/openvino/quantization/quantize_model.py	Wraps OpenVINO calibration-related algorithm steps in calibration_device_context() to ensure device propagation.
src/nncf/quantization/advanced_parameters.py	Adds calibration_device field + docstring entries to AdvancedQuantizationParameters and AdvancedCompressionParameters.
src/nncf/quantization/quantize_model.py	Adds ParameterNotSupportedError guards for calibration_device on ONNX/Torch/TorchFX backends.
tests/openvino/native/test_engine.py	Adds unit test verifying calibration_device_context() affects OVNativeEngine compile device and CPU-only FP32 config behavior.
tests/openvino/native/quantization/test_quantize_api.py	Adds OpenVINO device-propagation tests for quantize and quantize_with_accuracy_control; refactors subset size validation test.
tests/openvino/native/quantization/test_weights_compression.py	Adds OpenVINO device-propagation test for compress_weights(...calibration_device=...).
tests/cross_fw/test_templates/template_test_quantize_api.py	Adds a cross-framework template asserting non-OV backends reject calibration_device for quantize.
tests/torch/function_hook/quantization/test_quantize_api.py	Instantiates the template for Torch Function Hook backend.
tests/torch/fx/test_quantize_api.py	Instantiates the template for Torch FX backend using a minimal exported model.
tests/onnx/quantization/test_quantize_api.py	Instantiates the template for ONNX and adds a quantize_with_accuracy_control rejection test.
tests/cross_fw/test_templates/template_test_weights_compression.py	Adds a cross-framework template test asserting non-OV backends reject calibration_device for compress_weights.